Upload 65 files

LICENSE

@@ -0,0 +1,20 @@
+MIT License
+
+Copyright (c) 2024 Phạm Huỳnh Anh
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+1. The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+2. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+
+Giấy phép MIT
+
+Bản quyền (c) 2024 Phạm Huỳnh Anh
+
+Được cấp phép miễn phí cho bất kỳ ai nhận được bản sao của phần mềm này và tài liệu liên quan (gọi chung là "Phần mềm"), để sử dụng Phần mềm mà không bị hạn chế, bao gồm nhưng không giới hạn quyền sử dụng, sao chép, chỉnh sửa, hợp nhất, xuất bản, phân phối, cấp phép lại và/hoặc bán bản sao của Phần mềm, và cho phép những người mà Phần mềm được cung cấp cũng được thực hiện điều đó, với điều kiện tuân thủ các điều khoản sau:
+
+1. Thông báo bản quyền ở trên và thông báo cấp phép này phải được bao gồm trong tất cả các bản sao hoặc phần đáng kể của Phần mềm.
+
+2. PHẦN MỀM ĐƯỢC CUNG CẤP "NHƯ LÀ", KHÔNG CÓ BẢO ĐẢM DƯỚI DẠNG NÀO, RÕ RÀNG HAY NGỤ Ý, BAO GỒM NHƯNG KHÔNG GIỚI HẠN Ở CÁC BẢO ĐẢM VỀ TÍNH THƯƠNG MẠI, ĐỘ BỀN ĐỂ SỬ DỤNG VÀ KHÔNG VI PHẠM. TRONG BẤT KỲ TRƯỜNG HỢP NÀO, CÁC TÁC GIẢ HOẶC CÁC CHỦ SỞ HỮU BẢN QUYỀN KHÔNG CHỊU TRÁCH NHIỆM VỀ BẤT KỲ KHIẾU NẠI, THIỆT HẠI HOẶC TRÁCH NHIỆM KHÁC, DÙ TRONG HÀNH ĐỘNG HỢP ĐỒNG, TIẾN HÀNH HOẶC KHÁC NỮA, PHÁT SINH TỪ HOẶC LIÊN QUAN ĐẾN PHẦN MỀM HOẶC SỰ SỬ DỤNG HAY CÁC THƯƠNG THẢO KHÁC TRONG PHẦN MỀM.

README.md

@@ -1,14 +1,80 @@
----
-title: RVC GUI
-emoji: 🚀
-colorFrom: green
-colorTo: blue
-sdk: gradio
-sdk_version: 5.6.0
-app_file: app.py
-pinned: false
-license: mit
-short_description: RVC
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+<div align="center">
+
+
+# Vietnamese RVC By Anh
+Công cụ chuyển đổi giọng nói chất lượng và hiệu suất cao đơn giản dành cho người Việt.
+
+
+[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/drive/18Ed5HbwcX0di6aJymX0EaUNz-xXU5uUc?hl=vi#scrollTo=ers351v_CMGN)
+[![Licence](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/PhamHuynhAnh16/Vietnamese-RVC/blob/main/LICENSE)
+
+</div>
+
+
+# Mô tả
+Dự án này là một công cụ chuyển đổi giọng nói đơn giản, dễ sử dụng, được thiết kế đặc biệt cho người Việt Nam. Với mục tiêu tạo ra các sản phẩm chuyển đổi giọng nói chất lượng cao và hiệu suất tối ưu, dự án cho phép người dùng thay đổi giọng nói một cách mượt mà, tự nhiên.
+
+# Các tính năng của dự án
+- Tách nhạc (MDX-Net/Demucs)
+
+- Chuyển đổi giọng nói (Chuyển đổi tệp/Chuyển đổi hàng loạt/Chuyển đổi văn bản)
+
+- Chỉnh sửa hiệu ứng âm thanh
+
+- Tạo dữ liệu huấn luyện (Từ đường dẫn liên kết)
+
+- Huấn luyện mô hình (v1/v2)
+
+- Dung hợp mô hình
+
+- Đọc thông tin mô hình
+
+- Tải xuống từ kho mô hình có sẳn
+
+# Cách sử dụng
+- Khởi động Command Prompt hoặc Terminal trong thư mục chính.
+- **Cài đặt các thư viện cần thiết:**
+```
+pip install -r requirements.txt của dự án
+```
+
+- **Khi cài đặt xong, để mở giao diện sử dụng:**
+```
+Chạy tệp: run_app.bat
+```
+**hoặc:**
+```
+Nhập vào Command Prompt hoặc Terminal: python main/app/app.py
+```
+- **Để khởi động biểu đồ kiểm tra huấn luyện:**
+```
+Chạy tệp: run_tensorboard.bat
+```
+**hoặc:**
+```
+Nhập vào Command Prompt hoặc Terminal: python main/app/run_tensorboard.py
+```
+
+# Điều khoản sử dụng: 
+- Bạn phải đảm bảo rằng các nội dung âm thanh bạn tải lên và chuyển đổi qua dự án này không vi phạm quyền sở hữu trí tuệ của bên thứ ba.
+
+- Không được phép sử dụng dự án này cho bất kỳ hoạt động nào bất hợp pháp, bao gồm nhưng không giới hạn ở việc sử dụng để lừa đảo, quấy rối, hay gây tổn hại đến người khác.
+
+- Bạn chịu trách nhiệm hoàn toàn đối với bất kỳ thiệt hại nào phát sinh từ việc sử dụng sản phẩm không đúng cách.
+
+- Tôi sẽ không chịu trách nhiệm với bất kỳ thiệt hại trực tiếp hoặc gián tiếp nào phát sinh từ việc sử dụng dự án này.
+
+# Yêu cầu để sử dụng
+- Python 3.10
+- Pip < 24.0
+
+- Đã cài đặt ffmpeg
+- Đã cài đặt requirements.txt
+
+# Dự án dựa trên các dự án chính như
+- **[Applio](https://github.com/IAHispano/Applio/tree/main)**
+- **[Retrieval-based-Voice-Conversion-WebUI](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/tree/main)**
+- **[Python-audio-separator](https://github.com/nomadkaraoke/python-audio-separator/tree/main)**
+
+# ☎️ Liên hệ tôi
+- Discord: **pham_huynh_anh**

assets/languages/en-US.json

@@ -0,0 +1,710 @@
+{
+    "set_precision": "Override preprocessing and config.json to use {precision}.",
+    "set_lang": "Display language set to {lang}.",
+    "no_support_gpu": "Unfortunately, no compatible GPU is available to support your training.",
+    "text": "text",
+    "upload_success": "File {name} uploaded successfully.",
+    "download_url": "Download from the link",
+    "download_from_csv": "Download from the CSV model repository",
+    "download_from_applio": "Download model from Applio",
+    "upload": "Upload",
+    "option_not_valid": "Invalid option!",
+    "list_model": "Model list",
+    "success": "Completed!",
+    "index": "index",
+    "model": "model",
+    "zip": "compress",
+    "search": "search",
+    "provide_file": "Please provide a valid {filename} file!",
+    "start": "Starting {start}...",
+    "not_found": "Not found {name}.",
+    "found": "Found {results} results!",
+    "download_music": "download music",
+    "download": "download",
+    "provide_url": "Please provide a link.",
+    "provide_name_is_save": "Please provide a model name to save.",
+    "not_support_url": "Your model link is not supported.",
+    "error_occurred": "An error occurred: {e}.",
+    "not_model": "The file you uploaded is not a model file!",
+    "unable_analyze_model": "Unable to analyze the model!",
+    "credits": "\n**This project is brewed by [Phạm Huỳnh Anh]({author})**\n\n**The project is built based on several major projects such as:**\n\n**Conversion, Processing, Extraction, Training, Model Reading, Model Fusion, Training Model, Model Repository...: [Applio]({applio}) by the [AI Hispano]({ai_hispano}) group**\n\n**Extraction Methods, Information Display, Logging, Training Models...: [Retrieval-based-Voice-Conversion-WebUI]({rvc_webui}) by author [RVC BOSS]({rvc_boss})**\n\n**Music Separation Models MDX-Net and Demucs: [Python-audio-separator]({python_audio_separator}) by author [Andrew Beveridge]({andrew_beveridge})**\n",
+    "download_pretrain": "Downloading pre-trained model...",
+    "provide_pretrain": "Please provide a pre-trained model link {dg}.",
+    "provide_hubert": "Please provide a link to the embedding model.",
+    "sr_not_same": "The sample rates of the two models are not the same.",
+    "architectures_not_same": "Cannot merge models. The architectures are not the same.",
+    "fushion_model": "model fusion",
+    "model_fushion_info": "The model {name} is fused from {pth_1} and {pth_2} with a ratio of {ratio}.",
+    "not_found_create_time": "Creation time not found.",
+    "format_not_valid": "Invalid format.",
+    "read_info": "Models trained on different applications may produce different information or may not be readable!",
+    "epoch": "epoch.",
+    "step": "step",
+    "sr": "sample rate.",
+    "f0": "pitch training",
+    "version": "version.",
+    "not_f0": "Pitch training not performed",
+    "trained_f0": "Pitch training performed",
+    "model_info": "Model Name: {model_name}\n\n Model Creator: {model_author}\n\nEpoch: {epochs}\n\nSteps: {steps}\n\nVersion: {version}\n\nSample Rate: {sr}\n\nPitch Training: {pitch_guidance}\n\nHash (ID): {model_hash}\n\nCreation Time: {creation_date_str}\n",
+    "input_not_valid": "Please provide valid input!",
+    "output_not_valid": "Please provide valid output!",
+    "apply_effect": "apply effect",
+    "enter_the_text": "Please enter the text to read!",
+    "choose_voice": "Please choose a voice!",
+    "convert": "Converting {name}...",
+    "separator_music": "music separation",
+    "notfound": "Not found",
+    "turn_on_use_audio": "Please enable using separated audio to proceed",
+    "turn_off_convert_backup": "Disable backup voice conversion to use the original voice",
+    "turn_off_merge_backup": "Disable merging backup voice to use the original voice",
+    "not_found_original_vocal": "Original vocal not found!",
+    "convert_vocal": "Converting voice...",
+    "convert_success": "Voice conversion completed!",
+    "convert_backup": "Converting backup voice...",
+    "convert_backup_success": "Backup voice conversion completed!",
+    "merge_backup": "Merging main voice with backup voice...",
+    "merge_success": "Merge completed.",
+    "is_folder": "Input is a folder: Converting all audio files in the folder...",
+    "not_found_in_folder": "No audio files found in the folder!",
+    "batch_convert": "Batch conversion in progress...",
+    "batch_convert_success": "Batch conversion successful!",
+    "create": "create",
+    "provide_name": "Please provide a model name.",
+    "not_found_data": "Data not found",
+    "not_found_data_preprocess": "Processed audio data not found, please reprocess.",
+    "not_found_data_extract": "Extracted audio data not found, please re-extract.",
+    "provide_pretrained": "Please provide pre-trained {dg}.",
+    "download_pretrained": "Download pre-trained {dg}{rvc_version} original",
+    "not_found_pretrain": "Pre-trained {dg} not found",
+    "not_use_pretrain": "No pre-trained model will be used",
+    "training": "training",
+    "display_title": "<h1> 🎵 Voice model conversion and training interface created by Anh 🎵 <h1>",
+    "rick_roll": "Click here if you want to be Rick Rolled :) ---> [RickRoll]({rickroll})",
+    "terms_of_use": "**Please do not use the project for any unethical, illegal, or harmful purposes to individuals or organizations...**",
+    "exemption": "**In cases where users do not comply with the terms or violate them, I will not be responsible for any claims, damages, or liabilities, whether in contract, negligence, or other causes arising from, outside of, or related to the software, its use, or other transactions associated with it.**",
+    "separator_tab": "Music Separation",
+    "4_part": "A simple music separation system can separate into 4 parts: Music, vocals, main vocals, backup vocals",
+    "clear_audio": "Clean audio",
+    "separator_backing": "Separate backup vocals",
+    "denoise_backing": "Denoise backup vocals",
+    "denoise_mdx": "Denoise MDX separation",
+    "use_mdx": "Use MDX",
+    "dereveb_audio": "Remove vocal reverb",
+    "dereveb_backing": "Remove backup reverb",
+    "denoise_dereveb": "Denoise reverb",
+    "separator_model": "Music separation model",
+    "separator_backing_model": "Backup separation model",
+    "shift": "Number of predictions",
+    "shift_info": "Higher is better quality but slower and uses more resources",
+    "segments_size": "Segment Size",
+    "segments_size_info": "Higher is better quality but uses more resources",
+    "batch_size": "Batch size",
+    "batch_size_info": "Number of samples processed simultaneously in one training cycle. Higher can cause memory overflow",
+    "mdx_batch_size_info": "Number of samples processed at a time. Batch processing optimizes calculations. Large batches can cause memory overflow; small batches reduce resource efficiency",
+    "overlap": "Overlap",
+    "overlap_info": "Overlap amount between prediction windows",
+    "export_format": "Audio format",
+    "export_info": "The audio format to export the audio file in",
+    "output_separator": "Separated audio",
+    "hop_length_info": "Indicates the time interval for shifting the analysis window on the audio signal when performing transformations. Smaller values increase detail but require more computational resources",
+    "drop_audio": "Drop audio here",
+    "drop_text": "Drop text file here",
+    "use_url": "Use YouTube link",
+    "url_audio": "Link to audio",
+    "downloads": "Downloads",
+    "clean_strength": "Audio cleaning strength",
+    "clean_strength_info": "Strength of the audio cleaner for filtering vocals during export",
+    "input_output": "Audio input, output",
+    "audio_path": "Input audio path",
+    "refesh": "Refresh",
+    "output_folder": "Output audio folder path",
+    "output_folder_info": "Enter the folder path where the audio will be exported",
+    "input_audio": "Audio input",
+    "instruments": "Instruments",
+    "original_vocal": "Original vocal",
+    "main_vocal": "Main vocal",
+    "backing_vocal": "Backup vocal",
+    "convert_audio": "Convert Audio",
+    "convert_info": "Convert audio using a trained voice model",
+    "upscale_audio": "Upscale quality",
+    "autotune": "Auto-tune",
+    "use_audio": "Use separated audio",
+    "convert_original": "Convert original voice",
+    "convert_backing": "Convert backup voice",
+    "not_merge_backing": "Do not merge backup voice",
+    "merge_instruments": "Merge instruments",
+    "pitch": "Pitch",
+    "pitch_info": "Recommendation: set to 12 to change male voice to female and vice versa",
+    "model_accordion": "Model and index",
+    "model_name": "Model file",
+    "index_path": "Index file",
+    "index_strength": "Index strength",
+    "index_strength_info": "Higher values increase strength. However, lower values may reduce artificial effects in the audio",
+    "output_path": "Audio output path",
+    "output_path_info": "Enter the output path (leave it as .wav format; it will auto-correct during conversion)",
+    "setting": "General settings",
+    "f0_method": "Extraction method",
+    "f0_method_info": "Method used for data extraction",
+    "f0_method_hybrid": "HYBRID extraction method",
+    "f0_method_hybrid_info": "Combines two different extraction types",
+    "hubert_model": "Embedding model",
+    "hubert_info": "Pre-trained model to assist embedding",
+    "modelname": "Model name",
+    "modelname_info": "If you have your own model, just upload it and input the name here",
+    "split_audio": "Split audio",
+    "split_audio_info": "Splitting audio into smaller parts for conversion can speed up the process",
+    "batch_process": "Batch processing",
+    "batch_process_info": "Batch processing can reduce training time but may overload resources",
+    "autotune_rate": "Auto-tune level",
+    "autotune_rate_info": "Level of auto-tuning adjustment",
+    "resample": "Resample",
+    "resample_info": "Resample post-processing to the final sample rate; 0 means no resampling",
+    "filter_radius": "Median filter",
+    "filter_radius_info": "If greater than three, median filtering is applied. The value represents the filter radius and can reduce breathiness or noise.",
+    "volume_envelope": "Volume envelope",
+    "volume_envelope_info": "Use the input volume envelope to replace or mix with the output volume envelope. The closer to 1, the more the output envelope is used",
+    "protect": "Consonant protection",
+    "protect_info": "Protect distinct consonants and breathing sounds to prevent audio tearing and other artifacts. Increasing this value provides comprehensive protection. Reducing it may reduce protection but also minimize indexing effects",
+    "output_convert": "Converted audio",
+    "main_convert": "Convert main voice",
+    "main_or_backing": "Main voice + Backup voice",
+    "voice_or_instruments": "Voice + Instruments",
+    "convert_text": "Convert Text",
+    "convert_text_markdown": "## Convert Text to Speech",
+    "convert_text_markdown_2": "Convert text to speech and read aloud using the trained voice model",
+    "input_txt": "Input data from a text file (.txt)",
+    "text_to_speech": "Text to read",
+    "voice_speed": "Reading speed",
+    "voice_speed_info": "Speed of the voice",
+    "tts_1": "1. Convert Text",
+    "tts_2": "2. Convert Speech",
+    "voice": "Voices by country",
+    "output_tts": "Output speech path",
+    "output_tts_convert": "Converted speech output path",
+    "tts_output": "Enter the output path",
+    "output_tts_markdown": "Unconverted and converted audio",
+    "output_text_to_speech": "Generated speech from text-to-speech conversion",
+    "output_file_tts_convert": "Speech converted using the model",
+    "output_audio": "Audio output",
+    "provide_output": "Enter the output path",
+    "audio_effects": "Audio Effects",
+    "apply_audio_effects": "## Add Additional Audio Effects",
+    "audio_effects_edit": "Edit and add effects to audio",
+    "reverb": "Reverb effect",
+    "chorus": "Chorus effect",
+    "delay": "Delay effect",
+    "more_option": "Additional options",
+    "phaser": "Phaser effect",
+    "compressor": "Compressor effect",
+    "apply": "Apply",
+    "reverb_freeze": "Freeze mode",
+    "reverb_freeze_info": "Create a continuous echo effect when this mode is enabled",
+    "room_size": "Room size",
+    "room_size_info": "Adjust the room space to create reverberation",
+    "damping": "Damping",
+    "damping_info": "Adjust the level of absorption to control the amount of reverberation",
+    "wet_level": "Reverb signal level",
+    "wet_level_info": "Adjust the level of the reverb signal effect",
+    "dry_level": "Original signal level",
+    "dry_level_info": "Adjust the level of the signal without effects",
+    "width": "Audio width",
+    "width_info": "Adjust the width of the audio space",
+    "chorus_depth": "Chorus depth",
+    "chorus_depth_info": "Adjust the intensity of the chorus to create a wider sound",
+    "chorus_rate_hz": "Frequency",
+    "chorus_rate_hz_info": "Adjust the oscillation speed of the chorus effect",
+    "chorus_mix": "Mix signals",
+    "chorus_mix_info": "Adjust the mix level between the original and the processed signal",
+    "chorus_centre_delay_ms": "Center delay (ms)",
+    "chorus_centre_delay_ms_info": "The delay time between stereo channels to create the chorus effect",
+    "chorus_feedback": "Feedback",
+    "chorus_feedback_info": "Adjust the amount of the effect signal fed back into the original signal",
+    "delay_seconds": "Delay time",
+    "delay_seconds_info": "Adjust the delay time between the original and the processed signal",
+    "delay_feedback": "Delay feedback",
+    "delay_feedback_info": "Adjust the amount of feedback signal, creating a repeating effect",
+    "delay_mix": "Delay signal mix",
+    "delay_mix_info": "Adjust the mix level between the original and delayed signal",
+    "fade": "Fade effect",
+    "bass_or_treble": "Bass and treble",
+    "limiter": "Threshold limiter",
+    "distortion": "Distortion effect",
+    "gain": "Audio gain",
+    "bitcrush": "Bit reduction effect",
+    "clipping": "Clipping effect",
+    "fade_in": "Fade-in effect (ms)",
+    "fade_in_info": "Time for the audio to gradually increase from 0 to normal level",
+    "fade_out": "Fade-out effect (ms)",
+    "fade_out_info": "Time for the audio to gradually decrease to 0",
+    "bass_boost": "Bass boost level (dB)",
+    "bass_boost_info": "Boost level for bass frequencies",
+    "bass_frequency": "Low-pass filter cutoff frequency (Hz)",
+    "bass_frequency_info": "Frequency below which the sound will start to reduce. Lower frequencies make bass clearer",
+    "treble_boost": "Treble boost level (dB)",
+    "treble_boost_info": "Boost level for high frequencies",
+    "treble_frequency": "High-pass filter cutoff frequency (Hz)",
+    "treble_frequency_info": "Sounds below this frequency will be filtered out. Higher frequencies retain more treble",
+    "limiter_threashold_db": "Limiter threshold",
+    "limiter_threashold_db_info": "Limit the maximum audio level to prevent it from exceeding the threshold",
+    "limiter_release_ms": "Release time",
+    "limiter_release_ms_info": "Time for the audio to return after being limited",
+    "distortion_info": "Adjust the level of distortion to create a noisy effect",
+    "gain_info": "Adjust the volume level of the signal",
+    "clipping_threashold_db": "Clipping threshold",
+    "clipping_threashold_db_info": "Trim signals exceeding the threshold, creating a distorted sound",
+    "bitcrush_bit_depth": "Bit depth",
+    "bitcrush_bit_depth_info": "Reduce audio quality by decreasing bit depth, creating a distorted effect",
+    "phaser_depth": "Phaser depth",
+    "phaser_depth_info": "Adjust the depth of the effect, impacting its intensity",
+    "phaser_rate_hz": "Frequency",
+    "phaser_rate_hz_info": "Adjust the frequency of the phaser effect",
+    "phaser_mix": "Mix signal",
+    "phaser_mix_info": "Adjust the mix level between the original and processed signals",
+    "phaser_centre_frequency_hz": "Center frequency",
+    "phaser_centre_frequency_hz_info": "The center frequency of the phaser effect, affecting the adjusted frequencies",
+    "phaser_feedback": "Feedback",
+    "phaser_feedback_info": "Adjust the feedback level of the effect, creating a stronger or lighter phaser feel",
+    "compressor_threashold_db": "Compressor threshold",
+    "compressor_threashold_db_info": "The threshold level above which the audio will be compressed",
+    "compressor_ratio": "Compression ratio",
+    "compressor_ratio_info": "Adjust the level of audio compression when exceeding the threshold",
+    "compressor_attack_ms": "Attack time (ms)",
+    "compressor_attack_ms_info": "Time for compression to start taking effect after the audio exceeds the threshold",
+    "compressor_release_ms": "Release time",
+    "compressor_release_ms_info": "Time for the audio to return to normal after being compressed",
+    "create_dataset_url": "Link to audio (use commas for multiple links)",
+    "createdataset": "Create dataset",
+    "create_dataset_markdown": "## Create Training Data from YouTube",
+    "create_dataset_markdown_2": "Process and create training datasets using YouTube links",
+    "denoise": "Model denoise",
+    "skip": "Skip seconds",
+    "model_ver": "Voice separation version",
+    "model_ver_info": "The model version for separating vocals",
+    "create_dataset_info": "Dataset creation information",
+    "output_data": "Dataset output",
+    "output_data_info": "Output data after creation",
+    "skip_start": "Skip beginning",
+    "skip_start_info": "Skip the initial seconds of the audio; use commas for multiple audios",
+    "skip_end": "Skip end",
+    "skip_end_info": "Skip the final seconds of the audio; use commas for multiple audios",
+    "training_model": "Train Model",
+    "training_markdown": "Train and build a voice model with a set of voice data",
+    "training_model_name": "Name of the model during training (avoid special characters or spaces)",
+    "sample_rate": "Sample rate",
+    "sample_rate_info": "Sample rate of the model",
+    "training_version": "Model version",
+    "training_version_info": "Version of the model during training",
+    "training_pitch": "Pitch training",
+    "training_pitch_info": "Train the pitch for the model",
+    "upload_dataset": "Upload training data",
+    "preprocess_split": "Should be disabled if data has already been processed",
+    "preprocess_effect": "Processing effects",
+    "preprocess_effect_info": "Should be disabled if data has already been processed",
+    "clear_dataset": "Clean dataset",
+    "clear_dataset_info": "Clean data segments",
+    "preprocess_info": "Preprocessing information",
+    "preprocess_button": "1. Process data",
+    "extract_button": "2. Extract data",
+    "extract_info": "Data extraction information",
+    "total_epoch": "Total epochs",
+    "total_epoch_info": "Total training epochs",
+    "save_epoch": "Save frequency",
+    "save_epoch_info": "Frequency of saving the model during training to allow retraining",
+    "create_index": "Create index",
+    "index_algorithm": "Index algorithm",
+    "index_algorithm_info": "Algorithm for creating the index",
+    "custom_dataset": "Custom folder",
+    "custom_dataset_info": "Custom folder for training data",
+    "overtraining_detector": "Overtraining detector",
+    "overtraining_detector_info": "Check for overtraining during model training",
+    "sync_graph": "Synchronize graph",
+    "sync_graph_info": "Only enable if you need to retrain the model from scratch.",
+    "cache_in_gpu": "Cache model in GPU",
+    "cache_in_gpu_info": "Store the model in GPU cache memory",
+    "dataset_folder": "Folder containing data",
+    "threshold": "Overtraining threshold",
+    "setting_cpu_gpu": "CPU/GPU settings",
+    "gpu_number": "Number of GPUs used",
+    "gpu_number_info": "Number of GPUs used during training",
+    "save_only_latest": "Save only the latest",
+    "save_only_latest_info": "Save only the latest D and G models",
+    "save_every_weights": "Save all models",
+    "save_every_weights_info": "Save all models after each epoch",
+    "gpu_info": "GPU information",
+    "gpu_info_2": "Information about the GPU used during training",
+    "cpu_core": "Number of CPU cores available",
+    "cpu_core_info": "Number of CPU cores used during training",
+    "not_use_pretrain_2": "Do not use pretraining",
+    "not_use_pretrain_info": "Do not use pre-trained models",
+    "custom_pretrain": "Custom pretraining",
+    "custom_pretrain_info": "Customize pre-training settings",
+    "pretrain_file": "Pre-trained model file {dg}",
+    "train_info": "Training information",
+    "export_model": "5. Export Model",
+    "zip_model": "2. Compress model",
+    "output_zip": "Output file after compression",
+    "refesh_pretrain": "Refresh pre-trained model",
+    "model_path": "Model path",
+    "model_ratio": "Model ratio",
+    "model_ratio_info": "Adjusting towards one side will make the model more like that side",
+    "output_model_path": "Model output path",
+    "fushion": "Model Fusion",
+    "fushion_markdown": "## Merge Two Models",
+    "fushion_markdown_2": "Combine two voice models into a single model",
+    "read_model": "Read Information",
+    "read_model_markdown": "## Read Model Information",
+    "read_model_markdown_2": "Retrieve recorded information within the model",
+    "drop_model": "Drop model here",
+    "readmodel": "Read model",
+    "model_path_info": "Enter the path to the model file",
+    "modelinfo": "Model Information",
+    "download_markdown": "## Download Model",
+    "download_markdown_2": "Download voice models, pre-trained models, and embedding models",
+    "model_download": "Download voice model",
+    "model_url": "Link to the model",
+    "30s": "Please wait about 30 seconds. The system will restart automatically!",
+    "fp_select": "Setting precision...",
+    "fp_select_2": "Precision set to: {fp}",
+    "model_download_select": "Choose a model download method",
+    "model_warehouse": "Model repository",
+    "get_model": "Retrieve model",
+    "name_to_search": "Name to search",
+    "search_2": "Search",
+    "select_download_model": "Choose a searched model (Click to select)",
+    "download_pretrainec": "Download pre-trained model",
+    "only_huggingface": "Supports only huggingface.co",
+    "pretrained_url": "Pre-trained model link {dg}",
+    "select_pretrain": "Choose pre-trained model",
+    "select_pretrain_info": "Choose a pre-trained model to download",
+    "pretrain_sr": "Model sample rate",
+    "drop_pretrain": "Drop pre-trained model {dg} here",
+    "hubert_download": "Download embedding model",
+    "hubert_url": "Link to embedding model",
+    "drop_hubert": "Drop embedding model here",
+    "source": "Origin",
+    "source_info": "## Project Origin and Author",
+    "settings": "Settings",
+    "settings_markdown": "## Additional Settings",
+    "settings_markdown_2": "Customize additional features of the project",
+    "lang": "Language",
+    "lang_restart": "The display language in the project (When changing the language, the system will automatically restart after 30 seconds to update)",
+    "change_lang": "Change Language",
+    "fp_train": "Training precision",
+    "fp_info": "Use fp32 for higher precision but requires more memory and strong computation. Use fp16 to reduce memory usage and increase processing speed.",
+    "fp_button": "Change Precision",
+    "theme": "Theme",
+    "theme_restart": "Theme type displayed in the interface (When changing the theme, the system will automatically restart after 30 seconds to update)",
+    "theme_button": "Change Theme",
+    "change_light_dark": "Switch Light/Dark Mode",
+    "tensorboard_url": "Graph URL",
+    "errors_loading_audio": "Error loading audio: {e}",
+    "apply_error": "An error occurred while applying effects: {e}",
+    "indexpath": "Index path",
+    "skip_file": "Part {i} skipped because it is too short: {chunk}ms",
+    "split_total": "Total parts split",
+    "process_audio_error": "An error occurred while processing the audio",
+    "merge_error": "An error occurred while merging audio",
+    "not_found_convert_file": "Processed file not found",
+    "convert_batch": "Batch conversion...",
+    "found_audio": "Found {audio_files} audio files for conversion.",
+    "not_found_audio": "No audio files found!",
+    "error_convert": "An error occurred during audio conversion: {e}",
+    "error_convert_batch": "An error occurred during the conversion of audio segments: {e}",
+    "error_convert_batch_2": "An error occurred during batch audio conversion: {e}",
+    "convert_batch_success": "Batch conversion completed successfully in {elapsed_time} seconds. {output_path}",
+    "convert_audio_success": "File {input_path} converted successfully in {elapsed_time} seconds. {output_path}",
+    "hybrid_methods": "Estimating f0 pitch using methods {methods}",
+    "method_not_valid": "Invalid method",
+    "read_faiss_index_error": "An error occurred while reading the FAISS index: {e}",
+    "read_model_error": "Failed to load model: {e}",
+    "denoise_error": "An error occurred while denoising: {e}",
+    "change_format": "Converting audio to format {output_format}...",
+    "change_format_error": "An error occurred while converting audio format: {e}",
+    "starting_download": "Starting download",
+    "version_not_valid": "Invalid vocal separation version",
+    "create_dataset_value_not_valid": "Enable music separation option to use reverb removal options",
+    "skip<audio": "Cannot skip as skip time is less than audio file length",
+    "skip>audio": "Cannot skip as skip time is greater than audio file length",
+    "=<0": "Skip time is less than or equal to 0 and has been skipped",
+    "skip_warning": "Skip duration ({seconds} seconds) exceeds audio length ({total_duration} seconds). Skipping.",
+    "download_success": "Download completed successfully",
+    "create_dataset_error": "An error occurred while creating the training dataset",
+    "create_dataset_success": "Training dataset creation completed in {elapsed_time} seconds",
+    "skip_start_audio": "Successfully skipped start of audio: {input_file}",
+    "skip_end_audio": "Successfully skipped end of audio: {input_file}",
+    "merge_audio": "Merged all parts containing audio",
+    "separator_process": "Separating vocals: {input}...",
+    "not_found_main_vocal": "Main vocal not found!",
+    "not_found_backing_vocal": "Backup vocal not found!",
+    "not_found_instruments": "Instruments not found",
+    "merge_instruments_process": "Merging vocals with instruments...",
+    "dereverb": "Removing vocal reverb",
+    "dereverb_success": "Successfully removed vocal reverb",
+    "save_index": "Index file saved",
+    "create_index_error": "An error occurred while creating the index",
+    "sr_not_16000": "Sample rate must be 16000",
+    "gpu_not_valid": "Invalid GPU index. Switching to CPU.",
+    "extract_file_error": "An error occurred while extracting the file",
+    "extract_f0_method": "Starting pitch extraction using {num_processes} cores with method {f0_method}...",
+    "extract_f0": "Pitch Extraction",
+    "extract_f0_success": "Pitch extraction completed in {elapsed_time} seconds.",
+    "NaN": "contains NaN values and will be ignored.",
+    "start_extract_hubert": "Starting Hubert embedding extraction...",
+    "not_found_audio_file": "Audio file not found. Please ensure you provided the correct audio.",
+    "extract_hubert": "Extract Hubert embeddings",
+    "process_error": "An error occurred during processing",
+    "extract_hubert_success": "Hubert embedding extraction completed in {elapsed_time} seconds.",
+    "export_process": "Model path",
+    "extract_error": "An error occurred during data extraction",
+    "extract_success": "Data extraction successful",
+    "min_length>=min_interval>=hop_size": "min_length must be greater than or equal to min_interval and hop_size",
+    "max_sil_kept>=hop_size": "max_sil_kept must be greater than or equal to hop_size",
+    "start_preprocess": "Starting data preprocessing with {num_processes} cores...",
+    "not_integer": "Voice ID folder must be an integer; instead got",
+    "preprocess": "Preprocessing Data",
+    "preprocess_success": "Preprocessing completed in {elapsed_time} seconds.",
+    "preprocess_model_success": "Preprocessing data for the model completed successfully",
+    "turn_on_dereverb": "Reverb removal for backup vocals requires enabling reverb removal",
+    "turn_on_separator_backing": "Backup vocal separation requires enabling vocal separation",
+    "demucs_model": "Demucs model",
+    "backing_model_ver": "Backup vocal separation model version",
+    "mdx_model": "MDX model",
+    "clean_audio_success": "Audio cleaned successfully!",
+    "separator_error": "An error occurred during music separation",
+    "separator_success": "Music separation completed in {elapsed_time} seconds",
+    "separator_process_2": "Processing music separation",
+    "separator_success_2": "Music separation successful!",
+    "separator_process_backing": "Processing backup vocal separation",
+    "separator_process_backing_success": "Backup vocal separation successful!",
+    "process_original": "Processing original vocal reverb removal...",
+    "process_original_success": "Original vocal reverb removal successful!",
+    "process_main": "Processing main vocal reverb removal...",
+    "process_main_success": "Main vocal reverb removal successful!",
+    "process_backing": "Processing backup vocal reverb removal...",
+    "process_backing_success": "Backup vocal reverb removal successful!",
+    "save_every_epoch": "Save model after: {save_every_epoch} epochs",
+    "total_e": "Total training epochs: {total_epoch} epochs",
+    "dorg": "Pre-trained G: {pretrainG} | Pre-trained D: {pretrainD}",
+    "training_f0": "Pitch training",
+    "not_gpu": "No GPU detected, reverting to CPU (not recommended)",
+    "sync": "Graph synchronization is now enabled! When synchronization is active, the model will undergo a single training phase. After graphs are synchronized, training will resume with previously specified epochs.",
+    "sync_success": "Graph synchronization successful!",
+    "not_found_checkpoint": "Checkpoint file not found: {checkpoint_path}",
+    "save_checkpoint": "Reloaded checkpoint '{checkpoint_path}' (epoch {checkpoint_dict})",
+    "save_model": "Saved model '{checkpoint_path}' (epoch {iteration})",
+    "sid_error": "Error converting speaker ID '{sid}' to integer. Exception: {e}",
+    "sr_does_not_match": "{sample_rate} Sample rate does not match target {sample_rate2} Sample rate",
+    "spec_error": "An error occurred while retrieving specifications from {spec_filename}: {e}",
+    "time_or_speed_training": "time={current_time} | training speed={elapsed_time_str}",
+    "savemodel": "Saved model '{model_dir}' (epoch {epoch} and step {step})",
+    "model_author": "Credit model to {model_author}",
+    "unregistered": "Model unregistered",
+    "not_author": "Model not credited",
+    "training_author": "Model creator name",
+    "training_author_info": "To credit the model, enter your name here",
+    "extract_model_error": "An error occurred while extracting the model",
+    "start_training": "Starting training",
+    "import_pretrain": "Loaded pre-trained model ({dg}) '{pretrain}'",
+    "not_using_pretrain": "No pre-trained model ({dg}) will be used",
+    "training_warning": "WARNING: Generated loss is lower than the lower threshold loss for the next epoch.",
+    "overtraining_find": "Overtraining detected at epoch {epoch} with smoothed generator loss {smoothed_value_gen} and smoothed discriminator loss {smoothed_value_disc}",
+    "best_epoch": "New best epoch {epoch} with smoothed generator loss {smoothed_value_gen} and smoothed discriminator loss {smoothed_value_disc}",
+    "success_training": "Training completed with {epoch} epochs, {global_step} steps, and {loss_gen_all} total generator loss.",
+    "training_info": "Lowest generator loss: {lowest_value_rounded} at epoch {lowest_value_epoch}, step {lowest_value_step}",
+    "model_training_info": "{model_name} | epoch={epoch} | step={global_step} | {epoch_recorder} | lowest value={lowest_value_rounded} (epoch {lowest_value_epoch} and step {lowest_value_step}) | remaining epochs for overtraining: g/total: {remaining_epochs_gen} d/total: {remaining_epochs_disc} | smoothed generator loss={smoothed_value_gen} | smoothed discriminator loss={smoothed_value_disc}",
+    "model_training_info_2": "{model_name} | epoch={epoch} | step={global_step} | {epoch_recorder} | lowest value={lowest_value_rounded} (epoch {lowest_value_epoch} and step {lowest_value_step})",
+    "model_training_info_3": "{model_name} | epoch={epoch} | step={global_step} | {epoch_recorder}",
+    "training_error": "An error occurred while training the model:",
+    "separator_info": "Initializing with output path: {output_dir}, output format: {output_format}",
+    "output_dir_is_none": "Output folder not specified. Using current working directory.",
+    ">0or=1": "Normalization threshold must be greater than 0 and less than or equal to 1.",
+    "output_single": "Single root output requested; only one file ({output_single_stem}) will be written",
+    "step2": "The second step will be reversed using spectrogram instead of waveform. This may improve quality but is slightly slower.",
+    "other_than_zero": "Sample rate set to {sample_rate} but must be a nonzero integer.",
+    "too_high": "Sample rate set to {sample_rate}. Enter a less ambitious value.",
+    "sr_not_valid": "Sample rate must be a nonzero integer. Please provide a valid integer.",
+    "name_ver": "Version {name}",
+    "os": "Operating System",
+    "platform_info": "System: {system_info} Name: {node} Release: {release} Machine: {machine} Processor: {processor}",
+    "install_ffmpeg": "FFmpeg installed",
+    "none_ffmpeg": "FFmpeg is not installed. Please install FFmpeg to use this package.",
+    "install_onnx": "ONNX Runtime package {pu} installed with version",
+    "running_in_cpu": "Unable to configure hardware acceleration, running in CPU mode",
+    "running_in_cuda": "CUDA available in Torch, setting Torch device to CUDA",
+    "onnx_have": "ONNXruntime available {have}, enabling acceleration",
+    "onnx_not_have": "{have} not available in ONNXruntime; acceleration will NOT be enabled",
+    "python_not_install": "Python package: {package_name} is not installed",
+    "hash": "Calculating hash for model file {model_path}",
+    "ioerror": "IOError while seeking -10 MB or reading model file to compute hash: {e}",
+    "cancel_download": "File already exists at {output_path}, skipping download",
+    "download_model": "Downloading file from {url} to {output_path} with a timeout of 300 seconds",
+    "download_error": "Failed to download file from {url}, response code: {status_code}",
+    "vip_model": "Model: '{model_friendly_name}' is a premium model intended by Anjok07 only for paid subscriber access.",
+    "vip_print": "Hey there, if you haven't subscribed, please consider supporting UVR's developer, Anjok07, by subscribing here: https://patreon.com/uvr",
+    "search_model": "Searching for model {model_filename} in the list of supported models in the group",
+    "load_download_json": "Downloaded model list loaded",
+    "single_model": "Identified single model file: {model_friendly_name}",
+    "not_found_model": "Model not found in the UVR repository, attempting to download from the audio model separation repository...",
+    "single_model_path": "Returning path for single model file: {model_path}",
+    "find_model": "Input file name {model_filename} found in multi-file model: {model_friendly_name}",
+    "find_models": "Identified multi-file model: {model_friendly_name}, iterating through files to download",
+    "find_path": "Attempting to determine download PATH for config pair",
+    "not_found_model_warehouse": "Model not found in the UVR repository, attempting to download from the audio model separation repository...",
+    "yaml_warning": "The model name you specified, {model_filename}, is actually a model config file rather than a model file.",
+    "yaml_warning_2": "We found a model matching this config file: {config_key}, so we’ll use that model file for this run.",
+    "yaml_warning_3": "To avoid confusing/inconsistent behavior in the future, specify the actual model file name instead.",
+    "yaml_debug": "Config YAML model file not found in UVR repository, attempting to download from the audio model separation repository...",
+    "download_model_friendly": "All files downloaded for model {model_friendly_name}, returning original path {model_path}",
+    "not_found_model_2": "Model file {model_filename} not found in the supported files",
+    "load_yaml": "Loading model data from YAML at path {model_data_yaml_filepath}",
+    "load_yaml_2": "Model data loaded from YAML file: {model_data}",
+    "hash_md5": "Computing MD5 hash for model file to identify model parameters from UVR data...",
+    "model_hash": "Model {model_path} has hash {model_hash}",
+    "mdx_data": "MDX model data path set to {mdx_model_data_path}",
+    "load_mdx": "Loading MDX model parameters from UVR model data file...",
+    "model_not_support": "Unsupported model file: no parameters found for MD5 hash {model_hash} in UVR model data for MDX vault.",
+    "uvr_json": "Model data loaded from UVR JSON with hash {model_hash}: {model_data}",
+    "loading_model": "Loading model {model_filename}...",
+    "download_model_friendly_2": "Downloaded model, friendly name: {model_friendly_name}, Model path: {model_path}",
+    "model_type_not_support": "Unsupported model type: {model_type}",
+    "demucs_not_support_python<3.10": "Demucs models require Python version 3.10 or higher.",
+    "import_module": "Importing module for model type",
+    "initialization": "Initializing separator class for model type",
+    "loading_model_success": "Model loading completed.",
+    "loading_model_duration": "Model loading duration",
+    "starting_separator": "Starting separation process for audio file path",
+    "normalization": "Normalization threshold set to {normalization_threshold}, waveform will be scaled down to this maximum amplitude to prevent clipping.",
+    "loading_separator_model": "Downloading model {model_filename}...",
+    "separator_success_3": "Separation process completed.",
+    "separator_duration": "Separation duration",
+    "downloading_model": "Downloaded model, type: {model_type}, friendly name: {model_friendly_name}, Model path: {model_path}, Model data: {model_data_dict_size} items",
+    "not_found_model_signature": "Pre-trained model with signature {sig} not found.",
+    "repo_must_be_folder": "{repo} must exist and be a directory.",
+    "demucs_info": "Demucs parameters: Segment size = {segment_size}, Segment size active = {segments_enabled}",
+    "demucs_info_2": "Demucs parameters: Number of predictions = {shifts}, Overlap = {overlap}",
+    "start_demucs": "Demucs Separator initialization completed",
+    "start_separator": "Starting separation process...",
+    "prepare_mix": "Preparing mixture...",
+    "demix": "Mixture prepared for demixing. Shape: {shape}",
+    "cancel_mix": "Loading model for demixing...",
+    "model_review": "Model loaded and set to evaluation mode.",
+    "del_gpu_cache_after_demix": "Cleared model and GPU cache after demixing.",
+    "process_output_file": "Processing output file...",
+    "source_length": "Processing source array, source length is {source_length}",
+    "process_ver": "Processing source version...",
+    "set_map": "Set source map to {part} parts...",
+    "process_all_part": "Processing for all root parts...",
+    "skip_part": "Skipping root part {stem_name} as out_single_stem is set to {output_single_stem}...",
+    "starting_demix_demucs": "Starting the demix process in demix_demucs...",
+    "model_infer": "Running model inference...",
+    "del_all_but_one": "Duplicate pre-trained models detected for signature {xp_sig}. Please delete all but one.",
+    "name_not_pretrained": "{name} is not a pre-trained model or a model bundle.",
+    "invalid_checksum": "Invalid checksum for file {path}, expected {checksum} but got {actual_checksum}",
+    "mdx_info": "MDX parameters: Batch size = {batch_size}, Segment size = {segment_size}",
+    "mdx_info_2": "MDX parameters: Overlap = {overlap}, Hop_length = {hop_length}, Denoising enabled = {enable_denoise}",
+    "mdx_info_3": "MDX parameters",
+    "load_model_onnx": "Loading ONNX model for inference...",
+    "load_model_onnx_success": "Successfully loaded model using ONNXruntime inference session.",
+    "onnx_to_pytorch": "Model converted from ONNX to PyTorch due to mismatched segment size with dim_t, processing may be slower.",
+    "stft": "Inverse STFT applied. Returning result with shape",
+    "no_denoise": "Model running on spectrum without denoising.",
+    "mix": "Preparing mix for input audio file {audio_file_path}...",
+    "normalization_demix": "Normalizing mix prior to demixing...",
+    "mix_success": "Mix preparation completed.",
+    "primary_source": "Normalizing primary source...",
+    "secondary_source": "Producing secondary source: Mixing in compatible mode",
+    "invert_using_spec": "Inverting secondary stem using spectrum when invert_USE_spec is set to True",
+    "invert_using_spec_2": "Inverting secondary stem by subtracting transformed stem from the initial transformed mix",
+    "enable_denoise": "Model running on both positive and negative spectrums for denoising.",
+    "is_match_mix": "is_match_mix: Predicted spectrum obtained directly from STFT output.",
+    "save_secondary_stem_output_path": "Saving secondary stem {stem_name} to {stem_output_path}...",
+    "starting_model": "Initializing model settings...",
+    "input_info": "Model input parameters",
+    "model_settings": "Model settings",
+    "initialize_mix": "Initializing mix with is_ckpt = {is_ckpt}. Initial mix shape: {shape}",
+    "!=2": "Expected 2-channel audio signal but got {shape} channels",
+    "process_check": "Processing in checkpoint mode...",
+    "stft_2": "STFT applied to mix. Spectrum shape: {shape}",
+    "cache": "Computed padding",
+    "shape": "Mix shape after padding: {shape}, Number of parts: {num_chunks}",
+    "process_no_check": "Processing in no-checkpoint mode...",
+    "n_sample_or_pad": "Number of samples: {n_sample}, Computed padding: {pad}",
+    "shape_2": "Mix shape after padding",
+    "process_part": "Processed part {mix_waves}: Start {i}, End {ii}",
+    "mix_waves_to_tensor": "Converted mix_waves to tensor. Tensor shape: {shape}",
+    "mix_match": "Mix mode Match; applying compensation factor.",
+    "tar_waves": "tar_waves. Shape",
+    "normalization_2": "Normalizing result by dividing it by divisor.",
+    "mix_wave": "Processing mix_wave batch",
+    "mix_or_batch": "Mix parts into batches. Number of batches",
+    "demix_is_match_mix": "Starting demix process with is_match_mix,",
+    "mix_shape": "Root mix parts stored. Shape",
+    "chunk_size_or_overlap": "Chunk size for compatible mixing: {chunk_size}, Overlap: {overlap}",
+    "chunk_size_or_overlap_standard": "Standard chunk size: {chunk_size}, Overlap: {overlap}",
+    "calc_size": "Generated size calculated",
+    "window": "Window applied to this segment.",
+    "process_part_2": "Processing segment {total}/{total_chunks}: Start {start}, End {end}",
+    "all_process_part": "Total segments to process",
+    "step_or_overlap": "Step size to process parts: {step} with overlap set to {overlap}.",
+    "mix_cache": "Mix prepared with padding. Mix shape",
+    "dims": "Cannot use sin/cos position encoding with odd dimensions (dim={dims})",
+    "activation": "activation must be relu/gelu, not {activation}",
+    "length_or_training_length": "Provided length {length} exceeds training duration {training_length}",
+    "type_not_valid": "Invalid type for",
+    "del_parameter": "Removing non-existent parameter ",
+    "info": "Common parameters: Model name = {model_name}, Model path = {model_path}",
+    "info_2": "Common parameters: Output path = {output_dir}, Output format = {output_format}",
+    "info_3": "Common parameters: Normalization threshold = {normalization_threshold}",
+    "info_4": "Common parameters: Denoising enabled = {enable_denoise}, Single stem output = {output_single_stem}",
+    "info_5": "Common parameters: Inversion using specs = {invert_using_spec}, Sample rate = {sample_rate}",
+    "info_6": "Common parameters: Primary root name = {primary_stem_name}, Secondary root name = {secondary_stem_name}",
+    "info_7": "Common parameters: Karaoke mode = {is_karaoke}, BV model = {is_bv_model}, BV model rebalancing = {bv_model_rebalance}",
+    "success_process": "Completed processing root {stem_name} and writing audio...",
+    "load_audio": "Loading audio from file",
+    "load_audio_success": "Audio loaded. Sample rate: {sr}, Audio shape: {shape}",
+    "convert_mix": "Converting provided mix array.",
+    "convert_shape": "Converted mix shape: {shape}",
+    "audio_not_valid": "Audio file {audio_path} is empty or invalid",
+    "audio_valid": "Audio file is valid and contains data.",
+    "mix_single": "Mix is mono. Converting to stereo.",
+    "convert_mix_audio": "Converted to stereo mix.",
+    "mix_success_2": "Mix preparation completed.",
+    "duration": "Audio duration is {duration_hours} hours ({duration_seconds} seconds).",
+    "write": "Using {name} to write.",
+    "write_audio": "Writing {name} with root path:",
+    "original_not_valid": "Warning: Original source array is nearly silent or empty.",
+    "shape_audio": "Audio data shape before processing",
+    "convert_data": "Data type before conversion",
+    "original_source_to_int16": "Converted original_source to int16.",
+    "shape_audio_2": "Interleaved audio data shape",
+    "create_audiosegment": "Successfully created AudioSegment.",
+    "create_audiosegment_error": "Specific error while creating AudioSegment",
+    "export_error": "Error exporting audio file",
+    "export_success": "Successfully exported audio file to",
+    "clean": "Running garbage collection...",
+    "clean_cache": "Clearing {name} cache...",
+    "del_path": "Deleting path, source, and root of input audio file...",
+    "not_success": "Process was not successful: ",
+    "resample_error": "Error during resampling",
+    "shapes": "Shapes",
+    "wav_resolution": "Resolution type",
+    "warnings": "Warning: Extremely aggressive values detected",
+    "warnings_2": "Warning: NaN or infinite values detected in wave input. Shape",
+    "process_file": "Processing file... \n",
+    "save_instruments": "Saving reverse track...",
+    "assert": "Audio files must have the same shape - Mix: {mixshape}, Inst: {instrumentalshape}",
+    "rubberband": "Rubberband CLI cannot be executed. Please ensure Rubberband-CLI is installed.",
+    "rate": "Rate must be strictly positive",
+    "gdown_warning": "You provided a Google Drive link that is not a direct file download link. Consider using the `--fuzzy` option or the following URL: https://drive.google.com/uc?id={file_id}",
+    "gdown_error": "Could not retrieve the public link for the file. You may need to change its permissions to 'Anyone with the link' or there may already be excessive access permissions.",
+    "gdown_error_2": "Failed to retrieve file path:\n\n{error}\n\nYou can still access the file via a browser:\n\n\t{url_origin}\n\nbut Gdown cannot. Please check your connection and permissions.",
+    "temps": "Several temporary files exist for continuation:",
+    "del_all_temps": "Please delete all except one to continue downloading.",
+    "continue": "Continue:",
+    "to": "To:",
+    "gdown_value_error": "A path or ID must be specified",
+    "missing_url": "URL is missing",
+    "mac_not_match": "MAC does not match",
+    "file_not_access": "File is not accessible",
+    "int_resp==-3": "Request failed, retrying"
+}

assets/languages/vi-VN.json

@@ -0,0 +1,710 @@
+{
+    "set_precision": "Ghi đè tiền xử lý và config.json để sử dụng {precision}.",
+    "set_lang": "Đã đặt ngôn ngữ hiển thị là {lang}",
+    "no_support_gpu": "Thật không may, không có GPU tương thích để hỗ trợ việc đào tạo của bạn.",
+    "text": "văn bản",
+    "upload_success": "Đã tải lên tệp {name} thành công.",
+    "download_url": "Tải từ đường dẫn liên kết",
+    "download_from_csv": "Tải từ kho mô hình csv",
+    "download_from_applio": "Tải mô hình từ Applio",
+    "upload": "Tải lên",
+    "option_not_valid": "Tùy chọn không hợp lệ!",
+    "list_model": "Danh sách mô hình",
+    "success": "Hoàn thành!",
+    "index": "chỉ mục",
+    "model": "mô hình",
+    "zip": "nén",
+    "search": "tìm kiếm",
+    "provide_file": "Vui lòng cung cấp tệp {filename} hợp lệ!",
+    "start": "Bắt đầu {start}...",
+    "not_found": "Không tìm thấy {name}",
+    "found": "Đã tìm thấy {results} kết quả!",
+    "download_music": "tải nhạc",
+    "download": "tải xuống",
+    "provide_url": "Vui lòng cung cấp đường dẫn liên kết.",
+    "provide_name_is_save": "Vui lòng cung cấp tên mô hình để lưu.",
+    "not_support_url": "Liên kết mô hình của bạn không được hỗ trợ.",
+    "error_occurred": "Đã xảy ra lỗi: {e}",
+    "not_model": "Tệp bạn vừa tải lên không phải là tệp mô hình!",
+    "unable_analyze_model": "Không phân tích được mô hình!",
+    "credits": "\n**Dự án này được nấu bởi [Phạm Huỳnh Anh]({author})**\n\n**Dự án được nấu dựa trên một số dự án chính như:**\n\n**Chuyển đổi, Xử lý, Trích xuất, Huấn luyện, Đọc mô hình, dung hợp mô hình, mô hình huấn luyện, kho mô hình...: [Applio]({applio}) của nhóm [AI Hispano]({ai_hispano})**\n\n**Phương pháp trích xuất, cách hiển thị thông tin, cách ghi nhật ký, mô hình huấn luyện...: [Retrieval-based-Voice-Conversion-WebUI]({rvc_webui}) của tác giả [RVC BOSS]({rvc_boss})**\n\n**Mô hình tách nhạc MDX-Net và Demucs: [Python-audio-separator]({python_audio_separator}) của tác giả [Andrew Beveridge]({andrew_beveridge})**\n",
+    "download_pretrain": "Tải xuống huấn luyện trước...",
+    "provide_pretrain": "Vui lòng cung cấp đường dẫn mô hình huấn luyện trước {dg}.",
+    "provide_hubert": "Vui lòng đưa đường dẫn liên kết đến mô hình nhúng",
+    "sr_not_same": "Tốc độ lấy mẫu của hai mô hình không giống nhau",
+    "architectures_not_same": "Không thể hợp nhất các mô hình. Các kiến trúc mô hình không giống nhau.",
+    "fushion_model": "dung hợp mô hình",
+    "model_fushion_info": "Mô hình được {name} được dung hợp từ {pth_1} và {pth_2} với ratio {ratio}.",
+    "not_found_create_time": "Không tìm thấy thời gian tạo.",
+    "format_not_valid": "Định dạng không hợp lệ.",
+    "read_info": "Các mô hình được huấn luyện trên các ứng dụng khác nhau có thể đem lại các thông tin khác nhau hoặc không thể đọc!",
+    "epoch": "kỷ nguyên.",
+    "step": "bước",
+    "sr": "tốc độ lấy mẫu.",
+    "f0": "huấn luyện cao độ",
+    "version": "phiên bản.",
+    "not_f0": "Không được huấn luyện cao độ",
+    "trained_f0": "Được huấn luyện cao độ",
+    "model_info": "Tên mô hình: {model_name}\n\n Người tạo mô hình: {model_author}\n\nKỷ nguyên: {epochs}\n\nSố bước: {steps}\n\nPhiên bản của mô hình: {version}\n\nTốc độ lấy mẫu: {sr}\n\nHuấn luyện cao độ: {pitch_guidance}\n\nHash (ID): {model_hash}\n\nThời gian tạo: {creation_date_str}\n",
+    "input_not_valid": "Vui lòng nhập đầu vào hợp lệ!",
+    "output_not_valid": "Vui lòng nhập đầu ra hợp lệ!",
+    "apply_effect": "áp dụng hiệu ứng",
+    "enter_the_text": "Vui lòng nhập văn bản để đọc!",
+    "choose_voice": "Vui lòng chọn giọng!",
+    "convert": "Chuyển đổi {name}...",
+    "separator_music": "tách nhạc",
+    "notfound": "Không tìm thấy",
+    "turn_on_use_audio": "Vui lòng bật sử dụng âm thanh vừa tách để sử dụng",
+    "turn_off_convert_backup": "Tắt chuyển đổi giọng bè để có thể sử dụng giọng gốc",
+    "turn_off_merge_backup": "Tắt không kết hợp giọng bè để có thể sử dụng giọng gốc",
+    "not_found_original_vocal": "Không tìm thấy giọng gốc!",
+    "convert_vocal": "Đang chuyển đổi giọng nói...",
+    "convert_success": "Đã hoàn thành chuyển đổi giọng nói!",
+    "convert_backup": "Đang chuyển đổi giọng bè...",
+    "convert_backup_success": "Đã Hoàn thành chuyển đổi giọng bè!",
+    "merge_backup": "Kết hợp giọng với giọng bè...",
+    "merge_success": "Kết hợp Hoàn thành.",
+    "is_folder": "Đầu vào là một thư mục: Chuyển đổi tất cả tệp âm thanh trong thư mục...",
+    "not_found_in_folder": "Không tìm thấy tệp âm thanh trong thư mục!",
+    "batch_convert": "Đang chuyển đổi hàng loạt...",
+    "batch_convert_success": "Chuyển đổi hàng loạt thành công!",
+    "create": "tạo",
+    "provide_name": "Vui lòng cung cấp tên mô hình.",
+    "not_found_data": "Không tìm thấy dữ liệu",
+    "not_found_data_preprocess": "Không tìm thấy dữ liệu được xử lý, vui lòng xử lý lại âm thanh",
+    "not_found_data_extract": "Không tìm thấy dữ liệu được trích xuất, vui lòng trích xuất lại âm thanh",
+    "provide_pretrained": "Vui lòng nhập huấn luyện {dg}",
+    "download_pretrained": "Tải xuống huấn luyện trước {dg}{rvc_version} gốc",
+    "not_found_pretrain": "Không tìm thấy huấn luyện trước {dg}",
+    "not_use_pretrain": "Sẽ không có huấn luyện trước được sử dụng",
+    "training": "huấn luyện",
+    "display_title": "<h1> 🎵 Giao diện chuyển đổi và huấn luyện mô hình giọng nói được tạo bởi Anh 🎵 <h1>",
+    "rick_roll": "Bấm vào đây nếu bạn muốn bị Rick Roll:) ---> [RickRoll]({rickroll})",
+    "terms_of_use": "**Vui lòng không sử dụng Dự án với bất kỳ mục đích nào vi phạm đạo đức, pháp luật, hoặc gây tổn hại đến cá nhân, tổ chức...**",
+    "exemption": "**Trong trường hợp người sử dụng không tuân thủ các điều khoản hoặc vi phạm, tôi sẽ không chịu trách nhiệm về bất kỳ khiếu nại, thiệt hại, hay trách nhiệm pháp lý nào, dù là trong hợp đồng, do sơ suất, hay các lý do khác, phát sinh từ, ngoài, hoặc liên quan đến phần mềm, việc sử dụng phần mềm hoặc các giao dịch khác liên quan đến phần mềm.**",
+    "separator_tab": "Tách Nhạc",
+    "4_part": "Một hệ thống tách nhạc đơn giản có thể tách được 4 phần: Nhạc, giọng, giọng chính, giọng bè",
+    "clear_audio": "Làm sạch âm thanh",
+    "separator_backing": "Tách giọng bè",
+    "denoise_backing": "Khữ giọng bè",
+    "denoise_mdx": "Khữ tách MDX",
+    "use_mdx": "Sử dụng MDX",
+    "dereveb_audio": "Tách vang giọng",
+    "dereveb_backing": "Tách vang bè",
+    "denoise_dereveb": "Khữ tách vang",
+    "separator_model": "Mô hình tách nhạc",
+    "separator_backing_model": "Mô hình tách bè",
+    "shift": "Số lượng dự đoán",
+    "shift_info": "Càng cao chất lượng càng tốt nhưng lâu nhưng tốn tài nguyên",
+    "segments_size": "Kích Thước Phân Đoạn",
+    "segments_size_info": "Càng cao chất lượng càng tốt nhưng tốn tài nguyên",
+    "batch_size": "Kích thước lô",
+    "batch_size_info": "Số lượng mẫu xử lý đồng thời trong một lần huấn luyện. Cao có thể gây tràn bộ nhớ",
+    "mdx_batch_size_info": "Số lượng mẫu được xử lý cùng một lúc. Việc chia thành các lô giúp tối ưu hóa quá trình tính toán. Lô quá lớn có thể làm tràn bộ nhớ, khi lô quá nhỏ sẽ làm giảm hiệu quả dùng tài nguyên",
+    "overlap": "Chồng chéo",
+    "overlap_info": "Số lượng chồng chéo giữa các cửa sổ dự đoán",
+    "export_format": "Định dạng âm thanh",
+    "export_info": "Định dạng âm thanh khi xuất tệp âm thanh ra",
+    "output_separator": "Âm thanh đã được tách",
+    "hop_length_info": "Biểu thị khoảng thời gian di chuyển cửa sổ phân tích trên tín hiệu âm thanh khi thực hiện các phép biến đổi. Giá trị nhỏ hơn tăng độ chi tiết nhưng tốn tài nguyên tính toán hơn",
+    "drop_audio": "Thả âm thanh vào đây",
+    "drop_text": "Thả tệp văn bản vào đây",
+    "use_url": "Sử dụng đường dẫn youtube",
+    "url_audio": "Đường dẫn liên kết đến âm thanh",
+    "downloads": "Tải Xuống",
+    "clean_strength": "Mức độ làm sạch âm thanh",
+    "clean_strength_info": "Mức độ của bộ làm sạch âm thanh để lọc giọng hát khi xuất",
+    "input_output": "Đầu vào, đầu ra âm thanh",
+    "audio_path": "Đường dẫn đầu vào âm thanh",
+    "refesh": "Tải lại",
+    "output_folder": "Đường dẫn thư mục đầu ra âm thanh",
+    "output_folder_info": "Nhập đường dẫn thư mục âm thanh sẽ xuất ra ở đó",
+    "input_audio": "Đầu vào âm thanh",
+    "instruments": "Nhạc nền",
+    "original_vocal": "Giọng gốc",
+    "main_vocal": "Giọng chính",
+    "backing_vocal": "Giọng bè",
+    "convert_audio": "Chuyển Đổi Âm Thanh",
+    "convert_info": "Chuyển đổi âm thanh bằng mô hình giọng nói đã được huấn luyện",
+    "upscale_audio": "Tăng chất lượng",
+    "autotune": "Tự động điều chỉnh",
+    "use_audio": "Sử dụng âm thanh vừa tách",
+    "convert_original": "Chuyển đổi giọng gốc",
+    "convert_backing": "Chuyển đổi giọng bè",
+    "not_merge_backing": "Không kết hợp giọng bè",
+    "merge_instruments": "Kết hợp nhạc nền",
+    "pitch": "Cao độ",
+    "pitch_info": "Khuyến cáo: chỉnh lên 12 để chuyển giọng nam thành nữ và ngược lại",
+    "model_accordion": "Mô hình và chỉ mục",
+    "model_name": "Tệp mô hình",
+    "index_path": "Tệp chỉ mục",
+    "index_strength": "Ảnh hưởng của chỉ mục",
+    "index_strength_info": "Càng cao ảnh hưởng càng lớn. Tuy nhiên, việc chọn giá trị thấp hơn có thể giảm hiện tượng giả trong âm thanh",
+    "output_path": "Đường dẫn đầu ra âm thanh",
+    "output_path_info": "Nhập đường dẫn đầu ra(cứ để định dạng .wav khi chuyển đổi nó tự sửa)",
+    "setting": "Cài đặt chung",
+    "f0_method": "Phương pháp trích xuất",
+    "f0_method_info": "Phương pháp để trích xuất dữ liệu",
+    "f0_method_hybrid": "Phương pháp trích xuất HYBRID",
+    "f0_method_hybrid_info": "Sự kết hợp của hai loại trích xuất khác nhau",
+    "hubert_model": "Mô hình nhúng",
+    "hubert_info": "Mô hình được huấn luyện trước để giúp nhúng",
+    "modelname": "Tên của mô hình",
+    "modelname_info": "Nếu bạn có mô hình riêng chỉ cần tải và nhập tên của mô hình vào đây",
+    "split_audio": "Cắt âm thanh",
+    "split_audio_info": "Cắt âm thanh ra các phần nhỏ để chuyển đổi có thể giúp tăng tốc độ",
+    "batch_process": "Xử lý lô",
+    "batch_process_info": "Xử lý lô có thể giảm thời gian huấn luyện nhưng có thể bị quá tải",
+    "autotune_rate": "Mức độ điều chỉnh",
+    "autotune_rate_info": "Mức độ điều chỉnh tự động",
+    "resample": "Lấy mẫu lại",
+    "resample_info": "Lấy mẫu lại sau xử lý đến tốc độ lấy mẫu cuối cùng, 0 có nghĩa là không lấy mẫu lại",
+    "filter_radius": "Lọc trung vị",
+    "filter_radius_info": "Nếu giá trị lớn hơn ba sẽ áp dụng tính năng lọc trung vị. Giá trị đại diện cho bán kính bộ lọc và có thể làm giảm hơi thở hoặc tắt thở.",
+    "volume_envelope": "Đường bao âm thanh",
+    "volume_envelope_info": "Sử dụng đường bao âm lượng của đầu vào để thay thế hoặc trộn với đường bao âm lượng của đầu ra. Càng gần 1 thì đường bao đầu ra càng được sử dụng nhiều",
+    "protect": "Bảo vệ phụ âm",
+    "protect_info": "Bảo vệ các phụ âm riêng biệt và âm thanh thở ngăn chặn việc rách điện âm và các hiện tượng giả khác. Việc chỉnh tối đa sẽ bảo vệ toàn diện. Việc giảm giá trị này có thể giảm độ bảo vệ, đồng thời có khả năng giảm thiểu hiệu ứng lập chỉ mục",
+    "output_convert": "Âm thanh đã được chuyển đổi",
+    "main_convert": "Chuyển đổi giọng chính",
+    "main_or_backing": "Giọng chính + Giọng bè",
+    "voice_or_instruments": "Giọng + Nhạc nền",
+    "convert_text": "Chuyển Đổi Văn Bản",
+    "convert_text_markdown": "## Chuyển Đổi Văn Bản Thành Giọng Nói",
+    "convert_text_markdown_2": "Chuyển văn bản thành giọng nói và đọc lại bằng mô hình giọng nói được huấn luyện",
+    "input_txt": "Nhập dữ liệu từ tệp văn bản txt",
+    "text_to_speech": "Văn bản cần đọc",
+    "voice_speed": "Tốc độ đọc",
+    "voice_speed_info": "Tốc độ đọc của giọng nói",
+    "tts_1": "1. Chuyển Đổi Văn Bản",
+    "tts_2": "2. Chuyển Đổi Giọng Nói",
+    "voice": "Giọng nói của các nước",
+    "output_tts": "Đường dẫn đầu ra giọng nói",
+    "output_tts_convert": "Đường dẫn đầu ra giọng chuyển đổi",
+    "tts_output": "Nhập đường dẫn đầu ra",
+    "output_tts_markdown": "Âm thanh chưa được chuyển đổi và âm thanh đã được chuyển đổi",
+    "output_text_to_speech": "Giọng được tạo bởi chuyển đổi văn bản thành giọng nói",
+    "output_file_tts_convert": "Giọng được chuyển đổi bởi mô hình",
+    "output_audio": "Đầu ra âm thanh",
+    "provide_output": "Nhập đường dẫn đầu ra",
+    "audio_effects": "Hiệu Ứng Âm Thanh",
+    "apply_audio_effects": "## Áp Dụng Thêm Hiệu Ứng Cho Âm Thanh",
+    "audio_effects_edit": "Chỉnh sửa thêm hiệu ứng cho âm thanh",
+    "reverb": "Hiệu ứng vọng âm",
+    "chorus": "Hiệu ứng hòa âm",
+    "delay": "Hiệu ứng độ trễ",
+    "more_option": "Tùy chọn thêm",
+    "phaser": "Hiệu ứng xoay pha",
+    "compressor": "Hiệu ứng nén",
+    "apply": "Áp dụng",
+    "reverb_freeze": "Chế độ đóng băng",
+    "reverb_freeze_info": "Tạo hiệu ứng vang liên tục khi bật chế độ này",
+    "room_size": "Kích thước phòng",
+    "room_size_info": "Điều chỉnh không gian của phòng để tạo độ vang",
+    "damping": "Giảm âm",
+    "damping_info": "Điều chỉnh độ hút âm, kiểm soát mức độ vang",
+    "wet_level": "Mức độ tín hiệu vang",
+    "wet_level_info": "Điều chỉnh mức độ của tín hiệu có hiệu ứng vọng âm",
+    "dry_level": "Mức độ tín hiệu gốc",
+    "dry_level_info": "Điều chỉnh mức độ của tín hiệu không có hiệu ứng",
+    "width": "Chiều rộng âm thanh",
+    "width_info": "Điều chỉnh độ rộng của không gian âm thanh",
+    "chorus_depth": "Giảm âm",
+    "chorus_depth_info": "Điều chỉnh cường độ hòa âm, tạo ra cảm giác rộng cho âm thanh",
+    "chorus_rate_hz": "Tần số",
+    "chorus_rate_hz_info": "Điều chỉnh tốc độ dao động của hòa âm",
+    "chorus_mix": "Trộn tín hiệu",
+    "chorus_mix_info": "Điều chỉnh mức độ trộn giữa âm gốc và âm có hiệu ứng",
+    "chorus_centre_delay_ms": "Đỗ trễ trung tâm (mili giây)",
+    "chorus_centre_delay_ms_info": "Khoảng thời gian trễ giữa các kênh stereo để tạo hiệu ứng hòa âm",
+    "chorus_feedback": "Phản hồi",
+    "chorus_feedback_info": "Điều chỉnh lượng tín hiệu hiệu ứng được quay lại vào tín hiệu gốc",
+    "delay_seconds": "Thời gian trễ",
+    "delay_seconds_info": "Điều chỉnh khoảng thời gian trễ giữa âm gốc và âm có hiệu ứng",
+    "delay_feedback": "Phản hồi độ trễ",
+    "delay_feedback_info": "Điều chỉnh lượng tín hiệu được quay lại, tạo hiệu ứng lặp lại",
+    "delay_mix": "Trộn tín hiệu độ trễ",
+    "delay_mix_info": "Điều chỉnh mức độ trộn giữa âm gốc và âm trễ",
+    "fade": "Hiệu ứng mờ dần",
+    "bass_or_treble": "Âm trầm và âm cao",
+    "limiter": "Giới hạn ngưỡng",
+    "distortion": "Hiệu ứng nhiễu âm",
+    "gain": "Cường độ âm",
+    "bitcrush": "Hiệu ứng giảm bits",
+    "clipping": "Hiệu ứng méo âm",
+    "fade_in": "Hiệu ứng mờ dần vào (mili giây)",
+    "fade_in_info": "Thời gian mà âm thanh sẽ tăng dần từ mức 0 đến mức bình thường",
+    "fade_out": "Hiệu ứng mờ dần ra (mili giây)",
+    "fade_out_info": "thời gian mà âm thanh sẽ giảm dần từ xuống 0",
+    "bass_boost": "Độ khuếch đại âm trầm (db)",
+    "bass_boost_info": "mức độ tăng cường âm trầm",
+    "bass_frequency": "Tần số cắt của bộ lọc thông thấp (Hz)",
+    "bass_frequency_info": "tần số mà âm thanh bắt đầu bị giảm. Tần số thấp hơn sẽ làm âm trầm rõ hơn",
+    "treble_boost": "Độ khuếch đại âm cao (db)",
+    "treble_boost_info": "mức độ tăng cường âm cao",
+    "treble_frequency": "Tần số cắt của bộ lọc thông cao (Hz)",
+    "treble_frequency_info": "âm thanh dưới tần số này sẽ bị lọc bỏ. Tần số càng cao thì chỉ giữ lại âm càng cao",
+    "limiter_threashold_db": "Ngưỡng giới hạn",
+    "limiter_threashold_db_info": "Giới hạn mức độ âm thanh tối đa, ngăn không cho vượt quá ngưỡng",
+    "limiter_release_ms": "Thời gian thả",
+    "limiter_release_ms_info": "Khoảng thời gian để âm thanh trở lại sau khi bị giới hạn",
+    "distortion_info": "Điều chỉnh mức độ nhiễu âm, tạo hiệu ứng méo tiếng",
+    "gain_info": "Tăng giảm âm lượng của tín hiệu",
+    "clipping_threashold_db": "Ngưỡng cắt",
+    "clipping_threashold_db_info": "Cắt bớt tín hiệu vượt quá ngưỡng, tạo âm thanh méo",
+    "bitcrush_bit_depth": "Độ sâu bit",
+    "bitcrush_bit_depth_info": "Giảm chất lượng âm thanh bằng cách giảm số bit, tạo hiệu ứng âm thanh bị méo",
+    "phaser_depth": "Độ sâu",
+    "phaser_depth_info": "Điều chỉnh độ sâu của hiệu ứng, ảnh hưởng đến cường độ của hiệu ứng xoay pha",
+    "phaser_rate_hz": "Tần số",
+    "phaser_rate_hz_info": "Điều chỉnh tốc độ của hiệu ứng hiệu ứng xoay pha",
+    "phaser_mix": "Trộn tín hiệu",
+    "phaser_mix_info": "Điều chỉnh mức độ trộn giữa tín hiệu gốc và tín hiệu đã qua xử lý",
+    "phaser_centre_frequency_hz": "Tần số trung tâm",
+    "phaser_centre_frequency_hz_info": "Tần số trung tâm của hiệu ứng xoay pha, ảnh hưởng đến tần số bị điều chỉnh",
+    "phaser_feedback": "Phản hồi",
+    "phaser_feedback_info": "Điều chỉnh lượng phản hồi tín hiệu, tạo cảm giác xoay pha mạnh hoặc nhẹ",
+    "compressor_threashold_db": "Ngưỡng nén",
+    "compressor_threashold_db_info": "Ngưỡng mức âm thanh sẽ bị nén khi vượt qua ngưỡng này",
+    "compressor_ratio": "Tỉ lệ nén",
+    "compressor_ratio_info": "Điều chỉnh mức độ nén âm thanh khi vượt qua ngưỡng",
+    "compressor_attack_ms": "Thời gian tấn công (mili giây)",
+    "compressor_attack_ms_info": "Khoảng thời gian nén bắt đầu tác dụng sau khi âm thanh vượt ngưỡng",
+    "compressor_release_ms": "Thời gian thả",
+    "compressor_release_ms_info": "Thời gian để âm thanh trở lại trạng thái bình thường sau khi bị nén",
+    "create_dataset_url": "Đường dẫn liên kết đến âm thanh(sử dụng dấu , để sử dụng nhiều liên kết)",
+    "createdataset": "Tạo dữ liệu",
+    "create_dataset_markdown": "## Tạo Dữ Liệu Huấn Luyện Từ Youtube",
+    "create_dataset_markdown_2": "Xử lý và tạo tập tin dữ liệu huấn luyện bằng đường dẫn youtube",
+    "denoise": "Khử tách mô hình",
+    "skip": "Bỏ qua giây",
+    "model_ver": "Phiên bản tách giọng",
+    "model_ver_info": "Phiên bản của mô hình tách nhạc để tách giọng",
+    "create_dataset_info": "Thông tin tạo dữ liệu",
+    "output_data": "Đầu ra dữ liệu",
+    "output_data_info": "Đầu ra dữ liệu sau khi tạo xong dữ liệu",
+    "skip_start": "Bỏ qua phần đầu",
+    "skip_start_info": "Bỏ qua số giây đầu của âm thanh, dùng dấu , để sử dụng cho nhiều âm thanh",
+    "skip_end": "Bỏ qua phần cuối",
+    "skip_end_info": "Bỏ qua số giây cuối của âm thanh, dùng dấu , để sử dụng cho nhiều âm thanh",
+    "training_model": "Huấn Luyện Mô Hình",
+    "training_markdown": "Huấn luyện và đào tạo mô hình giọng nói bằng một lượng dữ liệu giọng nói",
+    "training_model_name": "Tên của mô hình khi huấn luyện(không sử dụng ký tự đặc biệt hay dấu cách)",
+    "sample_rate": "Tỉ lệ lấy mẫu",
+    "sample_rate_info": "Tỉ lệ lấy mẫu của mô hình",
+    "training_version": "Phiên bản mô hình",
+    "training_version_info": "Phiên bản mô hình khi huấn luyện",
+    "training_pitch": "Huấn luyện cao độ",
+    "training_pitch_info": "Huấn luyện cao độ cho mô hình",
+    "upload_dataset": "Tải lên dữ liệu huấn luyện",
+    "preprocess_split": "Nên tắt nếu dữ liệu đã được xử lý",
+    "preprocess_effect": "Hiệu ứng quá trình",
+    "preprocess_effect_info": "Nên tắt nếu dữ liệu đã được xử lý",
+    "clear_dataset": "Làm sạch dữ liệu",
+    "clear_dataset_info": "Làm sạch các đoạn dữ liệu",
+    "preprocess_info": "Thông tin phần xử lý trước",
+    "preprocess_button": "1. Xử lý dữ liệu",
+    "extract_button": "2. Trích xuất dữ liệu",
+    "extract_info": "Thông tin phần trích xuất dữ liệu",
+    "total_epoch": "Tổng số kỷ nguyên",
+    "total_epoch_info": "Tổng số kỷ nguyên huấn luyện đào tạo",
+    "save_epoch": "Tần suất lưu",
+    "save_epoch_info": "Tần suất lưu mô hình khi huấn luyện, giúp việc huấn luyện lại mô hình",
+    "create_index": "Tạo chỉ mục",
+    "index_algorithm": "Thuật toán chỉ mục",
+    "index_algorithm_info": "Thuật toán tạo chỉ mục",
+    "custom_dataset": "Tùy chọn thư mục",
+    "custom_dataset_info": "Tùy chọn thư mục dữ liệu huấn luyện",
+    "overtraining_detector": "Kiểm tra quá sức",
+    "overtraining_detector_info": "Kiểm tra huấn luyện mô hình quá sức",
+    "sync_graph": "Đồng bộ biểu đồ",
+    "sync_graph_info": "Chỉ bật khi cần huấn luyện mô hình lại từ đầu.",
+    "cache_in_gpu": "Lưu mô hình vào đệm",
+    "cache_in_gpu_info": "Lưu mô hình vào bộ nhớ đệm gpu",
+    "dataset_folder": "Thư mục chứa dữ liệu",
+    "threshold": "Ngưỡng huấn luyện quá sức",
+    "setting_cpu_gpu": "Tùy chọn CPU/GPU",
+    "gpu_number": "Số gpu được sử dụng",
+    "gpu_number_info": "Số của GPU được sử dụng trong huấn luyện",
+    "save_only_latest": "Chỉ lưu mới nhất",
+    "save_only_latest_info": "Chỉ lưu mô hình D và G mới nhất",
+    "save_every_weights": "Lưu mọi mô hình",
+    "save_every_weights_info": "Lưu mọi mô hình sau mỗi lượt kỷ nguyên",
+    "gpu_info": "Thông tin của GPU",
+    "gpu_info_2": "Thông tin của GPU được sử dụng trong huấn luyện",
+    "cpu_core": "Số lõi xử lý có thể sử d���ng",
+    "cpu_core_info": "Số lõi được sử dụng trong việc huấn luyện",
+    "not_use_pretrain_2": "Không dùng huấn luyện",
+    "not_use_pretrain_info": "Không dùng huấn luyện trước",
+    "custom_pretrain": "Tùy chỉnh huấn luyện",
+    "custom_pretrain_info": "Tùy chỉnh huấn luyện trước",
+    "pretrain_file": "Tệp mô hình huấn luyện trước {dg}",
+    "train_info": "Thông tin phần huấn luyện",
+    "export_model": "5. Xuất Mô hình",
+    "zip_model": "2. Nén mô hình",
+    "output_zip": "Đầu ra tệp khi nén",
+    "refesh_pretrain": "Tải lại huấn luyện trước",
+    "model_path": "Đường dẫn mô hình",
+    "model_ratio": "Tỉ lệ mô hình",
+    "model_ratio_info": "Chỉnh hướng về bên nào sẽ làm cho mô hình giống với bên đó",
+    "output_model_path": "Đầu ra mô hình",
+    "fushion": "Dung Hợp Mô Hình",
+    "fushion_markdown": "## Dung Hợp Hai Mô Hình Với Nhau",
+    "fushion_markdown_2": "Dung hợp hai mô hình giọng nói lại với nhau để tạo thành một mô hình duy nhất",
+    "read_model": "Đọc Thông Tin",
+    "read_model_markdown": "## Đọc Thông Tin Của Mô Hình",
+    "read_model_markdown_2": "Đọc các thông tin được ghi trong mô hình",
+    "drop_model": "Thả mô hình vào đây",
+    "readmodel": "Đọc mô hình",
+    "model_path_info": "Nhập đường dẫn đến tệp mô hình",
+    "modelinfo": "Thông Tin Mô Hình",
+    "download_markdown": "## Tải Xuống Mô Hình",
+    "download_markdown_2": "Tải xuống mô hình giọng nói, mô hình huấn luyện trước, mô hình nhúng",
+    "model_download": "Tải xuống mô hình giọng nói",
+    "model_url": "Đường dẫn liên kết đến mô hình",
+    "30s": "Vui lòng đợi khoảng 30 giây. Hệ thống sẽ tự khởi động lại!",
+    "fp_select": "Đang đặt lại độ chính xác...",
+    "fp_select_2": "Đã đặt độ chính xác là: {fp}",
+    "model_download_select": "Chọn cách tải mô hình",
+    "model_warehouse": "Kho mô hình",
+    "get_model": "Nhận mô hình",
+    "name_to_search": "Tên để tìm kiếm",
+    "search_2": "Tìm kiếm",
+    "select_download_model": "Chọn mô hình đã được tìm kiếm(Bấm vào để chọn)",
+    "download_pretrainec": "Tải xuống mô hình huấn luyện trước",
+    "only_huggingface": "Chỉ hỗ trợ huggingface.co",
+    "pretrained_url": "Đường dẫn liên kết đến mô hình huấn luyện trước {dg}",
+    "select_pretrain": "Chọn mô hình huấn luyện trước",
+    "select_pretrain_info": "Chọn mô hình huấn luyện trước để cài đặt về",
+    "pretrain_sr": "Tốc độ lấy mẫu của mô hình",
+    "drop_pretrain": "Thả mô hình huấn luyện trước {dg} vào đây",
+    "hubert_download": "Tải xuống mô hình nhúng",
+    "hubert_url": "Đường dẫn liên kết tới mô hình nhúng",
+    "drop_hubert": "Thả mô hình nhúng vào đây",
+    "source": "Nguồn Gốc",
+    "source_info": "## Nguồn Gốc Và Tác Giả Của Dự án",
+    "settings": "Tùy Chỉnh",
+    "settings_markdown": "## Tùy Chỉnh Thêm",
+    "settings_markdown_2": "Tùy chỉnh thêm một số tính năng của dự án",
+    "lang": "Ngôn ngữ",
+    "lang_restart": "Ngôn ngữ được hiển thị trong dự án(Khi đổi ngôn ngữ hệ thống sẽ tự khởi động lại sau 30 giây để cập nhật)",
+    "change_lang": "Đổi Ngôn Ngữ",
+    "fp_train": "Độ chính xác huấn luyện",
+    "fp_info": "Dùng fp32 để có độ chính xác cao, nhưng dùng nhiều bộ nhớ và yêu cầu tính toán mạnh mẽ. Dùng fp16 để giảm bộ nhớ, tăng tốc độ xử lý.",
+    "fp_button": "Đổi Độ Chính Xác",
+    "theme": "Chủ đề",
+    "theme_restart": "Loại Chủ đề của giao diện được hiển thị(Khi đổi chủ đề hệ thống sẽ tự khởi động lại sau 30 giây để cập nhật)",
+    "theme_button": "Đổi Chủ Đề",
+    "change_light_dark": "Đổi Chế Độ Sáng/Tối",
+    "tensorboard_url": "Đường dẫn biểu đồ",
+    "errors_loading_audio": "Lỗi khi tải âm thanh: {e}",
+    "apply_error": "Đã xảy ra lỗi khi áp dụng hiệu ứng: {e}",
+    "indexpath": "Đường dẫn chỉ mục",
+    "skip_file": "Phần {i} được bỏ qua vì quá ngắn: {chunk}ms",
+    "split_total": "Tổng số phần đã cắt",
+    "process_audio_error": "Đã xảy ra lỗi khi xử lý âm thanh",
+    "merge_error": "Đã xảy ra lỗi khi ghép âm thanh",
+    "not_found_convert_file": "Không tìm thấy tệp đã xử lý",
+    "convert_batch": "Chuyển đổi hàng loạt...",
+    "found_audio": "Tìm thấy {audio_files} tệp âm thanh cho việc chuyển đổi.",
+    "not_found_audio": "Không tìm thấy tệp âm thanh!",
+    "error_convert": "Đã xảy ra lỗi khi chuyển đổi âm thanh: {e}",
+    "error_convert_batch": "Đã xảy ra lỗi khi chuyển đổi các đoạn âm thanh cắt: {e}",
+    "error_convert_batch_2": "Đã xảy ra lỗi khi chuyển đổi âm thanh hàng loạt: {e}",
+    "convert_batch_success": "Đã chuyển đổi hàng loạt thành công sau {elapsed_time} giây. {output_path}",
+    "convert_audio_success": "Tệp {input_path} được chuyển đổi thành công sau {elapsed_time} giây. {output_path}",
+    "hybrid_methods": "Tính toán ước lượng cao độ f0 cho các phương pháp {methods}",
+    "method_not_valid": "Phương pháp không hợp lệ",
+    "read_faiss_index_error": "Đã xảy ra lỗi khi đọc chỉ mục FAISS: {e}",
+    "read_model_error": "Thất bại khi tải mô hình: {e}",
+    "denoise_error": "Đã xảy ra lỗi khi loại bỏ tiếng ồn: {e}",
+    "change_format": "Đang chuyển đổi âm thanh sang định dạng {output_format}...",
+    "change_format_error": "Đã xảy ra lỗi khi chuyển đổi định dạng âm thanh: {e}",
+    "starting_download": "Bắt đầu tải xuống",
+    "version_not_valid": "Phiên bản tách giọng không hợp lệ",
+    "create_dataset_value_not_valid": "Bật tùy chọn tách nhạc để có thể sử dụng tùy chọn tách vang",
+    "skip<audio": "Không thể bỏ qua vì số lượng thời gian bỏ qua thấp hơn số lượng tệp âm thanh",
+    "skip>audio": "Không thể bỏ qua vì số lượng thời gian bỏ qua cao hơn số lượng tệp âm thanh",
+    "=<0": "Thời gian bỏ qua bé hơn hoặc bằng 0 nên bỏ qua",
+    "skip_warning": "Thời lượng bỏ qua ({seconds} giây) vượt quá thời lượng âm thanh ({total_duration} giây). Bỏ qua.",
+    "download_success": "Đã tải xuống thành công",
+    "create_dataset_error": "Đã xảy ra lỗi khi tạo dữ liệu huấn luyện",
+    "create_dataset_success": "Quá trình tạo dữ liệu huấn huyện đã hoàn thành sau: {elapsed_time} giây",
+    "skip_start_audio": "Bỏ qua âm thanh đầu thành công: {input_file}",
+    "skip_end_audio": "Bỏ qua âm thanh cuối thành công: {input_file}",
+    "merge_audio": "Đã ghép các phần chứa âm thanh lại",
+    "separator_process": "Đang tách giọng: {input}...",
+    "not_found_main_vocal": "Không tìm thấy giọng chính!",
+    "not_found_backing_vocal": "Không tìm thấy giọng bè!",
+    "not_found_instruments": "Không tìm thấy nhạc nền",
+    "merge_instruments_process": "Kết hợp giọng với nhạc nền...",
+    "dereverb": "Đang tách âm vang",
+    "dereverb_success": "Đã tách âm vang thành công",
+    "save_index": "Đã lưu tệp chỉ mục",
+    "create_index_error": "Đã xảy ra lỗi khi tạo chỉ mục",
+    "sr_not_16000": "Tỉ lệ mẫu phải là 16000",
+    "gpu_not_valid": "Chỉ số GPU không hợp lệ. Chuyển sang CPU.",
+    "extract_file_error": "Đã xảy ra lỗi khi giải nén tập tin",
+    "extract_f0_method": "Bắt đầu trích xuất cao độ với {num_processes} lõi với phương pháp trích xuất {f0_method}...",
+    "extract_f0": "Trích Xuất Cao Độ",
+    "extract_f0_success": "Quá trình trích xuất cao độ đã hoàn tất vào {elapsed_time} giây.",
+    "NaN": "chứa giá trị NaN và sẽ bị bỏ qua.",
+    "start_extract_hubert": "Đang bắt đầu nhúng trích xuất...",
+    "not_found_audio_file": "Không tìm thấy tập tin âm thanh. Hãy chắc chắn rằng bạn đã cung cấp âm thanh chính xác.",
+    "extract_hubert": "Trích xuất nhúng",
+    "process_error": "Đã xảy ra lỗi khi xử lý",
+    "extract_hubert_success": "Quá trình trích xuất nhúng đã hoàn tất trong {elapsed_time} giây.",
+    "export_process": "Đường dẫn của mô hình",
+    "extract_error": "Đã xảy ra lỗi khi trích xuất dữ liệu",
+    "extract_success": "Đã trích xuất thành công mô hình",
+    "min_length>=min_interval>=hop_size": "min_length lớn hơn hoặc bằng min_interval lớn hơn hoặc bằng hop_size là bắt buộc",
+    "max_sil_kept>=hop_size": "max_sil_kept lớn hơn hoặc bằng hop_size là bắt buộc",
+    "start_preprocess": "Đang bắt đầu xử lý dữ liệu với {num_processes} lõi xử lý...",
+    "not_integer": "Thư mục ID giọng nói phải là số nguyên, thay vào đó có",
+    "preprocess": "Xử Lý Dữ Liệu",
+    "preprocess_success": "Quá trình xử lý hoàn tất sau {elapsed_time} giây.",
+    "preprocess_model_success": "Đã hoàn thành xử lý trước dữ liệu cho mô hình",
+    "turn_on_dereverb": "Điều kiện cần để sử dụng tách vang giọng bè là phải bật tách vang",
+    "turn_on_separator_backing": "Điều kiện cần để sử dụng tách vang giọng bè là phải bật tách bè",
+    "demucs_model": "Mô hình của Demucs",
+    "backing_model_ver": "Phiên bản mô hình của tách bè",
+    "mdx_model": "Mô hình mdx",
+    "clean_audio_success": "Đã làm sạch âm thành công!",
+    "separator_error": "Đã xảy ra lỗi khi tách nhạc",
+    "separator_success": "Quá trình tách nhạc đã hoàn thành sau: {elapsed_time} giây",
+    "separator_process_2": "Đang xử lý tách nhạc",
+    "separator_success_2": "Đã tách nhạc thành công!",
+    "separator_process_backing": "Đang xử lý tách giọng bè",
+    "separator_process_backing_success": "Đã tách giọng bè thành công!",
+    "process_original": "Đang xử lý tách âm vang giọng gốc...",
+    "process_original_success": "Đã tách âm vang giọng gốc thành công!",
+    "process_main": "Đang xử lý tách âm vang giọng chính...",
+    "process_main_success": "Đã tách âm vang giọng chính thành công!",
+    "process_backing": "Đang xử lý tách âm vang giọng bè...",
+    "process_backing_success": "Đã tách âm vang giọng bè thành công!",
+    "save_every_epoch": "Lưu mô hình sau: {save_every_epoch} kỷ nguyên",
+    "total_e": "Tổng số kỷ nguyên huấn luyện: {total_epoch} kỷ nguyên",
+    "dorg": "Huấn luyện trước G: {pretrainG} | Huấn luyện trước D: {pretrainD}",
+    "training_f0": "Huấn luyện cao độ",
+    "not_gpu": "Không phát hiện thấy GPU, hoàn nguyên về CPU (không khuyến nghị)",
+    "sync": "Biểu đồ đồng bộ hiện đã được kích hoạt! Khi bật biểu đồ đồng bộ, mô hình sẽ trải qua một giai đoạn huấn luyện duy nhất. Sau khi các biểu đồ được đồng bộ hóa, quá trình huấn luyện sẽ tiếp tục với số kỷ nguyên được chỉ định trước đó.",
+    "sync_success": "Đồ thị đã được đồng bộ hóa thành công!",
+    "not_found_checkpoint": "Không tìm thấy tệp điểm đã lưu: {checkpoint_path}",
+    "save_checkpoint": "Đã tải lại điểm đã lưu '{checkpoint_path}' (kỷ nguyên {checkpoint_dict})",
+    "save_model": "Đã lưu mô hình '{checkpoint_path}' (kỷ nguyên {iteration})",
+    "sid_error": "Lỗi chuyển đổi ID người nói '{sid}' thành số nguyên. Ngoại lệ: {e}",
+    "sr_does_not_match": "{sample_rate} Tỉ lệ mẫu không khớp với mục tiêu {sample_rate2} Tỉ lệ mẫu",
+    "spec_error": "Đã xảy ra lỗi khi nhận thông số kỹ thuật từ {spec_filename}: {e}",
+    "time_or_speed_training": "thời gian={current_time} | tốc độ huấn luyện={elapsed_time_str}",
+    "savemodel": "Đã lưu mô hình '{model_dir}' (kỷ nguyên {epoch} và bước {step})",
+    "model_author": "Ghi công mô hình cho {model_author}",
+    "unregistered": "Mô hình không được ghi chép",
+    "not_author": "Mô hình không được ghi chép",
+    "training_author": "Tên chủ mô hình",
+    "training_author_info": "Nếu bạn muốn ghi công mô hình hãy nhập tên của bạn vào đây",
+    "extract_model_error": "Đã xảy ra lỗi khi trích xuất mô hình",
+    "start_training": "Bắt đầu huấn luyện",
+    "import_pretrain": "Đã nạp huấn luyện trước ({dg}) '{pretrain}'",
+    "not_using_pretrain": "Sẽ không có huấn luyện trước ({dg}) được sử dụng",
+    "training_warning": "CẢNH BÁO: Tổn thất tạo ra thấp hơn đã bị vượt quá tổn thất thấp hơn trong kỷ nguyên tiếp theo.",
+    "overtraining_find": "Tập luyện quá sức được phát hiện ở kỷ nguyên {epoch} với mất mát g được làm mịn {smoothed_value_gen} và mất mát d được làm mịn {smoothed_value_disc}",
+    "best_epoch": "Kỷ nguyên mới tốt nhất {epoch} với mất mát g được làm mịn {smoothed_value_gen} và mất mát d được làm mịn {smoothed_value_disc}",
+    "success_training": "Đã đào tạo hoàn thành với {epoch} kỷ nguyên, {global_step} các bước và {loss_gen_all} mất mát gen.",
+    "training_info": "Tổn thất gen thấp nhất: {lowest_value_rounded} ở ký nguyên {lowest_value_epoch}, bước {lowest_value_step}",
+    "model_training_info": "{model_name} | kỷ nguyên={epoch} | bước={global_step} | {epoch_recorder} | giá trị thấp nhất={lowest_value_rounded} (kỷ nguyên {lowest_value_epoch} và bước {lowest_value_step}) | Số kỷ nguyên còn lại để tập luyện quá sức: g/total: {remaining_epochs_gen} d/total: {remaining_epochs_disc} | làm mịn mất mát gen={smoothed_value_gen} | làm mịn mất mát disc={smoothed_value_disc}",
+    "model_training_info_2": "{model_name} | kỷ nguyên={epoch} | bước={global_step} | {epoch_recorder} | giá trị thấp nhất={lowest_value_rounded} (kỷ nguyên {lowest_value_epoch} và bước {lowest_value_step})",
+    "model_training_info_3": "{model_name} | kỷ nguyên={epoch} | bước={global_step} | {epoch_recorder}",
+    "training_error": "Đã xảy ra lỗi khi huấn luyện mô hình:",
+    "separator_info": "Đang khởi tạo với đường dẫn đầu ra: {output_dir}, định dạng đầu ra: {output_format}",
+    "output_dir_is_none": "Thư mục đầu ra không được chỉ định. Sử dụng thư mục làm việc hiện tại.",
+    ">0or=1": "Ngưỡng chuẩn hóa phải lớn hơn 0 và nhỏ hơn hoặc bằng 1.",
+    "output_single": "Đã yêu cầu đầu ra một gốc nên chỉ có một tệp đầu ra ({output_single_stem}) sẽ được ghi",
+    "step2": "Bước thứ hai sẽ được đảo ngược bằng cách sử dụng quang phổ thay vì dạng sóng. Điều này có thể cải thiện chất lượng nhưng chậm hơn một chút.",
+    "other_than_zero": "Cài đặt tốc độ mẫu là {sample_rate} nhưng phải là số nguyên khác 0.",
+    "too_high": "Cài đặt tốc độ mẫu là {sample_rate}. Nhập một cái gì đó ít tham vọng hơn.",
+    "sr_not_valid": "Tốc độ mẫu phải là số nguyên khác 0. Vui lòng cung cấp số nguyên hợp lệ.",
+    "name_ver": "Phiên bản {name}",
+    "os": "Hệ điều hành",
+    "platform_info": "Hệ thống: {system_info} Tên: {node} Phát hành: {release} Máy: {machine} Vi xử lý: {processor}",
+    "install_ffmpeg": "Đã cài đặt FFmpeg",
+    "none_ffmpeg": "FFmpeg chưa được cài đặt. Vui lòng cài đặt FFmpeg để sử dụng gói này.",
+    "install_onnx": "Gói {pu} ONNX Runtime được cài đặt cùng với phiên bản",
+    "running_in_cpu": "Không thể cấu hình khả năng tăng tốc phần cứng, chạy ở chế độ CPU",
+    "running_in_cuda": "CUDA có sẵn trong Torch, cài đặt thiết bị Torch thành CUDA",
+    "onnx_have": "ONNXruntime có sẵn {have}, cho phép tăng tốc",
+    "onnx_not_have": "{have} không có sẵn trong ONNXruntime, do đó khả năng tăng tốc sẽ KHÔNG được bật",
+    "python_not_install": "Gói Python: {package_name} chưa được cài đặt",
+    "hash": "Tính hash của tệp mô hình {model_path}",
+    "ioerror": "IOError đang tìm kiếm -10 MB hoặc đọc tệp mô hình để tính toán hàm băm: {e}",
+    "cancel_download": "Tệp đã tồn tại tại {output_path}, bỏ qua quá trình tải xuống",
+    "download_model": "Đang tải tệp từ {url} xuống {output_path} với thời gian chờ 300 giây",
+    "download_error": "Không tải được tệp xuống từ {url}, mã phản hồi: {status_code}",
+    "vip_model": "Mô hình: '{model_friendly_name}' là mô hình cao cấp, được Anjok07 dự định chỉ dành cho những người đăng ký trả phí truy cập.",
+    "vip_print": "Này bạn, nếu bạn chưa đăng ký, vui lòng cân nhắc việc hỗ trợ cho nhà phát triển của UVR, Anjok07 bằng cách đăng ký tại đây: https://patreon.com/uvr",
+    "search_model": "Đang tìm kiếm mô hình {model_filename} trong tập tin các mô hình được hỗ trợ trong nhóm",
+    "load_download_json": "Đã tải danh sách tải xuống mô hình",
+    "single_model": "Đã xác định được tệp mô hình đơn: {model_friendly_name}",
+    "not_found_model": "Không tìm thấy mô hình trong kho lưu trữ UVR, đang cố tải xuống từ kho lưu trữ mô hình phân tách âm thanh...",
+    "single_model_path": "Đường dẫn trả về cho tệp mô hình đơn: {model_path}",
+    "find_model": "Đã tìm thấy tên tệp đầu vào {model_filename} trong mô hình nhiều tệp: {model_friendly_name}",
+    "find_models": "Đã xác định mô hình nhiều tệp: {model_friendly_name}, lặp qua các tệp để tải xuống",
+    "find_path": "Đang cố gắng xác định ĐƯỜNG DẪN tải xuống cho cặp cấu hình",
+    "not_found_model_warehouse": "Không tìm thấy mô hình trong kho lưu trữ UVR, đang cố tải xuống từ kho lưu trữ mô hình phân tách âm thanh...",
+    "yaml_warning": "Tên mô hình bạn đã chỉ định, {model_filename} thực sự là tệp cấu hình mô hình chứ không phải tệp mô hình.",
+    "yaml_warning_2": "Chúng tôi đã tìm thấy một mô hình khớp với tệp cấu hình này: {config_key} nên chúng tôi sẽ sử dụng tệp mô hình đó cho lần chạy này.",
+    "yaml_warning_3": "Để tránh hành vi gây nhầm lẫn/không nhất quán trong tương lai, thay vào đó hãy chỉ định tên tệp mô hình thực tế.",
+    "yaml_debug": "Không tìm thấy tệp cấu hình mô hình YAML trong kho lưu trữ UVR, đang cố tải xuống từ kho lưu trữ mô hình phân tách âm thanh...",
+    "download_model_friendly": "Tất cả các tệp đã tải xuống cho mô hình {model_friendly_name}, trả về đường dẫn ban đầu {model_path}",
+    "not_found_model_2": "Không tìm thấy tệp mô hình {model_filename} trong các tệp mô hình được hỗ trợ",
+    "load_yaml": "Đang tải dữ liệu mô hình từ YAML tại đường dẫn {model_data_yaml_filepath}",
+    "load_yaml_2": "Dữ liệu mô hình được tải từ tệp YAML: {model_data}",
+    "hash_md5": "Tính hash MD5 cho tệp mô hình để xác định các tham số mô hình từ dữ liệu UVR...",
+    "model_hash": "Mô hình {model_path} có hash {model_hash}",
+    "mdx_data": "Đường dẫn dữ liệu mô hình MDX được đặt thành {mdx_model_data_path}",
+    "load_mdx": "Đang tải các tham số mô hình MDX từ tệp dữ liệu mô hình UVR...",
+    "model_not_support": "Tệp mô hình không được hỗ trợ: không thể tìm thấy tham số cho hash MD5 {model_hash} trong tệp dữ liệu mô hình UVR cho vòm MDX.",
+    "uvr_json": "Dữ liệu mô hình được tải từ UVR JSON bằng hàm băm {model_hash}: {model_data}",
+    "loading_model": "Đang tải mô hình {model_filename}...",
+    "download_model_friendly_2": "Đã tải xuống mô hình, tên thân thiện: {model_friendly_name}, Đường dẫn mô hình: {model_path}",
+    "model_type_not_support": "Loại mô hình không được hỗ trợ: {model_type}",
+    "demucs_not_support_python<3.10": "Các mô hình Demucs yêu cầu phiên bản Python 3.10 trở lên.",
+    "import_module": "Nhập mô-đun cho loại mô hình",
+    "initialization": "Khởi tạo lớp phân cách cho loại mô hình",
+    "loading_model_success": "Đang tải mô hình hoàn tất.",
+    "loading_model_duration": "Tải thời lượng mô hình",
+    "starting_separator": "Bắt đầu quá trình tách cho đường dẫn tập tin âm thanh",
+    "normalization": "Ngưỡng chuẩn hóa được đặt thành {normalization_threshold}, dạng sóng sẽ hạ xuống biên độ tối đa này để tránh bị cắt.",
+    "loading_separator_model": "Đang tải xuống mô hình {model_filename}...",
+    "separator_success_3": "Quá trình tách hoàn tất.",
+    "separator_duration": "Thời gian tách",
+    "downloading_model": "Đã tải xuống mô hình, loại: {model_type}, tên thân thiện: {model_friendly_name}, đường dẫn mô hình: {model_path}, dữ liệu mô hình: {model_data_dict_size} mục",
+    "not_found_model_signature": "Không thể tìm thấy mô hình được đào tạo trước có chữ ký {sig}.",
+    "repo_must_be_folder": "{repo} phải tồn tại và là một thư mục.",
+    "demucs_info": "Thông số Demucs: Kích thước phân đoạn = {segment_size}, Kích hoạt kích thước phân đoạn = {segments_enabled}",
+    "demucs_info_2": "Thông số Demucs: Số lượng dự đoán = {shifts}, Chồng chéo = {overlap}",
+    "start_demucs": "Khởi tạo hoàn tất Demucs Separator",
+    "start_separator": "Bắt đầu quá trình tách...",
+    "prepare_mix": "Chuẩn bị hỗn hợp...",
+    "demix": "Hỗn hợp đã chuẩn bị để khử trộn. Hình dạng: {shape}",
+    "cancel_mix": "Đang tải mô hình để hủy trộn...",
+    "model_review": "Mô hình được tải và đặt ở chế độ đánh giá.",
+    "del_gpu_cache_after_demix": "Đã xóa bộ nhớ đệm mô hình và GPU sau khi hủy trộn.",
+    "process_output_file": "Đang xử lý tập tin đầu ra...",
+    "source_length": "Đang xử lý mảng nguồn, độ dài nguồn là {source_length}",
+    "process_ver": "Đang xử lý nguồn phiên bản...",
+    "set_map": "Đặt bản đồ nguồn thành {part} phần gốc...",
+    "process_all_part": "Xử lý cho tất cả các phần gốc...",
+    "skip_part": "Bỏ qua phần viết gốc {stem_name} vì out_single_stem được đặt thành {output_single_stem}...",
+    "starting_demix_demucs": "Đang bắt đầu quá trình trộn trong demix_demucs...",
+    "model_infer": "Chạy mô hình suy luận...",
+    "del_all_but_one": "Có mô hình được đào tạo trước trùng lặp cho chữ ký {xp_sig}. Xin vui lòng xóa tất cả trừ một.",
+    "name_not_pretrained": "{name} không phải là một mô hình được đào tạo trước hay một túi mô hình.",
+    "invalid_checksum": "Tổng kiểm tra không hợp lệ cho tệp {path}, dự kiến {checksum} nhưng lại nhận được {actual_checksum}",
+    "mdx_info": "Thông số MDX: Kích thước lô = {batch_size}, Kích thước phân đoạn = {segment_size}",
+    "mdx_info_2": "Thông số MDX: Chồng chéo = {overlap}, Hop_length = {hop_length}, Kích hoạt khữ nhiễu = {enable_denoise}",
+    "mdx_info_3": "Thông số MDX",
+    "load_model_onnx": "Đang tải mô hình ONNX để suy luận...",
+    "load_model_onnx_success": "Đã tải mô hình thành công bằng phiên suy luận ONNXruntime.",
+    "onnx_to_pytorch": "Mô hình được chuyển đổi từ onnx sang pytorch do kích thước phân đoạn không khớp với dim_t, quá trình xử lý có thể chậm hơn.",
+    "stft": "STFT nghịch đảo được áp dụng. Trả về kết quả có hình dạng",
+    "no_denoise": "Mô hình chạy trên quang phổ mà không khử nhiễu.",
+    "mix": "Đang chuẩn bị trộn cho tệp âm thanh đầu vào {audio_file_path}...",
+    "normalization_demix": "Chuẩn hóa hỗn hợp trước khi khử trộn...",
+    "mix_success": "Quá trình trộn hoàn tất.",
+    "primary_source": "Bình thường hóa nguồn chính...",
+    "secondary_source": "Sản xuất nguồn thứ cấp: Trộn ở chế độ trộn phù hợp",
+    "invert_using_spec": "Đảo ngược thân thứ cấp bằng cách sử dụng quang phổ khi invert_USE_spec được đặt thành True",
+    "invert_using_spec_2": "Đảo ngược thân thứ cấp bằng cách trừ đi thân cây được chuyển đổi từ hỗn hợp ban đầu được chuyển đổi",
+    "enable_denoise": "Mô hình chạy trên cả phổ âm và dương để khử nhiễu.",
+    "is_match_mix": "is_match_mix: dự đoán phổ thu được trực tiếp từ đầu ra STFT.",
+    "save_secondary_stem_output_path": "Đang lưu phần gốc {stem_name} vào {stem_output_path}...",
+    "starting_model": "Đang khởi tạo cài đặt mô hình...",
+    "input_info": "Thông số đầu vào của mô hình",
+    "model_settings": "Cài đặt mô hình",
+    "initialize_mix": "Đang khởi tạo kết hợp với is_ckpt = {is_ckpt}. Hình dạng trộn ban đầu: {shape}",
+    "!=2": "Dự kiến có tín hiệu âm thanh 2 kênh nhưng lại có {shape} kênh",
+    "process_check": "Xử lý ở chế độ điểm kiểm tra...",
+    "stft_2": "STFT được áp dụng trên hỗn hợp. Hình dạng quang phổ: {shape}",
+    "cache": "Khoảng đệm được tính toán",
+    "shape": "Hình dạng hỗn hợp sau khi đệm: {shape}, Số phần: {num_chunks}",
+    "process_no_check": "Xử lý ở chế độ không có điểm kiểm tra...",
+    "n_sample_or_pad": "Số lượng mẫu: {n_sample}, Đã tính đệm: {pad}",
+    "shape_2": "Hình dạng hỗn hợp sau khi đệm",
+    "process_part": "Đoạn đã xử lý {mix_waves}: Bắt đầu {i}, Kết thúc {ii}",
+    "mix_waves_to_tensor": "Đã chuyển đổi mix_waves thành tensor. Hình dạng tensor: {shape}",
+    "mix_match": "Chế độ trộn Match; áp dụng hệ số bù.",
+    "tar_waves": "Sóng tar_waves. Hình dạng",
+    "normalization_2": "Chuẩn hóa kết quả bằng cách chia kết quả cho số chia.",
+    "mix_wave": "Đang xử lý lô mix_wave",
+    "mix_or_batch": "Trộn phần chia thành từng đợt. Số lượng lô",
+    "demix_is_match_mix": "Bắt đầu quá trình hủy trộn với is_match_mix,",
+    "mix_shape": "Hỗn hợp phần gốc được lưu trữ. Hình dạng",
+    "chunk_size_or_overlap": "Kích thước đoạn để trộn phù hợp: {chunk_size}, Chồng chéo: {overlap}",
+    "chunk_size_or_overlap_standard": "Kích thước phần tiêu chuẩn: {chunk_size}, Chồng chéo: {overlap}",
+    "calc_size": "Kích thước được tạo được tính toán",
+    "window": "Cửa sổ được áp dụng cho đoạn này.",
+    "process_part_2": "Đang xử lý đoạn {total}/{total_chunks}: Bắt đầu {start}, Kết thúc {end}",
+    "all_process_part": "Tổng số phần cần xử lý",
+    "step_or_overlap": "Kích thước bước để xử lý các phần: {step} khi chồng chéo được đặt thành {overlap}.",
+    "mix_cache": "Hỗn hợp được chuẩn bị với lớp đệm. Hình dạng hỗn hợp",
+    "dims": "Không thể sử dụng mã hóa vị trí sin/cos với thứ nguyên lẻ (có dim={dims})",
+    "activation": "kích hoạt phải là relu/gelu, không phải {activation}",
+    "length_or_training_length": "Độ dài cho trước {length} dài hơn thời lượng huấn luyện {training_length}",
+    "type_not_valid": "Loại không hợp lệ cho",
+    "del_parameter": "Bỏ tham số không tồn tại ",
+    "info": "Các thông số phổ biến: Tên mô hình = {model_name}, Đường dẫn mô hình = {model_path}",
+    "info_2": "Các thông số phổ biến: Đường dẫn đầu ra = {output_dir}, Định dạng đầu ra = {output_format}",
+    "info_3": "Các thông số phổ biến: ngưỡng chuẩn hóa = {normalization_threshold}",
+    "info_4": "Các thông số phổ biến: Kích hoạt khữ nhiễu = {enable_denoise}, Đầu ra một phần = {output_single_stem}",
+    "info_5": "Các thông số phổ biến: Đảo ngược bằng cách sử dụng thông số kỹ thuật = {invert_using_spec}, tỷ lệ mẫu = {sample_rate}",
+    "info_6": "Các thông số phổ biến: Tên phần gốc chính = {primary_stem_name}, Tên phần gốc phụ = {secondary_stem_name}",
+    "info_7": "Các thông số phổ biến: Là Karaoke = {is_karaoke}, là mô hình bv = {is_bv_model}, tái cân bằng mô hình bv = {bv_model_rebalance}",
+    "success_process": "Đang hoàn tất quá trình xử lý phần gốc {stem_name} và ghi âm thanh...",
+    "load_audio": "Đang tải âm thanh từ tập tin",
+    "load_audio_success": "Đã tải âm thanh. Tốc độ mẫu: {sr}, Hình dạng âm thanh: {shape}",
+    "convert_mix": "Chuyển đổi mảng hỗn hợp được cung cấp.",
+    "convert_shape": "Hình dạng hỗn hợp chuyển đổi: {shape}",
+    "audio_not_valid": "Tệp âm thanh {audio_path} trống hoặc không hợp lệ",
+    "audio_valid": "Tệp âm thanh hợp lệ và chứa dữ liệu.",
+    "mix_single": "Hỗn hợp là đơn sắc. Chuyển đổi sang âm thanh nổi.",
+    "convert_mix_audio": "Đã chuyển đổi thành bản trộn âm thanh nổi.",
+    "mix_success_2": "Công tác chuẩn bị hỗn hợp đã hoàn tất.",
+    "duration": "Thời lượng âm thanh là {duration_hours} giờ ({duration_seconds} giây).",
+    "write": "Sử dụng {name} để viết.",
+    "write_audio": "Đang nhập {name} bằng đường dẫn gốc:",
+    "original_not_valid": "Cảnh báo: mảng nguồn gốc gần như im lặng hoặc trống.",
+    "shape_audio": "Hình dạng dữ liệu âm thanh trước khi xử lý",
+    "convert_data": "Kiểu dữ liệu trước khi chuyển đổi",
+    "original_source_to_int16": "Đã chuyển đổi original_source thành int16.",
+    "shape_audio_2": "Hình dạng dữ liệu âm thanh xen kẽ",
+    "create_audiosegment": "Đã tạo AudioSegment thành công.",
+    "create_audiosegment_error": "Lỗi cụ thể khi tạo AudioSegment",
+    "export_error": "Lỗi xuất file âm thanh",
+    "export_success": "Đã xuất thành công tệp âm thanh sang",
+    "clean": "Chạy thu gom rác...",
+    "clean_cache": "Xóa bộ nhớ đệm {name}...",
+    "del_path": "Xóa đường dẫn, nguồn và gốc của tệp âm thanh đầu vào...",
+    "not_success": "Quá trình đăng không thành công: ",
+    "resample_error": "Lỗi trong quá trình lấy mẫu lại",
+    "shapes": "Hình dạng",
+    "wav_resolution": "Loại độ phân giải",
+    "warnings": "Cảnh báo: Đã phát hiện các giá trị cực kỳ hung hãn",
+    "warnings_2": "Cảnh báo: Đã phát hiện NaN hoặc giá trị vô hạn trong đầu vào sóng. Hình dạng",
+    "process_file": "Đang xử lý tập tin... \n",
+    "save_instruments": "Lưu bản nhạc ngược...",
+    "assert": "Các tệp âm thanh phải có hình dạng giống nhau - Mix: {mixshape}, Inst: {instrumentalshape}",
+    "rubberband": "Không thể thực Rubberband. Vui lòng xác minh rằng Rubberband-cli đã được cài đặt.",
+    "rate": "Tỉ lệ phải hoàn toàn tích cực",
+    "gdown_warning": "Bạn đã chỉ định một liên kết Google Drive không phải là liên kết chính xác để tải xuống một tập tin. Bạn có thể muốn thử tùy chọn `--fuzzy` hoặc url sau: https://drive.google.com/uc?id={file_id}",
+    "gdown_error": "Không thể truy xuất liên kết công khai của tệp. Bạn có thể cần phải thay đổi quyền thành bất kỳ ai có liên kết hoặc đã có nhiều quyền truy cập.",
+    "gdown_error_2": "Không thể truy xuất đường dẫn tệp:\n\n{error}\n\nBạn vẫn có thể truy cập tệp từ trình duyệt:\n\n\t{url_origin}\n\nnhưng Gdown không thể. Vui lòng kiểm tra kết nối và quyền.",
+    "temps": "Có nhiều tệp tạm thời để tiếp tục:",
+    "del_all_temps": "Vui lòng xóa chúng ngoại trừ một để tiếp tục tải xuống.",
+    "continue": "Tiếp tục:",
+    "to": "Đến:",
+    "gdown_value_error": "Phải chỉ định đường dẫn hoặc id",
+    "missing_url": "Thiếu đường dẫn",
+    "mac_not_match": "MAC không khớp",
+    "file_not_access": "Tệp tin không thể truy cập",
+    "int_resp==-3": "Yêu cầu không thành công, đang thử lại"
+}

assets/logs/mute/f0/mute.wav.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b9acf9ab7facdb032e1d687fe35182670b0b94566c4b209ae48c239d19956a6
+size 1332

assets/logs/mute/f0_voiced/mute.wav.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30792849c8e72d67e6691754077f2888b101cb741e9c7f193c91dd9692870c87
+size 2536

assets/logs/mute/v1_extracted/mute.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64d5abbac078e19a3f649c0d78a02cb33a71407ded3ddf2db78e6b803d0c0126
+size 152704

assets/logs/mute/v2_extracted/mute.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16ef62b957887ac9f0913aa5158f18983afff1ef5a3e4c5fd067ac20fc380d54
+size 457856

assets/model/uvr5/download_checks.json

@@ -0,0 +1,169 @@
+{
+        "current_version": "UVR_Patch_10_6_23_4_27",
+        
+        "mdx_download_list": {      
+                "MDX-Net Model: UVR-MDX-NET Inst HQ 1": "UVR-MDX-NET-Inst_HQ_1.onnx",
+                "MDX-Net Model: UVR-MDX-NET Inst HQ 2": "UVR-MDX-NET-Inst_HQ_2.onnx",
+                "MDX-Net Model: UVR-MDX-NET Inst HQ 3": "UVR-MDX-NET-Inst_HQ_3.onnx",
+                "MDX-Net Model: UVR-MDX-NET Inst HQ 4": "UVR-MDX-NET-Inst_HQ_4.onnx",
+                "MDX-Net Model: UVR-MDX-NET Main": "UVR_MDXNET_Main.onnx",
+                "MDX-Net Model: UVR-MDX-NET Inst Main": "UVR-MDX-NET-Inst_Main.onnx",
+                "MDX-Net Model: UVR-MDX-NET 1": "UVR_MDXNET_1_9703.onnx",
+                "MDX-Net Model: UVR-MDX-NET 2": "UVR_MDXNET_2_9682.onnx",
+                "MDX-Net Model: UVR-MDX-NET 3": "UVR_MDXNET_3_9662.onnx",
+                "MDX-Net Model: UVR-MDX-NET Inst 1": "UVR-MDX-NET-Inst_1.onnx",
+                "MDX-Net Model: UVR-MDX-NET Inst 2": "UVR-MDX-NET-Inst_2.onnx",
+                "MDX-Net Model: UVR-MDX-NET Inst 3": "UVR-MDX-NET-Inst_3.onnx",
+                "MDX-Net Model: UVR-MDX-NET Karaoke": "UVR_MDXNET_KARA.onnx",
+                "MDX-Net Model: UVR-MDX-NET Karaoke 2": "UVR_MDXNET_KARA_2.onnx",
+                "MDX-Net Model: UVR_MDXNET_9482": "UVR_MDXNET_9482.onnx",
+                "MDX-Net Model: UVR-MDX-NET Voc FT": "UVR-MDX-NET-Voc_FT.onnx",
+                "MDX-Net Model: Kim Vocal 1": "Kim_Vocal_1.onnx",
+                "MDX-Net Model: Kim Vocal 2": "Kim_Vocal_2.onnx",
+                "MDX-Net Model: Kim Inst": "Kim_Inst.onnx",
+                "MDX-Net Model: Reverb HQ By FoxJoy": "Reverb_HQ_By_FoxJoy.onnx",
+                "MDX-Net Model: UVR-MDX-NET Crowd HQ 1 By Aufr33": "UVR-MDX-NET_Crowd_HQ_1.onnx"
+        },
+
+        "demucs_download_list":{
+                "Demucs v4: htdemucs_ft":{
+                        "f7e0c4bc-ba3fe64a.th":"https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/f7e0c4bc-ba3fe64a.th",
+                        "d12395a8-e57c48e6.th":"https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/d12395a8-e57c48e6.th",
+                        "92cfc3b6-ef3bcb9c.th":"https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/92cfc3b6-ef3bcb9c.th",
+                        "04573f0d-f3cf25b2.th":"https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/04573f0d-f3cf25b2.th",
+                        "htdemucs_ft.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/htdemucs_ft.yaml"
+                },
+
+                "Demucs v4: htdemucs":{
+                        "955717e8-8726e21a.th": "https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/955717e8-8726e21a.th",
+                        "htdemucs.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/htdemucs.yaml"
+                },
+
+                "Demucs v4: hdemucs_mmi":{
+                        "75fc33f5-1941ce65.th": "https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/75fc33f5-1941ce65.th",
+                        "hdemucs_mmi.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/hdemucs_mmi.yaml"
+                },
+
+                "Demucs v4: htdemucs_6s":{
+                        "5c90dfd2-34c22ccb.th": "https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/5c90dfd2-34c22ccb.th",
+                        "htdemucs_6s.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/htdemucs_6s.yaml"
+                },
+
+                "Demucs v3: mdx":{
+                        "0d19c1c6-0f06f20e.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/0d19c1c6-0f06f20e.th", 
+                        "7ecf8ec1-70f50cc9.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/7ecf8ec1-70f50cc9.th",
+                        "c511e2ab-fe698775.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/c511e2ab-fe698775.th",
+                        "7d865c68-3d5dd56b.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/7d865c68-3d5dd56b.th",
+                        "mdx.yaml": "https://raw.githubusercontent.com/facebookresearch/demucs/main/demucs/remote/mdx.yaml"
+                },
+                
+                "Demucs v3: mdx_q":{
+                        "6b9c2ca1-3fd82607.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/6b9c2ca1-3fd82607.th",
+                        "b72baf4e-8778635e.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/b72baf4e-8778635e.th",
+                        "42e558d4-196e0e1b.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/42e558d4-196e0e1b.th",
+                        "305bc58f-18378783.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/305bc58f-18378783.th",
+                        "mdx_q.yaml": "https://raw.githubusercontent.com/facebookresearch/demucs/main/demucs/remote/mdx_q.yaml"
+                },
+                
+                "Demucs v3: mdx_extra":{
+                        "e51eebcc-c1b80bdd.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/e51eebcc-c1b80bdd.th",
+                        "a1d90b5c-ae9d2452.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/a1d90b5c-ae9d2452.th",
+                        "5d2d6c55-db83574e.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/5d2d6c55-db83574e.th",
+                        "cfa93e08-61801ae1.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/cfa93e08-61801ae1.th",
+                        "mdx_extra.yaml": "https://raw.githubusercontent.com/facebookresearch/demucs/main/demucs/remote/mdx_extra.yaml"
+                },
+                
+                "Demucs v3: mdx_extra_q": {
+                        "83fc094f-4a16d450.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/83fc094f-4a16d450.th",
+                        "464b36d7-e5a9386e.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/464b36d7-e5a9386e.th",
+                        "14fc6a69-a89dd0ee.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/14fc6a69-a89dd0ee.th",
+                        "7fd6ef75-a905dd85.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/7fd6ef75-a905dd85.th",
+                        "mdx_extra_q.yaml": "https://raw.githubusercontent.com/facebookresearch/demucs/main/demucs/remote/mdx_extra_q.yaml"
+                },
+                
+                "Demucs v3: UVR Model":{
+                        "ebf34a2db.th": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/ebf34a2db.th",
+                        "UVR_Demucs_Model_1.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/UVR_Demucs_Model_1.yaml"
+                },
+
+                "Demucs v3: repro_mdx_a":{
+                        "9a6b4851-03af0aa6.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/9a6b4851-03af0aa6.th", 
+                        "1ef250f1-592467ce.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/1ef250f1-592467ce.th",
+                        "fa0cb7f9-100d8bf4.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/fa0cb7f9-100d8bf4.th",
+                        "902315c2-b39ce9c9.th": "https://dl.fbaipublicfiles.com/demucs/mdx_final/902315c2-b39ce9c9.th",
+                        "repro_mdx_a.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/repro_mdx_a.yaml"
+                },
+
+                "Demucs v3: repro_mdx_a_time_only":{
+                        "9a6b4851-03af0aa6.th":"https://dl.fbaipublicfiles.com/demucs/mdx_final/9a6b4851-03af0aa6.th",
+                        "1ef250f1-592467ce.th":"https://dl.fbaipublicfiles.com/demucs/mdx_final/1ef250f1-592467ce.th",
+                        "repro_mdx_a_time_only.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/repro_mdx_a_time_only.yaml"
+                },
+
+                "Demucs v3: repro_mdx_a_hybrid_only":{
+                        "fa0cb7f9-100d8bf4.th":"https://dl.fbaipublicfiles.com/demucs/mdx_final/fa0cb7f9-100d8bf4.th",
+                        "902315c2-b39ce9c9.th":"https://dl.fbaipublicfiles.com/demucs/mdx_final/902315c2-b39ce9c9.th",
+                        "repro_mdx_a_hybrid_only.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/repro_mdx_a_hybrid_only.yaml"
+                },
+
+                "Demucs v2: demucs": {
+                        "demucs-e07c671f.th": "https://dl.fbaipublicfiles.com/demucs/v3.0/demucs-e07c671f.th"
+                },
+        
+                "Demucs v2: demucs_extra": {
+                        "demucs_extra-3646af93.th":"https://dl.fbaipublicfiles.com/demucs/v3.0/demucs_extra-3646af93.th"
+                },
+
+                "Demucs v2: demucs48_hq": {
+                        "demucs48_hq-28a1282c.th":"https://dl.fbaipublicfiles.com/demucs/v3.0/demucs48_hq-28a1282c.th"
+                },
+
+                "Demucs v2: tasnet": {
+                        "tasnet-beb46fac.th":"https://dl.fbaipublicfiles.com/demucs/v3.0/tasnet-beb46fac.th"
+                },
+
+                "Demucs v2: tasnet_extra": {
+                        "tasnet_extra-df3777b2.th":"https://dl.fbaipublicfiles.com/demucs/v3.0/tasnet_extra-df3777b2.th"
+                },
+                                
+                "Demucs v2: demucs_unittest": {
+                        "demucs_unittest-09ebc15f.th":"https://dl.fbaipublicfiles.com/demucs/v3.0/demucs_unittest-09ebc15f.th"
+                },
+
+                "Demucs v1: demucs": {
+                        "demucs.th":"https://dl.fbaipublicfiles.com/demucs/v2.0/demucs.th"
+                },
+
+                "Demucs v1: demucs_extra": {
+                        "demucs_extra.th":"https://dl.fbaipublicfiles.com/demucs/v2.0/demucs_extra.th"
+                },
+
+                "Demucs v1: light": {
+                        "light.th":"https://dl.fbaipublicfiles.com/demucs/v2.0/light.th"
+                },
+
+                "Demucs v1: light_extra": {
+                        "light_extra.th":"https://dl.fbaipublicfiles.com/demucs/v2.0/light_extra.th"
+                },
+
+                "Demucs v1: tasnet": {
+                        "tasnet.th":"https://dl.fbaipublicfiles.com/demucs/v2.0/tasnet.th"
+                },
+                                
+                "Demucs v1: tasnet_extra": {
+                        "tasnet_extra.th":"https://dl.fbaipublicfiles.com/demucs/v2.0/tasnet_extra.th"
+                }
+        },
+
+        "mdx_download_vip_list": {
+                "MDX-Net Model VIP: UVR-MDX-NET_Main_340": "UVR-MDX-NET_Main_340.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET_Main_390": "UVR-MDX-NET_Main_390.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET_Main_406": "UVR-MDX-NET_Main_406.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET_Main_427": "UVR-MDX-NET_Main_427.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET_Main_438": "UVR-MDX-NET_Main_438.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET_Inst_82_beta": "UVR-MDX-NET_Inst_82_beta.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET_Inst_90_beta": "UVR-MDX-NET_Inst_90_beta.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET_Inst_187_beta": "UVR-MDX-NET_Inst_187_beta.onnx",
+                "MDX-Net Model VIP: UVR-MDX-NET-Inst_full_292": "UVR-MDX-NET-Inst_full_292.onnx"
+        }
+}

main/app/clean.py

@@ -0,0 +1,40 @@
+import time
+import threading
+
+from googleapiclient.discovery import build
+
+
+class Clean:
+    def __init__(self, every=300):
+        self.service = build('drive', 'v3')
+        self.every = every  
+        self.trash_cleanup_thread = None
+
+    def delete(self):
+        page_token = None
+
+        while 1:
+            response = self.service.files().list(q="trashed=true", spaces='drive', fields="nextPageToken, files(id, name)", pageToken=page_token).execute()
+            
+            for file in response.get('files', []):
+                if file['name'].startswith("G_") or file['name'].startswith("D_"):
+                    try:
+                        self.service.files().delete(fileId=file['id']).execute()
+                    except Exception as e:
+                        raise RuntimeError(e)
+
+            page_token = response.get('nextPageToken', None)
+            if page_token is None: break
+
+    def clean(self):
+        while 1:
+            self.delete()
+            time.sleep(self.every)
+
+    def start(self):
+        self.trash_cleanup_thread = threading.Thread(target=self.clean)
+        self.trash_cleanup_thread.daemon = True 
+        self.trash_cleanup_thread.start()
+
+    def stop(self):
+        if self.trash_cleanup_thread: self.trash_cleanup_thread.join()

main/app/run_tensorboard.py

@@ -0,0 +1,30 @@
+import os
+import sys
+import json
+import logging
+import webbrowser
+
+from tensorboard import program
+
+sys.path.append(os.getcwd())
+
+from main.configs.config import Config
+translations = Config().translations
+
+with open(os.path.join("main", "configs", "config.json"), "r") as f:
+    configs = json.load(f)
+
+def launch_tensorboard_pipeline():
+    logging.getLogger("root").setLevel(logging.ERROR)
+    logging.getLogger("tensorboard").setLevel(logging.ERROR)
+
+    tb = program.TensorBoard()
+    tb.configure(argv=[None, "--logdir", "assets/logs", f"--port={configs["tensorboard_port"]}"])
+    url = tb.launch()
+
+    print(f"{translations['tensorboard_url']}: {url}")
+    webbrowser.open(url)
+
+    return f"{translations['tensorboard_url']}: {url}"
+
+if __name__ == "__main__": launch_tensorboard_pipeline()

main/app/sync.py

@@ -0,0 +1,80 @@
+import time
+import threading
+import subprocess
+
+from typing import List, Union
+
+
+class Channel:
+    def __init__(self, source, destination, sync_deletions=False, every=60, exclude: Union[str, List, None] = None):
+        self.source = source
+        self.destination = destination
+        self.event = threading.Event()
+        self.syncing_thread = threading.Thread(target=self._sync, args=())
+        self.sync_deletions = sync_deletions
+        self.every = every
+
+
+        if not exclude: exclude = []
+        if isinstance(exclude,str): exclude = [exclude]
+
+        self.exclude = exclude
+        self.command = ['rsync', '-aP']
+
+
+    def alive(self):
+        if self.syncing_thread.is_alive(): return True
+        else: return False
+
+
+    def _sync(self):
+        command = self.command
+
+        for exclusion in self.exclude:
+            command.append(f'--exclude={exclusion}')
+
+        command.extend([f'{self.source}/', f'{self.destination}/'])
+
+        if self.sync_deletions: command.append('--delete')
+
+        while not self.event.is_set():
+            subprocess.run(command)
+            time.sleep(self.every)
+
+
+    def copy(self):
+        command = self.command
+
+        for exclusion in self.exclude:
+            command.append(f'--exclude={exclusion}')
+
+        command.extend([f'{self.source}/', f'{self.destination}/'])
+
+        if self.sync_deletions: command.append('--delete')
+        subprocess.run(command)
+
+        return True
+    
+
+    def start(self):
+        if self.syncing_thread.is_alive():
+            self.event.set()
+            self.syncing_thread.join()
+
+        if self.event.is_set(): self.event.clear()
+        if self.syncing_thread._started.is_set(): self.syncing_thread = threading.Thread(target=self._sync, args=())
+
+        self.syncing_thread.start()
+
+        return self.alive()
+
+
+    def stop(self):
+        if self.alive():
+            self.event.set()
+            self.syncing_thread.join()
+
+            while self.alive():
+                if not self.alive(): break
+
+        return not self.alive()

main/configs/config.json

@@ -0,0 +1,229 @@
+{
+    "language": "vi-VN",
+    "support_language": [
+        "en-US",
+        "vi-VN"
+    ],
+
+
+    "theme": "NoCrypt/miku",
+    "themes": [
+        "NoCrypt/miku",
+        "gstaff/xkcd",
+        "JohnSmith9982/small_and_pretty",
+        "ParityError/Interstellar",
+        "earneleh/paris",
+        "shivi/calm_seafoam",
+        "Hev832/Applio",
+        "YTheme/Minecraft",
+        "gstaff/sketch",
+        "SebastianBravo/simci_css",
+        "allenai/gradio-theme",
+        "Nymbo/Nymbo_Theme_5",
+        "lone17/kotaemon",
+        "Zarkel/IBM_Carbon_Theme",
+        "SherlockRamos/Feliz",
+        "freddyaboulton/dracula_revamped",
+        "freddyaboulton/bad-theme-space",
+        "gradio/dracula_revamped",
+        "abidlabs/dracula_revamped",
+        "gradio/dracula_test",
+        "gradio/seafoam",
+        "gradio/glass",
+        "gradio/monochrome",
+        "gradio/soft",
+        "gradio/default",
+        "gradio/base",
+        "abidlabs/pakistan",
+        "dawood/microsoft_windows",
+        "ysharma/steampunk",
+        "ysharma/huggingface",
+        "abidlabs/Lime",
+        "freddyaboulton/this-theme-does-not-exist-2",
+        "aliabid94/new-theme",
+        "aliabid94/test2",
+        "aliabid94/test3",
+        "aliabid94/test4",
+        "abidlabs/banana",
+        "freddyaboulton/test-blue",
+        "gstaff/whiteboard",
+        "ysharma/llamas",
+        "abidlabs/font-test",
+        "YenLai/Superhuman",
+        "bethecloud/storj_theme",
+        "sudeepshouche/minimalist",
+        "knotdgaf/gradiotest",
+        "ParityError/Anime",
+        "Ajaxon6255/Emerald_Isle",
+        "ParityError/LimeFace",
+        "finlaymacklon/smooth_slate",
+        "finlaymacklon/boxy_violet",
+        "derekzen/stardust",
+        "EveryPizza/Cartoony-Gradio-Theme",
+        "Ifeanyi/Cyanister",
+        "Tshackelton/IBMPlex-DenseReadable",
+        "snehilsanyal/scikit-learn",
+        "Himhimhim/xkcd",
+        "nota-ai/theme",
+        "rawrsor1/Everforest",
+        "rottenlittlecreature/Moon_Goblin",
+        "abidlabs/test-yellow",
+        "abidlabs/test-yellow3",
+        "idspicQstitho/dracula_revamped",
+        "kfahn/AnimalPose",
+        "HaleyCH/HaleyCH_Theme",
+        "simulKitke/dracula_test",
+        "braintacles/CrimsonNight",
+        "wentaohe/whiteboardv2",
+        "reilnuud/polite",
+        "remilia/Ghostly",
+        "Franklisi/darkmode",
+        "coding-alt/soft",
+        "xiaobaiyuan/theme_land",
+        "step-3-profit/Midnight-Deep",
+        "xiaobaiyuan/theme_demo",
+        "Taithrah/Minimal",
+        "Insuz/SimpleIndigo",
+        "zkunn/Alipay_Gradio_theme",
+        "Insuz/Mocha",
+        "xiaobaiyuan/theme_brief",
+        "Ama434/434-base-Barlow",
+        "Ama434/def_barlow",
+        "Ama434/neutral-barlow",
+        "dawood/dracula_test",
+        "nuttea/Softblue",
+        "BlueDancer/Alien_Diffusion",
+        "naughtondale/monochrome",
+        "Dagfinn1962/standard",
+        "default"
+    ],
+
+
+    "tts_voice": [
+        "af-ZA-AdriNeural", "af-ZA-WillemNeural", "sq-AL-AnilaNeural", 
+        "sq-AL-IlirNeural", "am-ET-AmehaNeural", "am-ET-MekdesNeural", 
+        "ar-DZ-AminaNeural", "ar-DZ-IsmaelNeural", "ar-BH-AliNeural", 
+        "ar-BH-LailaNeural", "ar-EG-SalmaNeural", "ar-EG-ShakirNeural", 
+        "ar-IQ-BasselNeural", "ar-IQ-RanaNeural", "ar-JO-SanaNeural", 
+        "ar-JO-TaimNeural", "ar-KW-FahedNeural", "ar-KW-NouraNeural", 
+        "ar-LB-LaylaNeural", "ar-LB-RamiNeural", "ar-LY-ImanNeural", 
+        "ar-LY-OmarNeural", "ar-MA-JamalNeural", "ar-MA-MounaNeural", 
+        "ar-OM-AbdullahNeural", "ar-OM-AyshaNeural", "ar-QA-AmalNeural", 
+        "ar-QA-MoazNeural", "ar-SA-HamedNeural", "ar-SA-ZariyahNeural", 
+        "ar-SY-AmanyNeural", "ar-SY-LaithNeural", "ar-TN-HediNeural", 
+        "ar-TN-ReemNeural", "ar-AE-FatimaNeural", "ar-AE-HamdanNeural", 
+        "ar-YE-MaryamNeural", "ar-YE-SalehNeural", "az-AZ-BabekNeural", 
+        "az-AZ-BanuNeural", "bn-BD-NabanitaNeural", "bn-BD-PradeepNeural", 
+        "bn-IN-BashkarNeural", "bn-IN-TanishaaNeural", "bs-BA-GoranNeural", 
+        "bs-BA-VesnaNeural", "bg-BG-BorislavNeural", "bg-BG-KalinaNeural", 
+        "my-MM-NilarNeural", "my-MM-ThihaNeural", "ca-ES-EnricNeural", 
+        "ca-ES-JoanaNeural", "zh-HK-HiuGaaiNeural", "zh-HK-HiuMaanNeural", 
+        "zh-HK-WanLungNeural", "zh-CN-XiaoxiaoNeural", "zh-CN-XiaoyiNeural", 
+        "zh-CN-YunjianNeural", "zh-CN-YunxiNeural", "zh-CN-YunxiaNeural", 
+        "zh-CN-YunyangNeural", "zh-CN-liaoning-XiaobeiNeural", "zh-TW-HsiaoChenNeural", 
+        "zh-TW-YunJheNeural", "zh-TW-HsiaoYuNeural", "zh-CN-shaanxi-XiaoniNeural", 
+        "hr-HR-GabrijelaNeural", "hr-HR-SreckoNeural", "cs-CZ-AntoninNeural", 
+        "cs-CZ-VlastaNeural", "da-DK-ChristelNeural", "da-DK-JeppeNeural", 
+        "nl-BE-ArnaudNeural", "nl-BE-DenaNeural", "nl-NL-ColetteNeural", 
+        "nl-NL-FennaNeural", "nl-NL-MaartenNeural", "en-AU-NatashaNeural", 
+        "en-AU-WilliamNeural", "en-CA-ClaraNeural", "en-CA-LiamNeural", 
+        "en-HK-SamNeural", "en-HK-YanNeural", "en-IN-NeerjaExpressiveNeural", 
+        "en-IN-NeerjaNeural", "en-IN-PrabhatNeural", "en-IE-ConnorNeural", 
+        "en-IE-EmilyNeural", "en-KE-AsiliaNeural", "en-KE-ChilembaNeural", 
+        "en-NZ-MitchellNeural", "en-NZ-MollyNeural", "en-NG-AbeoNeural", 
+        "en-NG-EzinneNeural", "en-PH-JamesNeural", "en-PH-RosaNeural", 
+        "en-SG-LunaNeural", "en-SG-WayneNeural", "en-ZA-LeahNeural", 
+        "en-ZA-LukeNeural", "en-TZ-ElimuNeural", "en-TZ-ImaniNeural", 
+        "en-GB-LibbyNeural", "en-GB-MaisieNeural", "en-GB-RyanNeural", 
+        "en-GB-SoniaNeural", "en-GB-ThomasNeural", "en-US-AvaMultilingualNeural", 
+        "en-US-AndrewMultilingualNeural", "en-US-EmmaMultilingualNeural", 
+        "en-US-BrianMultilingualNeural", "en-US-AvaNeural", "en-US-AndrewNeural", 
+        "en-US-EmmaNeural", "en-US-BrianNeural", "en-US-AnaNeural", "en-US-AriaNeural", 
+        "en-US-ChristopherNeural", "en-US-EricNeural", "en-US-GuyNeural", 
+        "en-US-JennyNeural", "en-US-MichelleNeural", "en-US-RogerNeural", 
+        "en-US-SteffanNeural", "et-EE-AnuNeural", "et-EE-KertNeural", 
+        "fil-PH-AngeloNeural", "fil-PH-BlessicaNeural", "fi-FI-HarriNeural", 
+        "fi-FI-NooraNeural", "fr-BE-CharlineNeural", "fr-BE-GerardNeural", 
+        "fr-CA-ThierryNeural", "fr-CA-AntoineNeural", "fr-CA-JeanNeural", 
+        "fr-CA-SylvieNeural", "fr-FR-VivienneMultilingualNeural", "fr-FR-RemyMultilingualNeural", 
+        "fr-FR-DeniseNeural", "fr-FR-EloiseNeural", "fr-FR-HenriNeural", 
+        "fr-CH-ArianeNeural", "fr-CH-FabriceNeural", "gl-ES-RoiNeural", 
+        "gl-ES-SabelaNeural", "ka-GE-EkaNeural", "ka-GE-GiorgiNeural", 
+        "de-AT-IngridNeural", "de-AT-JonasNeural", "de-DE-SeraphinaMultilingualNeural", 
+        "de-DE-FlorianMultilingualNeural", "de-DE-AmalaNeural", "de-DE-ConradNeural", 
+        "de-DE-KatjaNeural", "de-DE-KillianNeural", "de-CH-JanNeural", 
+        "de-CH-LeniNeural", "el-GR-AthinaNeural", "el-GR-NestorasNeural", 
+        "gu-IN-DhwaniNeural", "gu-IN-NiranjanNeural", "he-IL-AvriNeural", 
+        "he-IL-HilaNeural", "hi-IN-MadhurNeural", "hi-IN-SwaraNeural", 
+        "hu-HU-NoemiNeural", "hu-HU-TamasNeural", "is-IS-GudrunNeural", 
+        "is-IS-GunnarNeural", "id-ID-ArdiNeural", "id-ID-GadisNeural", 
+        "ga-IE-ColmNeural", "ga-IE-OrlaNeural", "it-IT-GiuseppeNeural", 
+        "it-IT-DiegoNeural", "it-IT-ElsaNeural", "it-IT-IsabellaNeural", 
+        "ja-JP-KeitaNeural", "ja-JP-NanamiNeural", "jv-ID-DimasNeural", 
+        "jv-ID-SitiNeural", "kn-IN-GaganNeural", "kn-IN-SapnaNeural", 
+        "kk-KZ-AigulNeural", "kk-KZ-DauletNeural", "km-KH-PisethNeural", 
+        "km-KH-SreymomNeural", "ko-KR-HyunsuNeural", "ko-KR-InJoonNeural", 
+        "ko-KR-SunHiNeural", "lo-LA-ChanthavongNeural", "lo-LA-KeomanyNeural", 
+        "lv-LV-EveritaNeural", "lv-LV-NilsNeural", "lt-LT-LeonasNeural", 
+        "lt-LT-OnaNeural", "mk-MK-AleksandarNeural", "mk-MK-MarijaNeural", 
+        "ms-MY-OsmanNeural", "ms-MY-YasminNeural", "ml-IN-MidhunNeural", 
+        "ml-IN-SobhanaNeural", "mt-MT-GraceNeural", "mt-MT-JosephNeural", 
+        "mr-IN-AarohiNeural", "mr-IN-ManoharNeural", "mn-MN-BataaNeural", 
+        "mn-MN-YesuiNeural", "ne-NP-HemkalaNeural", "ne-NP-SagarNeural", 
+        "nb-NO-FinnNeural", "nb-NO-PernilleNeural", "ps-AF-GulNawazNeural", 
+        "ps-AF-LatifaNeural", "fa-IR-DilaraNeural", "fa-IR-FaridNeural", 
+        "pl-PL-MarekNeural", "pl-PL-ZofiaNeural", "pt-BR-ThalitaNeural", 
+        "pt-BR-AntonioNeural", "pt-BR-FranciscaNeural", "pt-PT-DuarteNeural", 
+        "pt-PT-RaquelNeural", "ro-RO-AlinaNeural", "ro-RO-EmilNeural", 
+        "ru-RU-DmitryNeural", "ru-RU-SvetlanaNeural", "sr-RS-NicholasNeural", 
+        "sr-RS-SophieNeural", "si-LK-SameeraNeural", "si-LK-ThiliniNeural", 
+        "sk-SK-LukasNeural", "sk-SK-ViktoriaNeural", "sl-SI-PetraNeural", 
+        "sl-SI-RokNeural", "so-SO-MuuseNeural", "so-SO-UbaxNeural", 
+        "es-AR-ElenaNeural", "es-AR-TomasNeural", "es-BO-MarceloNeural", 
+        "es-BO-SofiaNeural", "es-CL-CatalinaNeural", "es-CL-LorenzoNeural", 
+        "es-ES-XimenaNeural", "es-CO-GonzaloNeural", "es-CO-SalomeNeural", 
+        "es-CR-JuanNeural", "es-CR-MariaNeural", "es-CU-BelkysNeural", 
+        "es-CU-ManuelNeural", "es-DO-EmilioNeural", "es-DO-RamonaNeural", 
+        "es-EC-AndreaNeural", "es-EC-LuisNeural", "es-SV-LorenaNeural", 
+        "es-SV-RodrigoNeural", "es-GQ-JavierNeural", "es-GQ-TeresaNeural", 
+        "es-GT-AndresNeural", "es-GT-MartaNeural", "es-HN-CarlosNeural", 
+        "es-HN-KarlaNeural", "es-MX-DaliaNeural", "es-MX-JorgeNeural", 
+        "es-NI-FedericoNeural", "es-NI-YolandaNeural", "es-PA-MargaritaNeural", 
+        "es-PA-RobertoNeural", "es-PY-MarioNeural", "es-PY-TaniaNeural", 
+        "es-PE-AlexNeural", "es-PE-CamilaNeural", "es-PR-KarinaNeural", 
+        "es-PR-VictorNeural", "es-ES-AlvaroNeural", "es-ES-ElviraNeural", 
+        "es-US-AlonsoNeural", "es-US-PalomaNeural", "es-UY-MateoNeural", 
+        "es-UY-ValentinaNeural", "es-VE-PaolaNeural", "es-VE-SebastianNeural", 
+        "su-ID-JajangNeural", "su-ID-TutiNeural", "sw-KE-RafikiNeural", 
+        "sw-KE-ZuriNeural", "sw-TZ-DaudiNeural", "sw-TZ-RehemaNeural", 
+        "sv-SE-MattiasNeural", "sv-SE-SofieNeural", "ta-IN-PallaviNeural", 
+        "ta-IN-ValluvarNeural", "ta-MY-KaniNeural", "ta-MY-SuryaNeural", 
+        "ta-SG-AnbuNeural", "ta-SG-VenbaNeural", "ta-LK-KumarNeural", 
+        "ta-LK-SaranyaNeural", "te-IN-MohanNeural", "te-IN-ShrutiNeural", 
+        "th-TH-NiwatNeural", "th-TH-PremwadeeNeural", "tr-TR-AhmetNeural", 
+        "tr-TR-EmelNeural", "uk-UA-OstapNeural", "uk-UA-PolinaNeural", 
+        "ur-IN-GulNeural", "ur-IN-SalmanNeural", "ur-PK-AsadNeural", 
+        "ur-PK-UzmaNeural", "uz-UZ-MadinaNeural", "uz-UZ-SardorNeural", 
+        "vi-VN-HoaiMyNeural", "vi-VN-NamMinhNeural", "cy-GB-AledNeural", 
+        "cy-GB-NiaNeural", "zu-ZA-ThandoNeural", "zu-ZA-ThembaNeural"
+    ],
+
+    
+    "separator_tab": true,
+    "convert_tab": true,
+    "tts_tab": true,
+    "effects_tab": true,
+    "create_dataset_tab": true,
+    "training_tab": true,
+    "fushion_tab": true,
+    "read_tab": true,
+    "downloads_tab": true,
+    "settings_tab": true,
+
+    "app_port": 7860,
+    "tensorboard_port": 6870,
+    "num_of_restart": 5,
+    "server_name": "0.0.0.0",
+    "app_show_error": true,
+    "share": true
+}

main/configs/config.py

@@ -0,0 +1,120 @@
+import os
+import json
+import torch
+
+
+version_config_paths = [
+    os.path.join("v1", "32000.json"),
+    os.path.join("v1", "40000.json"),
+    os.path.join("v1", "48000.json"),
+    os.path.join("v2", "48000.json"),
+    os.path.join("v2", "40000.json"),
+    os.path.join("v2", "32000.json"),
+]
+
+
+def singleton(cls):
+    instances = {}
+
+    def get_instance(*args, **kwargs):
+        if cls not in instances: instances[cls] = cls(*args, **kwargs)
+
+        return instances[cls]
+    return get_instance
+
+
+@singleton
+class Config:
+    def __init__(self):
+        self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
+        self.is_half = self.device != "cpu"
+        self.gpu_name = (torch.cuda.get_device_name(int(self.device.split(":")[-1])) if self.device.startswith("cuda") else None)
+        self.json_config = self.load_config_json()
+        self.translations = self.multi_language()
+        self.gpu_mem = None
+        self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
+
+
+    def load_config_json(self) -> dict:
+        configs = {}
+        
+        for config_file in version_config_paths:
+            config_path = os.path.join("main", "configs", config_file)
+
+            with open(config_path, "r") as f:
+                configs[config_file] = json.load(f)
+
+        return configs
+    
+    def multi_language(self):
+        with open(os.path.join("main", "configs", "config.json"), "r") as f:
+            configs = json.load(f)
+
+        lang = configs["language"]
+        
+        if len([l for l in os.listdir(os.path.join("assets", "languages")) if l.endswith(".json")]) < 1: raise FileNotFoundError("Không tìm thấy bất cứ gói ngôn ngữ nào(No package languages found)")
+        
+        if not lang: lang = "vi-VN"
+        if lang not in configs["support_language"]: raise ValueError("Ngôn ngữ không được hỗ trợ(Language not supported)")
+
+        lang_path = os.path.join("assets", "languages", f"{lang}.json")
+        if not os.path.exists(lang_path): lang_path = os.path.join("assets", "languages", f"vi-VN.json")
+
+        with open(lang_path, encoding="utf-8") as f:
+            translations = json.load(f)
+        
+        
+        return translations
+
+    def set_precision(self, precision):
+        if precision not in ["fp32", "fp16"]: raise ValueError("Loại chính xác không hợp lệ. Phải là 'fp32' hoặc 'fp16'(Invalid precision type. Must be 'fp32' or 'fp16').")
+
+        fp16_run_value = precision == "fp16"
+
+        for config_path in version_config_paths:
+            full_config_path = os.path.join("main", "configs", config_path)
+
+            try:
+                with open(full_config_path, "r") as f:
+                    config = json.load(f)
+
+                config["train"]["fp16_run"] = fp16_run_value
+
+                with open(full_config_path, "w") as f:
+                    json.dump(config, f, indent=4)
+            except FileNotFoundError:
+                print(self.translations["not_found"].format(name=full_config_path))
+
+        return self.translations["set_precision"].format(precision=precision)
+
+    def device_config(self) -> tuple:
+        if self.device.startswith("cuda"):
+            self.set_cuda_config()
+        elif self.has_mps():
+            self.device = "mps"
+            self.is_half = False
+            self.set_precision("fp32")
+        else:
+            self.device = "cpu"
+            self.is_half = False
+            self.set_precision("fp32")
+
+        x_pad, x_query, x_center, x_max = ((3, 10, 60, 65) if self.is_half else (1, 6, 38, 41))
+        
+        if self.gpu_mem is not None and self.gpu_mem <= 4: x_pad, x_query, x_center, x_max = (1, 5, 30, 32)
+
+        return x_pad, x_query, x_center, x_max
+
+    def set_cuda_config(self):
+        i_device = int(self.device.split(":")[-1])
+        self.gpu_name = torch.cuda.get_device_name(i_device)
+        low_end_gpus = ["16", "P40", "P10", "1060", "1070", "1080"]
+
+        if (any(gpu in self.gpu_name for gpu in low_end_gpus) and "V100" not in self.gpu_name.upper()):
+            self.is_half = False
+            self.set_precision("fp32")
+
+        self.gpu_mem = torch.cuda.get_device_properties(i_device).total_memory // (1024**3)
+
+    def has_mps(self) -> bool:
+        return torch.backends.mps.is_available()

main/configs/v1/32000.json

@@ -0,0 +1,82 @@
+{
+    "train": {
+        "log_interval": 200,
+        "seed": 1234,
+        "epochs": 20000,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-09,
+        "batch_size": 4,
+        "fp16_run": false,
+        "lr_decay": 0.999875,
+        "segment_size": 12800,
+        "init_lr_ratio": 1,
+        "warmup_epochs": 0,
+        "c_mel": 45,
+        "c_kl": 1.0
+    },
+    "data": {
+        "max_wav_value": 32768.0,
+        "sample_rate": 32000,
+        "filter_length": 1024,
+        "hop_length": 320,
+        "win_length": 1024,
+        "n_mel_channels": 80,
+        "mel_fmin": 0.0,
+        "mel_fmax": null
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "text_enc_hidden_dim": 256,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            10,
+            4,
+            2,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            16,
+            16,
+            4,
+            4,
+            4
+        ],
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "spk_embed_dim": 109
+    }
+}

main/configs/v1/40000.json

@@ -0,0 +1,80 @@
+{
+    "train": {
+        "log_interval": 200,
+        "seed": 1234,
+        "epochs": 20000,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-09,
+        "batch_size": 4,
+        "fp16_run": false,
+        "lr_decay": 0.999875,
+        "segment_size": 12800,
+        "init_lr_ratio": 1,
+        "warmup_epochs": 0,
+        "c_mel": 45,
+        "c_kl": 1.0
+    },
+    "data": {
+        "max_wav_value": 32768.0,
+        "sample_rate": 40000,
+        "filter_length": 2048,
+        "hop_length": 400,
+        "win_length": 2048,
+        "n_mel_channels": 125,
+        "mel_fmin": 0.0,
+        "mel_fmax": null
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "text_enc_hidden_dim": 256,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            10,
+            10,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            16,
+            16,
+            4,
+            4
+        ],
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "spk_embed_dim": 109
+    }
+}

main/configs/v1/48000.json

@@ -0,0 +1,82 @@
+{
+    "train": {
+        "log_interval": 200,
+        "seed": 1234,
+        "epochs": 20000,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-09,
+        "batch_size": 4,
+        "fp16_run": false,
+        "lr_decay": 0.999875,
+        "segment_size": 11520,
+        "init_lr_ratio": 1,
+        "warmup_epochs": 0,
+        "c_mel": 45,
+        "c_kl": 1.0
+    },
+    "data": {
+        "max_wav_value": 32768.0,
+        "sample_rate": 48000,
+        "filter_length": 2048,
+        "hop_length": 480,
+        "win_length": 2048,
+        "n_mel_channels": 128,
+        "mel_fmin": 0.0,
+        "mel_fmax": null
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "text_enc_hidden_dim": 256,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            10,
+            6,
+            2,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            16,
+            16,
+            4,
+            4,
+            4
+        ],
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "spk_embed_dim": 109
+    }
+}

main/configs/v2/32000.json

@@ -0,0 +1,76 @@
+{
+    "train": {
+        "log_interval": 200,
+        "seed": 1234,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-09,
+        "fp16_run": false,
+        "lr_decay": 0.999875,
+        "segment_size": 12800,
+        "c_mel": 45,
+        "c_kl": 1.0
+    },
+    "data": {
+        "max_wav_value": 32768.0,
+        "sample_rate": 32000,
+        "filter_length": 1024,
+        "hop_length": 320,
+        "win_length": 1024,
+        "n_mel_channels": 80,
+        "mel_fmin": 0.0,
+        "mel_fmax": null
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "text_enc_hidden_dim": 768,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            10,
+            8,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            20,
+            16,
+            4,
+            4
+        ],
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "spk_embed_dim": 109
+    }
+}

main/configs/v2/40000.json

@@ -0,0 +1,76 @@
+{
+    "train": {
+        "log_interval": 200,
+        "seed": 1234,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-09,
+        "fp16_run": false,
+        "lr_decay": 0.999875,
+        "segment_size": 12800,
+        "c_mel": 45,
+        "c_kl": 1.0
+    },
+    "data": {
+        "max_wav_value": 32768.0,
+        "sample_rate": 40000,
+        "filter_length": 2048,
+        "hop_length": 400,
+        "win_length": 2048,
+        "n_mel_channels": 125,
+        "mel_fmin": 0.0,
+        "mel_fmax": null
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "text_enc_hidden_dim": 768,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            10,
+            10,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            16,
+            16,
+            4,
+            4
+        ],
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "spk_embed_dim": 109
+    }
+}

main/configs/v2/48000.json

@@ -0,0 +1,76 @@
+{
+    "train": {
+        "log_interval": 200,
+        "seed": 1234,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-09,
+        "fp16_run": false,
+        "lr_decay": 0.999875,
+        "segment_size": 17280,
+        "c_mel": 45,
+        "c_kl": 1.0
+    },
+    "data": {
+        "max_wav_value": 32768.0,
+        "sample_rate": 48000,
+        "filter_length": 2048,
+        "hop_length": 480,
+        "win_length": 2048,
+        "n_mel_channels": 128,
+        "mel_fmin": 0.0,
+        "mel_fmax": null
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "text_enc_hidden_dim": 768,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            12,
+            10,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            24,
+            20,
+            4,
+            4
+        ],
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "spk_embed_dim": 109
+    }
+}

main/inference/audio_effects.py

@@ -0,0 +1,170 @@
+import os
+import sys
+import librosa
+import argparse
+
+import numpy as np
+import soundfile as sf
+
+from distutils.util import strtobool
+from scipy.signal import butter, filtfilt
+from pedalboard import Pedalboard, Chorus, Distortion, Reverb, PitchShift, Delay, Limiter, Gain, Bitcrush, Clipping, Compressor, Phaser
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+translations = Config().translations
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input_path", type=str, required=True)
+    parser.add_argument("--output_path", type=str, default="./audios/apply_effects.wav")
+    parser.add_argument("--resample", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--resample_sr", type=int, default=0)
+    parser.add_argument("--chorus", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--chorus_depth", type=float, default=0.5)
+    parser.add_argument("--chorus_rate", type=float, default=1.5)
+    parser.add_argument("--chorus_mix", type=float, default=0.5)
+    parser.add_argument("--chorus_delay", type=int, default=10)
+    parser.add_argument("--chorus_feedback", type=float, default=0)
+    parser.add_argument("--distortion", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--drive_db", type=int, default=20)
+    parser.add_argument("--reverb", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--reverb_room_size", type=float, default=0.5)
+    parser.add_argument("--reverb_damping", type=float, default=0.5)
+    parser.add_argument("--reverb_wet_level", type=float, default=0.33)
+    parser.add_argument("--reverb_dry_level", type=float, default=0.67)
+    parser.add_argument("--reverb_width", type=float, default=1)
+    parser.add_argument("--reverb_freeze_mode", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--pitchshift", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--pitch_shift", type=int, default=0)
+    parser.add_argument("--delay", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--delay_seconds", type=float, default=0.5)
+    parser.add_argument("--delay_feedback", type=float, default=0.5)
+    parser.add_argument("--delay_mix", type=float, default=0.5)
+    parser.add_argument("--compressor", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--compressor_threshold", type=int, default=-20)
+    parser.add_argument("--compressor_ratio", type=float, default=4)
+    parser.add_argument("--compressor_attack_ms", type=float, default=10)
+    parser.add_argument("--compressor_release_ms", type=int, default=200)
+    parser.add_argument("--limiter", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--limiter_threshold", type=int, default=0)
+    parser.add_argument("--limiter_release", type=int, default=100)
+    parser.add_argument("--gain", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--gain_db", type=int, default=0)
+    parser.add_argument("--bitcrush", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--bitcrush_bit_depth", type=int, default=16)
+    parser.add_argument("--clipping", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--clipping_threshold", type=int, default=-10)
+    parser.add_argument("--phaser", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--phaser_rate_hz", type=float, default=0.5)
+    parser.add_argument("--phaser_depth", type=float, default=0.5)
+    parser.add_argument("--phaser_centre_frequency_hz", type=int, default=1000)
+    parser.add_argument("--phaser_feedback", type=float, default=0)
+    parser.add_argument("--phaser_mix", type=float, default=0.5)
+    parser.add_argument("--treble_bass_boost", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--bass_boost_db", type=int, default=0)
+    parser.add_argument("--bass_boost_frequency", type=int, default=100)
+    parser.add_argument("--treble_boost_db", type=int, default=0)
+    parser.add_argument("--treble_boost_frequency", type=int, default=3000)
+    parser.add_argument("--fade_in_out", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--fade_in_duration", type=float, default=2000)
+    parser.add_argument("--fade_out_duration", type=float, default=2000)
+    parser.add_argument("--export_format", type=str, default="wav")
+    
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_arguments()
+
+    process_audio(input_path=args.input_path, output_path=args.output_path, resample=args.resample, resample_sr=args.resample_sr, chorus_depth=args.chorus_depth, chorus_rate=args.chorus_rate, chorus_mix=args.chorus_mix, chorus_delay=args.chorus_delay, chorus_feedback=args.chorus_feedback, distortion_drive=args.drive_db, reverb_room_size=args.reverb_room_size, reverb_damping=args.reverb_damping, reverb_wet_level=args.reverb_wet_level, reverb_dry_level=args.reverb_dry_level, reverb_width=args.reverb_width, reverb_freeze_mode=args.reverb_freeze_mode, pitch_shift=args.pitch_shift, delay_seconds=args.delay_seconds, delay_feedback=args.delay_feedback, delay_mix=args.delay_mix, compressor_threshold=args.compressor_threshold, compressor_ratio=args.compressor_ratio, compressor_attack_ms=args.compressor_attack_ms, compressor_release_ms=args.compressor_release_ms, limiter_threshold=args.limiter_threshold, limiter_release=args.limiter_release, gain_db=args.gain_db, bitcrush_bit_depth=args.bitcrush_bit_depth, clipping_threshold=args.clipping_threshold, phaser_rate_hz=args.phaser_rate_hz, phaser_depth=args.phaser_depth, phaser_centre_frequency_hz=args.phaser_centre_frequency_hz, phaser_feedback=args.phaser_feedback, phaser_mix=args.phaser_mix, bass_boost_db=args.bass_boost_db, bass_boost_frequency=args.bass_boost_frequency, treble_boost_db=args.treble_boost_db, treble_boost_frequency=args.treble_boost_frequency, fade_in_duration=args.fade_in_duration, fade_out_duration=args.fade_out_duration, export_format=args.export_format, chorus=args.chorus, distortion=args.distortion, reverb=args.reverb, pitchshift=args.pitchshift, delay=args.delay, compressor=args.compressor, limiter=args.limiter, gain=args.gain, bitcrush=args.bitcrush, clipping=args.clipping, phaser=args.phaser, treble_bass_boost=args.treble_bass_boost, fade_in_out=args.fade_in_out)
+
+
+def process_audio(input_path, output_path, resample, resample_sr, chorus_depth, chorus_rate, chorus_mix, chorus_delay, chorus_feedback, distortion_drive, reverb_room_size, reverb_damping, reverb_wet_level, reverb_dry_level, reverb_width, reverb_freeze_mode, pitch_shift, delay_seconds, delay_feedback, delay_mix, compressor_threshold, compressor_ratio, compressor_attack_ms, compressor_release_ms, limiter_threshold, limiter_release, gain_db, bitcrush_bit_depth, clipping_threshold, phaser_rate_hz, phaser_depth, phaser_centre_frequency_hz, phaser_feedback, phaser_mix, bass_boost_db, bass_boost_frequency, treble_boost_db, treble_boost_frequency, fade_in_duration, fade_out_duration, export_format, chorus, distortion, reverb, pitchshift, delay, compressor, limiter, gain, bitcrush, clipping, phaser, treble_bass_boost, fade_in_out):
+    def bass_boost(audio, gain_db, frequency, sample_rate):
+        if gain_db >= 1:
+            b, a = butter(4, frequency / (0.5 * sample_rate), btype='low')
+
+            return filtfilt(b, a, audio) * 10 ** (gain_db / 20)
+        else: return audio
+    
+    def treble_boost(audio, gain_db, frequency, sample_rate):
+        if gain_db >=1:
+            b, a = butter(4, frequency / (0.5 * sample_rate), btype='high')
+            
+            return filtfilt(b, a, audio) * 10 ** (gain_db / 20)
+        else: return audio
+
+    def fade_out_effect(audio, sr, duration=3.0):
+        length = int(duration * sr)
+        end = audio.shape[0]
+        
+        if length > end: length = end  
+        start = end - length
+
+        audio[start:end] = audio[start:end] * np.linspace(1.0, 0.0, length)
+        return audio
+
+    def fade_in_effect(audio, sr, duration=3.0):
+        length = int(duration * sr)
+        start = 0
+
+        if length > audio.shape[0]: length = audio.shape[0]  
+        end = length
+        
+        audio[start:end] = audio[start:end] * np.linspace(0.0, 1.0, length)
+        return audio
+    
+    if not input_path or not os.path.exists(input_path): 
+        print(translations["input_not_valid"])
+        sys.exit(1)
+
+    if not output_path: 
+        print(translations["output_not_valid"])
+        sys.exit(1)
+    
+    if os.path.exists(output_path): os.remove(output_path)
+    
+    try:
+        audio, sample_rate = sf.read(input_path)
+    except Exception as e:
+        raise RuntimeError(translations["errors_loading_audio"].format(e=e))
+    
+    try:
+        board = Pedalboard()
+
+        if chorus: board.append(Chorus(depth=chorus_depth, rate_hz=chorus_rate, mix=chorus_mix, centre_delay_ms=chorus_delay, feedback=chorus_feedback))
+        if distortion: board.append(Distortion(drive_db=distortion_drive))
+        if reverb: board.append(Reverb(room_size=reverb_room_size, damping=reverb_damping, wet_level=reverb_wet_level, dry_level=reverb_dry_level, width=reverb_width, freeze_mode=1 if reverb_freeze_mode else 0))
+        if pitchshift: board.append(PitchShift(semitones=pitch_shift))
+        if delay: board.append(Delay(delay_seconds=delay_seconds, feedback=delay_feedback, mix=delay_mix))
+        if compressor: board.append(Compressor(threshold_db=compressor_threshold, ratio=compressor_ratio, attack_ms=compressor_attack_ms, release_ms=compressor_release_ms))
+        if limiter: board.append(Limiter(threshold_db=limiter_threshold, release_ms=limiter_release))
+        if gain: board.append(Gain(gain_db=gain_db))
+        if bitcrush: board.append(Bitcrush(bit_depth=bitcrush_bit_depth))
+        if clipping: board.append(Clipping(threshold_db=clipping_threshold)) 
+        if phaser: board.append(Phaser(rate_hz=phaser_rate_hz, depth=phaser_depth, centre_frequency_hz=phaser_centre_frequency_hz, feedback=phaser_feedback, mix=phaser_mix))
+
+        processed_audio = board(audio, sample_rate)
+
+        if treble_bass_boost:
+            processed_audio = bass_boost(processed_audio, bass_boost_db, bass_boost_frequency, sample_rate)
+            processed_audio = treble_boost(processed_audio, treble_boost_db, treble_boost_frequency, sample_rate)
+
+        if fade_in_out:
+            processed_audio = fade_in_effect(processed_audio, sample_rate, fade_in_duration)
+            processed_audio = fade_out_effect(processed_audio, sample_rate, fade_out_duration)
+            
+        if resample_sr != sample_rate and resample_sr > 0 and resample:
+            processed_audio = librosa.resample(processed_audio, orig_sr=sample_rate, target_sr=resample_sr)
+            sample_rate = resample_sr
+    except Exception as e:
+        raise RuntimeError(translations["apply_error"].format(e=e))
+
+    sf.write(output_path.replace(".wav", f".{export_format}"), processed_audio, sample_rate, format=export_format)
+    return output_path
+
+if __name__ == "__main__": main()

main/inference/convert.py

@@ -0,0 +1,1060 @@
+import gc
+import re
+import os
+import sys
+import time
+import torch
+import faiss
+import shutil
+import codecs
+import pyworld
+import librosa
+import logging
+import argparse
+import warnings
+import traceback
+import torchcrepe
+import subprocess
+import parselmouth
+import logging.handlers
+
+import numpy as np
+import soundfile as sf
+import noisereduce as nr
+import torch.nn.functional as F
+import torch.multiprocessing as mp
+
+from tqdm import tqdm
+from scipy import signal
+from torch import Tensor
+from scipy.io import wavfile
+from audio_upscaler import upscale
+from distutils.util import strtobool
+from fairseq import checkpoint_utils
+from pydub import AudioSegment, silence
+
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+from main.library.predictors.FCPE import FCPE
+from main.library.predictors.RMVPE import RMVPE
+from main.library.algorithm.synthesizers import Synthesizer
+
+
+warnings.filterwarnings("ignore", category=FutureWarning)
+warnings.filterwarnings("ignore", category=UserWarning)
+
+logging.getLogger("wget").setLevel(logging.ERROR)
+logging.getLogger("torch").setLevel(logging.ERROR)
+logging.getLogger("faiss").setLevel(logging.ERROR)
+logging.getLogger("httpx").setLevel(logging.ERROR)
+logging.getLogger("fairseq").setLevel(logging.ERROR)
+logging.getLogger("httpcore").setLevel(logging.ERROR)
+logging.getLogger("faiss.loader").setLevel(logging.ERROR)
+
+
+FILTER_ORDER = 5
+CUTOFF_FREQUENCY = 48  
+SAMPLE_RATE = 16000  
+
+bh, ah = signal.butter(N=FILTER_ORDER, Wn=CUTOFF_FREQUENCY, btype="high", fs=SAMPLE_RATE)
+input_audio_path2wav = {}
+
+log_file = os.path.join("assets", "logs", "convert.log")
+
+logger = logging.getLogger(__name__)
+logger.propagate = False
+
+translations = Config().translations
+
+
+if logger.hasHandlers(): logger.handlers.clear()
+else:
+    console_handler = logging.StreamHandler()
+    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    console_handler.setFormatter(console_formatter)
+    console_handler.setLevel(logging.INFO)
+
+    file_handler = logging.handlers.RotatingFileHandler(log_file, maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
+    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    file_handler.setFormatter(file_formatter)
+    file_handler.setLevel(logging.DEBUG)
+
+    logger.addHandler(console_handler)
+    logger.addHandler(file_handler)
+    logger.setLevel(logging.DEBUG)
+
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--pitch", type=int, default=0)
+    parser.add_argument("--filter_radius", type=int, default=3)
+    parser.add_argument("--index_rate", type=float, default=0.5)
+    parser.add_argument("--volume_envelope", type=float, default=1)
+    parser.add_argument("--protect", type=float, default=0.33)
+    parser.add_argument("--hop_length", type=int, default=64)
+    parser.add_argument( "--f0_method", type=str, default="rmvpe")
+    parser.add_argument("--input_path", type=str, required=True)
+    parser.add_argument("--output_path", type=str, default="./audios/output.wav")
+    parser.add_argument("--pth_path",  type=str,  required=True)
+    parser.add_argument("--index_path", type=str, required=True)
+    parser.add_argument("--f0_autotune", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--f0_autotune_strength", type=float, default=1)
+    parser.add_argument("--clean_audio", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--clean_strength", type=float, default=0.7)
+    parser.add_argument("--export_format", type=str, default="wav")
+    parser.add_argument("--embedder_model", type=str, default="contentvec_base")
+    parser.add_argument("--upscale_audio", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--resample_sr", type=int, default=0)
+    parser.add_argument("--batch_process", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--batch_size", type=int, default=2)
+    parser.add_argument("--split_audio", type=lambda x: bool(strtobool(x)), default=False)
+
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_arguments()
+    pitch = args.pitch
+    filter_radius = args.filter_radius
+    index_rate = args.index_rate
+    volume_envelope = args.volume_envelope 
+    protect = args.protect
+    hop_length = args.hop_length 
+    f0_method = args.f0_method 
+    input_path = args.input_path 
+    output_path = args.output_path 
+    pth_path = args.pth_path 
+    index_path = args.index_path 
+    f0_autotune = args.f0_autotune 
+    f0_autotune_strength = args.f0_autotune_strength 
+    clean_audio = args.clean_audio 
+    clean_strength = args.clean_strength 
+    export_format = args.export_format 
+    embedder_model = args.embedder_model 
+    upscale_audio = args.upscale_audio 
+    resample_sr = args.resample_sr 
+    batch_process = args.batch_process 
+    batch_size = args.batch_size 
+    split_audio = args.split_audio
+
+    logger.debug(f"{translations['pitch']}: {pitch}")
+    logger.debug(f"{translations['filter_radius']}: {filter_radius}")
+    logger.debug(f"{translations['index_strength']} {index_rate}")
+    logger.debug(f"{translations['volume_envelope']}: {volume_envelope}")
+    logger.debug(f"{translations['protect']}: {protect}")
+    if f0_method == "crepe" or f0_method == "crepe-tiny": logger.debug(f"Hop length: {hop_length}")
+    logger.debug(f"{translations['f0_method']}: {f0_method}")
+    logger.debug(f"f0_method: {input_path}")
+    logger.debug(f"{translations['audio_path']}: {input_path}")
+    logger.debug(f"{translations['output_path']}: {output_path.replace('.wav', f'.{export_format}')}")
+    logger.debug(f"{translations['model_path']}: {pth_path}")
+    logger.debug(f"{translations['indexpath']}: {index_path}")
+    logger.debug(f"{translations['autotune']}: {f0_autotune}")
+    logger.debug(f"{translations['clear_audio']}: {clean_audio}")
+    if clean_audio: logger.debug(f"{translations['clean_strength']}: {clean_strength}")
+    logger.debug(f"{translations['export_format']}: {export_format}")
+    logger.debug(f"{translations['hubert_model']}: {embedder_model}")
+    logger.debug(f"{translations['upscale_audio']}: {upscale_audio}")
+    if resample_sr != 0: logger.debug(f"{translations['sample_rate']}: {resample_sr}")
+    if split_audio: logger.debug(f"{translations['batch_process']}: {batch_process}")
+    if batch_process and split_audio: logger.debug(f"{translations['batch_size']}: {batch_size}")
+    logger.debug(f"{translations['split_audio']}: {split_audio}")
+    if f0_autotune: logger.debug(f"{translations['autotune_rate_info']}: {f0_autotune_strength}")
+
+
+    check_rmvpe_fcpe(f0_method)
+    check_hubert(embedder_model)
+
+    run_convert_script(pitch=pitch, filter_radius=filter_radius, index_rate=index_rate, volume_envelope=volume_envelope, protect=protect, hop_length=hop_length, f0_method=f0_method, input_path=input_path, output_path=output_path, pth_path=pth_path, index_path=index_path, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, clean_audio=clean_audio, clean_strength=clean_strength, export_format=export_format, embedder_model=embedder_model, upscale_audio=upscale_audio, resample_sr=resample_sr, batch_process=batch_process, batch_size=batch_size, split_audio=split_audio)
+
+
+def check_rmvpe_fcpe(method):
+    def download_rmvpe():
+        if not os.path.exists(os.path.join("assets", "model", "predictors", "rmvpe.pt")): subprocess.run(["wget", "-q", "--show-progress", "--no-check-certificate", codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Pbyno_EIP_Cebwrpg_2/erfbyir/znva/", "rot13") + "rmvpe.pt", "-P", os.path.join("assets", "model", "predictors")], check=True)
+
+    def download_fcpe():
+        if not os.path.exists(os.path.join("assets", "model", "predictors", "fcpe.pt")): subprocess.run(["wget", "-q", "--show-progress", "--no-check-certificate", codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Pbyno_EIP_Cebwrpg_2/erfbyir/znva/", "rot13") + "fcpe.pt", "-P", os.path.join("assets", "model", "predictors")], check=True)
+
+    if method == "rmvpe": download_rmvpe()
+    elif method == "fcpe": download_fcpe()
+    elif "hybrid" in method:
+        methods_str = re.search("hybrid\[(.+)\]", method)
+        if methods_str: methods = [method.strip() for method in methods_str.group(1).split("+")]
+
+        for method in methods:
+            if method == "rmvpe": download_rmvpe()
+            elif method == "fcpe": download_fcpe()
+
+
+def check_hubert(hubert):
+    if hubert == "contentvec_base" or hubert == "hubert_base" or hubert == "japanese_hubert_base" or hubert == "korean_hubert_base" or hubert == "chinese_hubert_base":
+        model_path = os.path.join(now_dir, "assets", "model", "embedders", hubert + '.pt')
+
+        if not os.path.exists(model_path): subprocess.run(["wget", "-q", "--show-progress", "--no-check-certificate", codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Pbyno_EIP_Cebwrpg_2/erfbyir/znva/", "rot13") + f"{hubert}.pt", "-P", os.path.join("assets", "model", "embedders")], check=True)
+
+
+def load_audio_infer(file, sample_rate):
+    try:
+        file = file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        if not os.path.isfile(file): raise FileNotFoundError(translations["not_found"].format(name=file))
+
+        audio, sr = sf.read(file)
+
+        if len(audio.shape) > 1: audio = librosa.to_mono(audio.T)
+        if sr != sample_rate: audio = librosa.resample(audio, orig_sr=sr, target_sr=sample_rate)
+    except Exception as e:
+        raise RuntimeError(f"{translations['errors_loading_audio']}: {e}") 
+     
+    return audio.flatten()
+
+
+def process_audio(file_path, output_path):
+    try:
+        song = AudioSegment.from_file(file_path)
+        nonsilent_parts = silence.detect_nonsilent(song, min_silence_len=750, silence_thresh=-70)
+
+        cut_files = []
+        time_stamps = []
+
+        min_chunk_duration = 30
+
+        for i, (start_i, end_i) in enumerate(nonsilent_parts):
+            chunk = song[start_i:end_i]
+
+            if len(chunk) >= min_chunk_duration:
+                chunk_file_path = os.path.join(output_path, f"chunk{i}.wav")
+
+                if os.path.exists(chunk_file_path): os.remove(chunk_file_path)
+                chunk.export(chunk_file_path, format="wav")
+
+                cut_files.append(chunk_file_path)
+                time_stamps.append((start_i, end_i))
+            else: logger.debug(translations["skip_file"].format(i=i, chunk=len(chunk)))
+
+        logger.info(f"{translations['split_total']}: {len(cut_files)}")
+        return cut_files, time_stamps
+    except Exception as e:
+        raise RuntimeError(f"{translations['process_audio_error']}: {e}")
+
+
+def merge_audio(files_list, time_stamps, original_file_path, output_path, format):
+    try:
+        def extract_number(filename):
+            match = re.search(r'_(\d+)', filename)
+
+            return int(match.group(1)) if match else 0
+
+        files_list = sorted(files_list, key=extract_number)
+        total_duration = len(AudioSegment.from_file(original_file_path))
+
+        combined = AudioSegment.empty() 
+        current_position = 0 
+
+        for file, (start_i, end_i) in zip(files_list, time_stamps):
+            if start_i > current_position:
+                silence_duration = start_i - current_position
+                combined += AudioSegment.silent(duration=silence_duration)  
+
+            combined += AudioSegment.from_file(file)  
+            current_position = end_i
+
+        if current_position < total_duration: combined += AudioSegment.silent(duration=total_duration - current_position)
+
+        combined.export(output_path, format=format)
+        return output_path
+    except Exception as e:
+        raise RuntimeError(f"{translations['merge_error']}: {e}")
+
+
+def run_batch_convert(params):
+    cvt = VoiceConverter()
+
+    path = params["path"]
+    audio_temp = params["audio_temp"]
+    export_format = params["export_format"]
+    cut_files = params["cut_files"]
+    pitch = params["pitch"]
+    filter_radius = params["filter_radius"]
+    index_rate = params["index_rate"]
+    volume_envelope = params["volume_envelope"]
+    protect = params["protect"]
+    hop_length = params["hop_length"]
+    f0_method = params["f0_method"]
+    pth_path = params["pth_path"]
+    index_path = params["index_path"]
+    f0_autotune = params["f0_autotune"]
+    f0_autotune_strength = params["f0_autotune_strength"]
+    clean_audio = params["clean_audio"]
+    clean_strength = params["clean_strength"]
+    upscale_audio = params["upscale_audio"]
+    embedder_model = params["embedder_model"]
+    resample_sr = params["resample_sr"]
+    
+
+    segment_output_path = os.path.join(audio_temp, f"output_{cut_files.index(path)}.{export_format}")
+    if os.path.exists(segment_output_path): os.remove(segment_output_path)
+
+    cvt.convert_audio(pitch=pitch, filter_radius=filter_radius, index_rate=index_rate, volume_envelope=volume_envelope, protect=protect, hop_length=hop_length, f0_method=f0_method, audio_input_path=path, audio_output_path=segment_output_path, model_path=pth_path, index_path=index_path, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, clean_audio=clean_audio, clean_strength=clean_strength, export_format=export_format, upscale_audio=upscale_audio, embedder_model=embedder_model, resample_sr=resample_sr)
+    os.remove(path)
+
+
+    if os.path.exists(segment_output_path): return segment_output_path
+    else: 
+        logger.warning(f"{translations['not_found_convert_file']}: {segment_output_path}")
+        sys.exit(1)
+
+
+def run_convert_script(pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0_method, input_path, output_path, pth_path, index_path, f0_autotune, f0_autotune_strength, clean_audio, clean_strength, export_format, upscale_audio, embedder_model, resample_sr, batch_process, batch_size, split_audio):
+    cvt = VoiceConverter()
+    start_time = time.time()
+
+
+    if not pth_path or not os.path.exists(pth_path) or os.path.isdir(pth_path) or not pth_path.endswith(".pth"):
+        logger.warning(translations["provide_file"].format(filename=translations["model"]))
+        sys.exit(1)
+    
+    if not index_path or not os.path.exists(index_path) or os.path.isdir(index_path) or not index_path.endswith(".index"):
+        logger.warning(translations["provide_file"].format(filename=translations["index"]))
+        sys.exit(1)
+
+
+    output_dir = os.path.dirname(output_path)
+    output_dir = output_path if not output_dir else output_dir
+
+    if output_dir is None: output_dir = "audios"
+
+    if not os.path.exists(output_dir): os.makedirs(output_dir, exist_ok=True)
+
+    audio_temp = os.path.join("audios_temp")
+    if not os.path.exists(audio_temp) and split_audio: os.makedirs(audio_temp, exist_ok=True)
+
+    processed_segments = []
+
+    if os.path.isdir(input_path):
+        try:
+            logger.info(translations["convert_batch"])
+
+            audio_files = [f for f in os.listdir(input_path) if f.endswith(("wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"))]
+            if not audio_files: 
+                logger.warning(translations["not_found_audio"])
+                sys.exit(1)
+
+            logger.info(translations["found_audio"].format(audio_files=len(audio_files)))
+
+            for audio in audio_files:
+                audio_path = os.path.join(input_path, audio)
+                output_audio = os.path.join(input_path, os.path.splitext(audio)[0] + f"_output.{export_format}")
+
+                if split_audio:
+                    try:
+                        cut_files, time_stamps = process_audio(audio_path, audio_temp)
+                        num_threads = min(batch_size, len(cut_files))
+
+                        params_list = [
+                            {
+                                "path": path,
+                                "audio_temp": audio_temp,
+                                "export_format": export_format,
+                                "cut_files": cut_files,
+                                "pitch": pitch,
+                                "filter_radius": filter_radius,
+                                "index_rate": index_rate,
+                                "volume_envelope": volume_envelope,
+                                "protect": protect,
+                                "hop_length": hop_length,
+                                "f0_method": f0_method,
+                                "pth_path": pth_path,
+                                "index_path": index_path,
+                                "f0_autotune": f0_autotune,
+                                "f0_autotune_strength": f0_autotune_strength,
+                                "clean_audio": clean_audio,
+                                "clean_strength": clean_strength,
+                                "upscale_audio": upscale_audio,
+                                "embedder_model": embedder_model,
+                                "resample_sr": resample_sr
+                            }
+                            for path in cut_files
+                        ]
+
+                        if batch_process:
+                            with mp.Pool(processes=num_threads) as pool:
+                                with tqdm(total=len(params_list), desc=translations["convert_audio"]) as pbar:
+                                    for results in pool.imap_unordered(run_batch_convert, params_list):
+                                        processed_segments.append(results)
+                                        pbar.update(1)
+                        else: 
+                            for params in tqdm(params_list, desc=translations["convert_audio"]):
+                                run_batch_convert(params)
+
+                        merge_audio(processed_segments, time_stamps, audio_path, output_audio, export_format)
+                    except Exception as e:
+                        logger.error(translations["error_convert_batch"].format(e=e))
+                    finally:
+                        if os.path.exists(audio_temp): shutil.rmtree(audio_temp, ignore_errors=True)
+                else:
+                    try:
+                        logger.info(f"{translations['convert_audio']} '{audio_path}'...")
+
+                        if os.path.exists(output_audio): os.remove(output_audio)
+
+                        with tqdm(total=1, desc=translations["convert_audio"]) as pbar:
+                            cvt.convert_audio(pitch=pitch, filter_radius=filter_radius, index_rate=index_rate, volume_envelope=volume_envelope, protect=protect, hop_length=hop_length, f0_method=f0_method, audio_input_path=audio_path, audio_output_path=output_audio, model_path=pth_path, index_path=index_path, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, clean_audio=clean_audio, clean_strength=clean_strength, export_format=export_format, upscale_audio=upscale_audio, embedder_model=embedder_model, resample_sr=resample_sr)
+                            pbar.update(1)
+                    except Exception as e:
+                        logger.error(translations["error_convert"].format(e=e))
+
+            elapsed_time = time.time() - start_time
+            logger.info(translations["convert_batch_success"].format(elapsed_time=f"{elapsed_time:.2f}", output_path=output_path.replace('.wav', f'.{export_format}')))
+        except Exception as e:
+            logger.error(translations["error_convert_batch_2"].format(e=e))
+    else:
+        logger.info(f"{translations['convert_audio']} '{input_path}'...")
+
+        if not os.path.exists(input_path):
+            logger.warning(translations["not_found_audio"])
+            sys.exit(1)
+        
+        if os.path.isdir(output_path): output_path = os.path.join(output_path, f"output.{export_format}")
+        if os.path.exists(output_path): os.remove(output_path)
+
+        if split_audio:
+            try:              
+                cut_files, time_stamps = process_audio(input_path, audio_temp)
+                num_threads = min(batch_size, len(cut_files))
+
+                params_list = [
+                    {
+                        "path": path,
+                        "audio_temp": audio_temp,
+                        "export_format": export_format,
+                        "cut_files": cut_files,
+                        "pitch": pitch,
+                        "filter_radius": filter_radius,
+                        "index_rate": index_rate,
+                        "volume_envelope": volume_envelope,
+                        "protect": protect,
+                        "hop_length": hop_length,
+                        "f0_method": f0_method,
+                        "pth_path": pth_path,
+                        "index_path": index_path,
+                        "f0_autotune": f0_autotune,
+                        "f0_autotune_strength": f0_autotune_strength,
+                        "clean_audio": clean_audio,
+                        "clean_strength": clean_strength,
+                        "upscale_audio": upscale_audio,
+                        "embedder_model": embedder_model,
+                        "resample_sr": resample_sr
+                    }
+                    for path in cut_files
+                ]
+
+                if batch_process:
+                    with mp.Pool(processes=num_threads) as pool:
+                        with tqdm(total=len(params_list), desc=translations["convert_audio"]) as pbar:
+                            for results in pool.imap_unordered(run_batch_convert, params_list):
+                                processed_segments.append(results)
+                                pbar.update(1)
+                else: 
+                    for params in tqdm(params_list, desc=translations["convert_audio"]):
+                        run_batch_convert(params)
+
+                merge_audio(processed_segments, time_stamps, input_path, output_path.replace(".wav", f".{export_format}"), export_format)
+            except Exception as e:
+                logger.error(translations["error_convert_batch"].format(e=e))
+            finally:
+                if os.path.exists(audio_temp): shutil.rmtree(audio_temp, ignore_errors=True)
+        else:
+            try:
+                with tqdm(total=1, desc=translations["convert_audio"]) as pbar:
+                    cvt.convert_audio(pitch=pitch, filter_radius=filter_radius, index_rate=index_rate, volume_envelope=volume_envelope, protect=protect, hop_length=hop_length, f0_method=f0_method, audio_input_path=input_path, audio_output_path=output_path, model_path=pth_path, index_path=index_path, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, clean_audio=clean_audio, clean_strength=clean_strength, export_format=export_format, upscale_audio=upscale_audio, embedder_model=embedder_model, resample_sr=resample_sr)
+                    pbar.update(1)
+            except Exception as e:
+                logger.error(translations["error_convert"].format(e=e))
+
+        elapsed_time = time.time() - start_time
+        logger.info(translations["convert_audio_success"].format(input_path=input_path, elapsed_time=f"{elapsed_time:.2f}", output_path=output_path.replace('.wav', f'.{export_format}')))
+
+
+def change_rms(source_audio: np.ndarray, source_rate: int, target_audio: np.ndarray, target_rate: int, rate: float) -> np.ndarray:
+    rms1 = librosa.feature.rms(
+        y=source_audio,
+        frame_length=source_rate // 2 * 2,
+        hop_length=source_rate // 2,
+    )
+    
+    rms2 = librosa.feature.rms(
+        y=target_audio,
+        frame_length=target_rate // 2 * 2,
+        hop_length=target_rate // 2,
+    )
+
+    rms1 = F.interpolate(
+        torch.from_numpy(rms1).float().unsqueeze(0),
+        size=target_audio.shape[0],
+        mode="linear",
+    ).squeeze()
+
+    rms2 = F.interpolate(
+        torch.from_numpy(rms2).float().unsqueeze(0),
+        size=target_audio.shape[0],
+        mode="linear",
+    ).squeeze()
+
+    rms2 = torch.maximum(rms2, torch.zeros_like(rms2) + 1e-6)
+
+
+    adjusted_audio = (target_audio * (torch.pow(rms1, 1 - rate) * torch.pow(rms2, rate - 1)).numpy())
+    return adjusted_audio
+
+
+class Autotune:
+    def __init__(self, ref_freqs):
+        self.ref_freqs = ref_freqs
+        self.note_dict = self.ref_freqs
+
+
+    def autotune_f0(self, f0, f0_autotune_strength):
+        autotuned_f0 = np.zeros_like(f0)
+
+
+        for i, freq in enumerate(f0):
+            closest_note = min(self.note_dict, key=lambda x: abs(x - freq))
+            autotuned_f0[i] = freq + (closest_note - freq) * f0_autotune_strength
+
+        return autotuned_f0
+
+
+class VC:
+    def __init__(self, tgt_sr, config):
+        self.x_pad = config.x_pad
+        self.x_query = config.x_query
+        self.x_center = config.x_center
+        self.x_max = config.x_max
+        self.is_half = config.is_half
+        self.sample_rate = 16000
+        self.window = 160
+        self.t_pad = self.sample_rate * self.x_pad
+        self.t_pad_tgt = tgt_sr * self.x_pad
+        self.t_pad2 = self.t_pad * 2
+        self.t_query = self.sample_rate * self.x_query
+        self.t_center = self.sample_rate * self.x_center
+        self.t_max = self.sample_rate * self.x_max
+        self.time_step = self.window / self.sample_rate * 1000
+        self.f0_min = 50
+        self.f0_max = 1100
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+        self.device = config.device
+        self.ref_freqs = [
+            49.00,  
+            51.91,  
+            55.00, 
+            58.27,  
+            61.74,  
+            65.41, 
+            69.30, 
+            73.42, 
+            77.78,  
+            82.41, 
+            87.31, 
+            92.50, 
+            98.00,  
+            103.83,  
+            110.00, 
+            116.54, 
+            123.47, 
+            130.81, 
+            138.59,  
+            146.83,  
+            155.56,  
+            164.81, 
+            174.61, 
+            185.00,  
+            196.00,  
+            207.65, 
+            220.00,  
+            233.08,  
+            246.94, 
+            261.63, 
+            277.18,  
+            293.66, 
+            311.13, 
+            329.63,  
+            349.23, 
+            369.99, 
+            392.00, 
+            415.30,  
+            440.00,  
+            466.16,  
+            493.88, 
+            523.25,  
+            554.37, 
+            587.33,  
+            622.25, 
+            659.25, 
+            698.46, 
+            739.99,  
+            783.99,  
+            830.61, 
+            880.00, 
+            932.33,  
+            987.77, 
+            1046.50
+        ]
+        self.autotune = Autotune(self.ref_freqs)
+        self.note_dict = self.autotune.note_dict
+
+
+    def get_f0_crepe(self, x, f0_min, f0_max, p_len, hop_length, model="full"):
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+
+        audio = torch.from_numpy(x).to(self.device, copy=True)
+        audio = torch.unsqueeze(audio, dim=0)
+
+
+        if audio.ndim == 2 and audio.shape[0] > 1: audio = torch.mean(audio, dim=0, keepdim=True).detach()
+
+        audio = audio.detach()
+        pitch: Tensor = torchcrepe.predict(audio, self.sample_rate, hop_length, f0_min, f0_max, model, batch_size=hop_length * 2, device=self.device, pad=True)
+
+        p_len = p_len or x.shape[0] // hop_length
+        source = np.array(pitch.squeeze(0).cpu().float().numpy())
+        source[source < 0.001] = np.nan
+        
+        target = np.interp(
+            np.arange(0, len(source) * p_len, len(source)) / p_len,
+            np.arange(0, len(source)),
+            source,
+        )
+
+        f0 = np.nan_to_num(target)
+        return f0
+
+
+    def get_f0_hybrid(self, methods_str, x, f0_min, f0_max, p_len, hop_length, filter_radius):
+        methods_str = re.search("hybrid\[(.+)\]", methods_str)
+        if methods_str: methods = [method.strip() for method in methods_str.group(1).split("+")]
+
+        f0_computation_stack = []
+        logger.debug(translations["hybrid_methods"].format(methods=methods))
+
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+
+
+        for method in methods:
+            f0 = None
+
+
+            if method == "pm":
+                f0 = (parselmouth.Sound(x, self.sample_rate).to_pitch_ac(time_step=self.window / self.sample_rate * 1000 / 1000, voicing_threshold=0.6, pitch_floor=self.f0_min, pitch_ceiling=self.f0_max).selected_array["frequency"])
+                pad_size = (p_len - len(f0) + 1) // 2
+
+                if pad_size > 0 or p_len - len(f0) - pad_size > 0: f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+            elif method == 'dio':
+                f0, t = pyworld.dio(x.astype(np.double), fs=self.sample_rate, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=10)
+                f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sample_rate)
+
+                f0 = signal.medfilt(f0, 3)
+            elif method == "crepe-tiny":
+                f0 = self.get_f0_crepe(x, self.f0_min, self.f0_max, p_len, int(hop_length), "tiny")
+            elif method == "crepe": 
+                f0 = self.get_f0_crepe(x, f0_min, f0_max, p_len, int(hop_length))
+            elif method == "fcpe":
+                self.model_fcpe = FCPE(os.path.join("assets", "model", "predictors", "fcpe.pt"), hop_length=int(hop_length), f0_min=int(f0_min), f0_max=int(f0_max), dtype=torch.float32, device=self.device, sample_rate=self.sample_rate, threshold=0.03)
+                f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+
+                del self.model_fcpe
+                gc.collect() 
+            elif method == "rmvpe":
+                f0 = RMVPE(os.path.join("assets", "model", "predictors", "rmvpe.pt"), is_half=self.is_half, device=self.device).infer_from_audio(x, thred=0.03)
+                f0 = f0[1:]
+            elif method == "harvest":
+                f0, t = pyworld.harvest(x.astype(np.double), fs=self.sample_rate, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=10)
+                f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sample_rate)
+
+                if filter_radius > 2: f0 = signal.medfilt(f0, 3)
+            else: raise ValueError(translations["method_not_valid"])
+   
+            f0_computation_stack.append(f0)
+            
+        resampled_stack = []
+
+        for f0 in f0_computation_stack:
+            resampled_f0 = np.interp(np.linspace(0, len(f0), p_len), np.arange(len(f0)), f0)
+            resampled_stack.append(resampled_f0)
+
+        f0_median_hybrid = resampled_stack[0] if len(resampled_stack) == 1 else np.nanmedian(np.vstack(resampled_stack), axis=0)
+        return f0_median_hybrid
+
+
+    def get_f0(self, input_audio_path, x, p_len, pitch, f0_method, filter_radius, hop_length, f0_autotune, f0_autotune_strength):
+        global input_audio_path2wav
+        
+
+        if f0_method == "pm":
+            f0 = (parselmouth.Sound(x, self.sample_rate).to_pitch_ac(time_step=self.window / self.sample_rate * 1000 / 1000, voicing_threshold=0.6, pitch_floor=self.f0_min, pitch_ceiling=self.f0_max).selected_array["frequency"])
+            pad_size = (p_len - len(f0) + 1) // 2
+
+            if pad_size > 0 or p_len - len(f0) - pad_size > 0: f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+        elif f0_method == "dio":
+            f0, t = pyworld.dio(x.astype(np.double), fs=self.sample_rate, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=10)
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sample_rate)
+
+            f0 = signal.medfilt(f0, 3)
+        elif f0_method == "crepe-tiny":
+            f0 = self.get_f0_crepe(x, self.f0_min, self.f0_max, p_len, int(hop_length), "tiny")
+        elif f0_method == "crepe":
+            f0 = self.get_f0_crepe(x, self.f0_min, self.f0_max, p_len, int(hop_length))
+        elif f0_method == "fcpe":
+            self.model_fcpe = FCPE(os.path.join("assets", "model", "predictors", "fcpe.pt"), hop_length=int(hop_length), f0_min=int(self.f0_min), f0_max=int(self.f0_max), dtype=torch.float32, device=self.device, sample_rate=self.sample_rate, threshold=0.03)
+            f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+
+            del self.model_fcpe
+            gc.collect()
+        elif f0_method == "rmvpe":
+            f0 = RMVPE(os.path.join("assets", "model", "predictors", "rmvpe.pt"), is_half=self.is_half, device=self.device).infer_from_audio(x, thred=0.03)
+        elif f0_method == "harvest":
+            f0, t = pyworld.harvest(x.astype(np.double), fs=self.sample_rate, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=10)
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sample_rate)
+
+            if filter_radius > 2: f0 = signal.medfilt(f0, 3)
+        elif "hybrid" in f0_method:
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = self.get_f0_hybrid(f0_method, x, self.f0_min, self.f0_max, p_len, hop_length, filter_radius)
+        else: raise ValueError(translations["method_not_valid"])
+
+        if f0_autotune: f0 = Autotune.autotune_f0(self, f0, f0_autotune_strength)
+
+        f0 *= pow(2, pitch / 12)
+
+        f0bak = f0.copy()
+
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * 254 / (self.f0_mel_max - self.f0_mel_min) + 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > 255] = 255
+
+        f0_coarse = np.rint(f0_mel).astype(np.int32)
+        return f0_coarse, f0bak
+
+
+    def voice_conversion(self, model, net_g, sid, audio0, pitch, pitchf, index, big_npy, index_rate, version, protect):
+        pitch_guidance = pitch != None and pitchf != None
+
+        feats = (torch.from_numpy(audio0).half() if self.is_half else torch.from_numpy(audio0).float())
+
+        if feats.dim() == 2: feats = feats.mean(-1)
+        assert feats.dim() == 1, feats.dim()
+
+        feats = feats.view(1, -1)
+
+        padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+
+        inputs = {
+            "source": feats.to(self.device),
+            "padding_mask": padding_mask,
+            "output_layer": 9 if version == "v1" else 12,
+        }
+
+        with torch.no_grad():
+            logits = model.extract_features(**inputs)
+            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+
+        if protect < 0.5 and pitch_guidance: feats0 = feats.clone()
+
+        if (not isinstance(index, type(None)) and not isinstance(big_npy, type(None)) and index_rate != 0):
+            npy = feats[0].cpu().numpy()
+
+            if self.is_half: npy = npy.astype("float32")
+
+            score, ix = index.search(npy, k=8)
+
+            weight = np.square(1 / score)
+            weight /= weight.sum(axis=1, keepdims=True)
+
+            npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
+
+            if self.is_half: npy = npy.astype("float16")
+
+            feats = (torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate + (1 - index_rate) * feats)
+
+        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+
+        if protect < 0.5 and pitch_guidance: feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+
+        p_len = audio0.shape[0] // self.window
+
+        if feats.shape[1] < p_len:
+            p_len = feats.shape[1]
+
+            if pitch_guidance:
+                pitch = pitch[:, :p_len]
+                pitchf = pitchf[:, :p_len]
+
+        if protect < 0.5 and pitch_guidance:
+            pitchff = pitchf.clone()
+            pitchff[pitchf > 0] = 1
+            pitchff[pitchf < 1] = protect
+            pitchff = pitchff.unsqueeze(-1)
+            
+            feats = feats * pitchff + feats0 * (1 - pitchff)
+            feats = feats.to(feats0.dtype)
+
+        p_len = torch.tensor([p_len], device=self.device).long()
+
+        with torch.no_grad():
+            audio1 = ((net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0]).data.cpu().float().numpy()) if pitch_guidance else ((net_g.infer(feats, p_len, sid)[0][0, 0]).data.cpu().float().numpy())
+
+        del feats, p_len, padding_mask
+
+        if torch.cuda.is_available(): torch.cuda.empty_cache()
+        return audio1
+    
+
+    def pipeline(self, model, net_g, sid, audio, input_audio_path, pitch, f0_method, file_index, index_rate, pitch_guidance, filter_radius, tgt_sr, resample_sr, volume_envelope, version, protect, hop_length, f0_autotune, f0_autotune_strength):
+        if file_index != "" and os.path.exists(file_index) and index_rate != 0:
+            try:
+                index = faiss.read_index(file_index)
+                big_npy = index.reconstruct_n(0, index.ntotal)
+            except Exception as e:
+                logger.error(translations["read_faiss_index_error"].format(e=e))
+                index = big_npy = None
+        else: index = big_npy = None
+
+        audio = signal.filtfilt(bh, ah, audio)
+        audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
+        opt_ts = []
+
+        if audio_pad.shape[0] > self.t_max:
+            audio_sum = np.zeros_like(audio)
+
+            for i in range(self.window):
+                audio_sum += audio_pad[i : i - self.window]
+
+            for t in range(self.t_center, audio.shape[0], self.t_center):
+                opt_ts.append(t - self.t_query + np.where(np.abs(audio_sum[t - self.t_query : t + self.t_query]) == np.abs(audio_sum[t - self.t_query : t + self.t_query]).min())[0][0])
+
+        s = 0
+        audio_opt = []
+        t = None
+
+        audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
+        p_len = audio_pad.shape[0] // self.window
+
+        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
+
+        if pitch_guidance:
+            pitch, pitchf = self.get_f0(input_audio_path, audio_pad, p_len, pitch, f0_method, filter_radius, hop_length, f0_autotune, f0_autotune_strength)
+            pitch = pitch[:p_len]
+            pitchf = pitchf[:p_len]
+
+            if self.device == "mps": pitchf = pitchf.astype(np.float32)
+
+            pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+            pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
+
+        for t in opt_ts:
+            t = t // self.window * self.window
+
+            if pitch_guidance: audio_opt.append(self.voice_conversion(model, net_g, sid, audio_pad[s : t + self.t_pad2 + self.window], pitch[:, s // self.window : (t + self.t_pad2) // self.window], pitchf[:, s // self.window : (t + self.t_pad2) // self.window], index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])
+            else: audio_opt.append(self.voice_conversion(model, net_g, sid, audio_pad[s : t + self.t_pad2 + self.window], None, None, index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])
+
+            s = t
+            
+        if pitch_guidance: audio_opt.append(self.voice_conversion(model, net_g, sid, audio_pad[t:], pitch[:, t // self.window :] if t is not None else pitch, pitchf[:, t // self.window :] if t is not None else pitchf, index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])
+        else: audio_opt.append(self.voice_conversion(model, net_g, sid, audio_pad[t:], None, None, index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])
+            
+        audio_opt = np.concatenate(audio_opt)
+
+        if volume_envelope != 1: audio_opt = change_rms(audio, self.sample_rate, audio_opt, tgt_sr, volume_envelope)
+        if resample_sr >= self.sample_rate and tgt_sr != resample_sr: audio_opt = librosa.resample(audio_opt, orig_sr=tgt_sr, target_sr=resample_sr)
+
+        audio_max = np.abs(audio_opt).max() / 0.99
+        max_int16 = 32768
+
+        if audio_max > 1: max_int16 /= audio_max
+
+        audio_opt = (audio_opt * max_int16).astype(np.int16)
+
+        if pitch_guidance: del pitch, pitchf
+        del sid
+
+        if torch.cuda.is_available(): torch.cuda.empty_cache()
+        return audio_opt
+
+
+class VoiceConverter:
+    def __init__(self):
+        self.config = Config()  
+        self.hubert_model = (None)
+
+        self.tgt_sr = None 
+        self.net_g = None 
+
+        self.vc = None
+        self.cpt = None  
+
+        self.version = None 
+        self.n_spk = None  
+
+        self.use_f0 = None  
+        self.loaded_model = None
+    
+
+    def load_hubert(self, embedder_model):
+        try:
+            models, _, _ = checkpoint_utils.load_model_ensemble_and_task([os.path.join(now_dir, "assets", "model", "embedders", embedder_model + '.pt')], suffix="")
+        except Exception as e:
+            raise ImportError(translations["read_model_error"].format(e=e))
+        
+        self.hubert_model = models[0].to(self.config.device)
+        self.hubert_model = (self.hubert_model.half() if self.config.is_half else self.hubert_model.float())
+        self.hubert_model.eval()
+
+
+    @staticmethod
+    def remove_audio_noise(input_audio_path, reduction_strength=0.7):
+        try:
+            rate, data = wavfile.read(input_audio_path)
+            reduced_noise = nr.reduce_noise(y=data, sr=rate, prop_decrease=reduction_strength)
+
+            return reduced_noise
+        except Exception as e:
+            logger.error(translations["denoise_error"].format(e=e))
+            return None
+
+
+    @staticmethod
+    def convert_audio_format(input_path, output_path, output_format):
+        try:
+            if output_format != "wav":
+                logger.debug(translations["change_format"].format(output_format=output_format))
+                audio, sample_rate = sf.read(input_path)
+
+
+                common_sample_rates = [
+                    8000, 
+                    11025, 
+                    12000, 
+                    16000, 
+                    22050, 
+                    24000, 
+                    32000, 
+                    44100, 
+                    48000
+                ]
+
+                target_sr = min(common_sample_rates, key=lambda x: abs(x - sample_rate))
+                audio = librosa.resample(audio, orig_sr=sample_rate, target_sr=target_sr)
+
+                sf.write(output_path, audio, target_sr, format=output_format)
+
+            return output_path
+        except Exception as e:
+            raise RuntimeError(translations["change_format_error"].format(e=e))
+
+
+    def convert_audio(self, audio_input_path, audio_output_path, model_path, index_path, embedder_model, pitch, f0_method, index_rate, volume_envelope, protect, hop_length, f0_autotune, f0_autotune_strength, filter_radius, clean_audio, clean_strength, export_format, upscale_audio, resample_sr = 0, sid = 0):
+        self.get_vc(model_path, sid)
+
+        try:
+            if upscale_audio: upscale(audio_input_path, audio_input_path)
+
+            audio = load_audio_infer(audio_input_path, 16000)
+
+            audio_max = np.abs(audio).max() / 0.95
+
+
+            if audio_max > 1: audio /= audio_max
+
+            if not self.hubert_model: 
+                if not os.path.exists(os.path.join(now_dir, "assets", "model", "embedders", embedder_model + '.pt')): raise FileNotFoundError(f"Không tìm thấy mô hình: {embedder_model}")
+                
+                self.load_hubert(embedder_model)
+
+            if self.tgt_sr != resample_sr >= 16000: self.tgt_sr = resample_sr
+
+            file_index = (index_path.strip().strip('"').strip("\n").strip('"').strip().replace("trained", "added"))
+
+            audio_opt = self.vc.pipeline(model=self.hubert_model, net_g=self.net_g, sid=sid, audio=audio, input_audio_path=audio_input_path, pitch=pitch, f0_method=f0_method, file_index=file_index, index_rate=index_rate, pitch_guidance=self.use_f0, filter_radius=filter_radius, tgt_sr=self.tgt_sr, resample_sr=resample_sr, volume_envelope=volume_envelope, version=self.version, protect=protect, hop_length=hop_length, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength)
+
+            if audio_output_path: sf.write(audio_output_path, audio_opt, self.tgt_sr, format="wav")
+
+            if clean_audio:
+                cleaned_audio = self.remove_audio_noise(audio_output_path, clean_strength)
+                if cleaned_audio is not None: sf.write(audio_output_path, cleaned_audio, self.tgt_sr, format="wav")
+
+            output_path_format = audio_output_path.replace(".wav", f".{export_format}")
+            audio_output_path = self.convert_audio_format(audio_output_path, output_path_format, export_format)
+        except Exception as e:
+            logger.error(translations["error_convert"].format(e=e))
+            logger.error(traceback.format_exc())
+
+
+    def get_vc(self, weight_root, sid):
+        if sid == "" or sid == []:
+            self.cleanup_model()
+            if torch.cuda.is_available(): torch.cuda.empty_cache()
+
+        if not self.loaded_model or self.loaded_model != weight_root:
+          self.load_model(weight_root)
+
+          if self.cpt is not None:
+              self.setup_network()
+              self.setup_vc_instance()
+
+          self.loaded_model = weight_root
+
+
+    def cleanup_model(self):
+        if self.hubert_model is not None:
+            del self.net_g, self.n_spk, self.vc, self.hubert_model, self.tgt_sr
+
+            self.hubert_model = self.net_g = self.n_spk = self.vc = self.tgt_sr = None
+
+            if torch.cuda.is_available(): torch.cuda.empty_cache()
+
+        del self.net_g, self.cpt
+
+        if torch.cuda.is_available(): torch.cuda.empty_cache()
+        self.cpt = None
+
+
+    def load_model(self, weight_root):
+        self.cpt = (torch.load(weight_root, map_location="cpu") if os.path.isfile(weight_root) else None)
+
+
+    def setup_network(self):
+        if self.cpt is not None:
+            self.tgt_sr = self.cpt["config"][-1]
+            self.cpt["config"][-3] = self.cpt["weight"]["emb_g.weight"].shape[0]
+            self.use_f0 = self.cpt.get("f0", 1)
+
+            self.version = self.cpt.get("version", "v1")
+            self.text_enc_hidden_dim = 768 if self.version == "v2" else 256
+
+            self.net_g = Synthesizer(*self.cpt["config"], use_f0=self.use_f0, text_enc_hidden_dim=self.text_enc_hidden_dim, is_half=self.config.is_half)
+
+            del self.net_g.enc_q
+
+            self.net_g.load_state_dict(self.cpt["weight"], strict=False)
+            self.net_g.eval().to(self.config.device)
+            self.net_g = (self.net_g.half() if self.config.is_half else self.net_g.float())
+
+
+    def setup_vc_instance(self):
+        if self.cpt is not None:
+            self.vc = VC(self.tgt_sr, self.config)
+            self.n_spk = self.cpt["config"][-3]
+
+if __name__ == "__main__": 
+    mp.set_start_method("spawn", force=True)
+    main()

main/inference/create_dataset.py

@@ -0,0 +1,370 @@
+import os
+import re
+import sys
+import time
+import yt_dlp
+import shutil
+import librosa
+import logging
+import argparse
+import warnings
+import logging.handlers
+
+import soundfile as sf
+import noisereduce as nr
+
+from distutils.util import strtobool
+from pydub import AudioSegment, silence
+
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+from main.library.algorithm.separator import Separator
+
+
+translations = Config().translations
+
+
+log_file = os.path.join("assets", "logs", "create_dataset.log")
+logger = logging.getLogger(__name__)
+
+if logger.hasHandlers(): logger.handlers.clear()
+else: 
+    console_handler = logging.StreamHandler()
+    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    console_handler.setFormatter(console_formatter)
+    console_handler.setLevel(logging.INFO)
+
+    file_handler = logging.handlers.RotatingFileHandler(log_file, maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
+    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    file_handler.setFormatter(file_formatter)
+    file_handler.setLevel(logging.DEBUG)
+
+    logger.addHandler(console_handler)
+    logger.addHandler(file_handler)
+    logger.setLevel(logging.DEBUG)
+
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input_audio", type=str, required=True)
+    parser.add_argument("--output_dataset", type=str, default="./dataset")
+    parser.add_argument("--resample", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--resample_sr", type=int, default=44100)
+    parser.add_argument("--clean_dataset", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--clean_strength", type=float, default=0.7)
+    parser.add_argument("--separator_music", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--separator_reverb", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--kim_vocal_version", type=int, default=2)
+    parser.add_argument("--overlap", type=float, default=0.25)
+    parser.add_argument("--segments_size", type=int, default=256)
+    parser.add_argument("--mdx_hop_length", type=int, default=1024)
+    parser.add_argument("--mdx_batch_size", type=int, default=1)
+    parser.add_argument("--denoise_mdx", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--skip", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--skip_start_audios", type=str, default="0")
+    parser.add_argument("--skip_end_audios", type=str, default="0")
+    
+    args = parser.parse_args()
+    return args
+
+
+dataset_temp = os.path.join("dataset_temp")
+
+
+def main():
+    args = parse_arguments()
+    input_audio = args.input_audio
+    output_dataset = args.output_dataset
+    resample = args.resample
+    resample_sr = args.resample_sr
+    clean_dataset = args.clean_dataset
+    clean_strength = args.clean_strength
+    separator_music = args.separator_music
+    separator_reverb = args.separator_reverb
+    kim_vocal_version = args.kim_vocal_version
+    overlap = args.overlap
+    segments_size = args.segments_size
+    hop_length = args.mdx_hop_length
+    batch_size = args.mdx_batch_size
+    denoise_mdx = args.denoise_mdx
+    skip = args.skip
+    skip_start_audios = args.skip_start_audios
+    skip_end_audios = args.skip_end_audios
+
+    logger.debug(f"{translations['audio_path']}: {input_audio}")
+    logger.debug(f"{translations['output_path']}: {output_dataset}")
+    logger.debug(f"{translations['resample']}: {resample}")
+    if resample: logger.debug(f"{translations['sample_rate']}: {resample_sr}")
+    logger.debug(f"{translations['clear_dataset']}: {clean_dataset}")
+    if clean_dataset: logger.debug(f"{translations['clean_strength']}: {clean_strength}")
+    logger.debug(f"{translations['separator_music']}: {separator_music}")
+    logger.debug(f"{translations['dereveb_audio']}: {separator_reverb}")
+    if separator_music: logger.debug(f"{translations['training_version']}: {kim_vocal_version}")
+    logger.debug(f"{translations['segments_size']}: {segments_size}")
+    logger.debug(f"{translations['overlap']}: {overlap}")
+    logger.debug(f"Hop length: {hop_length}")
+    logger.debug(f"{translations['batch_size']}: {batch_size}")
+    logger.debug(f"{translations['denoise_mdx']}: {denoise_mdx}")
+    logger.debug(f"{translations['skip']}: {skip}")
+    if skip: logger.debug(f"{translations['skip_start']}: {skip_start_audios}")
+    if skip: logger.debug(f"{translations['skip_end']}: {skip_end_audios}")
+
+
+    if kim_vocal_version != 1 and kim_vocal_version != 2: raise ValueError(translations["version_not_valid"])
+    if separator_reverb and not separator_music: raise ValueError(translations["create_dataset_value_not_valid"])
+
+    start_time = time.time()
+
+
+    try:
+        paths = []
+
+        if not os.path.exists(dataset_temp): os.makedirs(dataset_temp, exist_ok=True)
+
+        urls = input_audio.replace(", ", ",").split(",")
+
+        for url in urls:
+            path = downloader(url, urls.index(url))
+            paths.append(path)
+
+        if skip:
+            skip_start_audios = skip_start_audios.replace(", ", ",").split(",")
+            skip_end_audios = skip_end_audios.replace(", ", ",").split(",")
+
+            if len(skip_start_audios) < len(paths) or len(skip_end_audios) < len(paths): 
+                logger.warning(translations["skip<audio"])
+                sys.exit(1)
+            elif len(skip_start_audios) > len(paths) or len(skip_end_audios) > len(paths): 
+                logger.warning(translations["skip>audio"])
+                sys.exit(1)
+            else:
+                for audio, skip_start_audio, skip_end_audio in zip(paths, skip_start_audios, skip_end_audios):
+                    skip_start(audio, skip_start_audio)
+                    skip_end(audio, skip_end_audio)
+
+        if separator_music:
+            separator_paths = []
+
+            for audio in paths:
+                vocals = separator_music_main(audio, dataset_temp, segments_size, overlap, denoise_mdx, kim_vocal_version, hop_length, batch_size)
+
+                if separator_reverb: vocals = separator_reverb_audio(vocals, dataset_temp, segments_size, overlap, denoise_mdx, hop_length, batch_size)
+                separator_paths.append(vocals)
+            
+            paths = separator_paths
+
+        processed_paths = []
+
+        for audio in paths:
+            output = process_audio(audio)
+            processed_paths.append(output)
+
+        paths = processed_paths
+                
+        for audio_path in paths:
+            data, sample_rate = sf.read(audio_path)
+
+            if resample_sr != sample_rate and resample_sr > 0 and resample: 
+                data = librosa.resample(data, orig_sr=sample_rate, target_sr=resample_sr)
+                sample_rate = resample_sr
+
+            if clean_dataset: data = nr.reduce_noise(y=data, prop_decrease=clean_strength)
+
+
+            sf.write(audio_path, data, sample_rate)
+    except Exception as e:
+        raise RuntimeError(f"{translations['create_dataset_error']}: {e}")
+    finally:
+        for audio in paths:
+            shutil.move(audio, output_dataset)
+
+        if os.path.exists(dataset_temp): shutil.rmtree(dataset_temp, ignore_errors=True)
+
+
+    elapsed_time = time.time() - start_time
+    logger.info(translations["create_dataset_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
+
+
+def downloader(url, name):
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        
+        ydl_opts = {
+            'format': 'bestaudio/best',
+            'outtmpl': os.path.join(dataset_temp, f"{name}"),
+            'postprocessors': [{
+                'key': 'FFmpegExtractAudio',
+                'preferredcodec': 'wav',
+                'preferredquality': '192',
+            }],
+            'noplaylist': True,
+            'verbose': False, 
+        }
+
+        logger.info(f"{translations['starting_download']}: {url}...")
+        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+            ydl.extract_info(url)  
+            logger.info(f"{translations['download_success']}: {url}")
+        
+    return os.path.join(dataset_temp, f"{name}" + ".wav")
+
+
+def skip_start(input_file, seconds):
+    data, sr = sf.read(input_file)
+    
+    total_duration = len(data) / sr
+    
+    if seconds <= 0: logger.warning(translations["=<0"])
+    elif seconds >= total_duration: logger.warning(translations["skip_warning"].format(seconds=seconds, total_duration=f"{total_duration:.2f}"))
+    else: 
+        logger.info(f"{translations['skip_start']}: {input_file}...")
+
+        sf.write(input_file, data[int(seconds * sr):], sr)
+
+        logger.info(translations["skip_start_audio"].format(input_file=input_file))
+
+
+def skip_end(input_file, seconds):
+    data, sr = sf.read(input_file)
+    
+    total_duration = len(data) / sr
+
+    if seconds <= 0: logger.warning(translations["=<0"])
+    elif seconds > total_duration: logger.warning(translations["skip_warning"].format(seconds=seconds, total_duration=f"{total_duration:.2f}"))
+    else: 
+        logger.info(f"{translations['skip_end']}: {input_file}...")
+
+        sf.write(input_file, data[:-int(seconds * sr)], sr)
+
+        logger.info(translations["skip_end_audio"].format(input_file=input_file))
+
+
+def process_audio(file_path):
+    try:
+        song = AudioSegment.from_file(file_path)
+        nonsilent_parts = silence.detect_nonsilent(song, min_silence_len=750, silence_thresh=-70)
+
+        cut_files = []
+
+        for i, (start_i, end_i) in enumerate(nonsilent_parts):
+            chunk = song[start_i:end_i]
+
+            if len(chunk) >= 30:
+                chunk_file_path = os.path.join(os.path.dirname(file_path), f"chunk{i}.wav")
+                if os.path.exists(chunk_file_path): os.remove(chunk_file_path)
+                
+                chunk.export(chunk_file_path, format="wav")
+
+                cut_files.append(chunk_file_path)
+            else: logger.warning(translations["skip_file"].format(i=i, chunk=len(chunk)))
+
+        logger.info(f"{translations['split_total']}: {len(cut_files)}")
+
+        def extract_number(filename):
+            match = re.search(r'_(\d+)', filename)
+
+            return int(match.group(1)) if match else 0
+
+        cut_files = sorted(cut_files, key=extract_number)
+
+        combined = AudioSegment.empty()
+
+        for file in cut_files:
+            combined += AudioSegment.from_file(file)
+
+        output_path = os.path.splitext(file_path)[0] + "_processed" + ".wav"
+
+        logger.info(translations["merge_audio"])
+
+        combined.export(output_path, format="wav")
+
+        return output_path
+    except Exception as e:
+        raise RuntimeError(f"{translations['process_audio_error']}: {e}")
+
+
+def separator_music_main(input, output, segments_size, overlap, denoise, version, hop_length, batch_size):
+    if not os.path.exists(input): 
+        logger.warning(translations["input_not_valid"])
+        return None
+    
+    if not os.path.exists(output): 
+        logger.warning(translations["output_not_valid"])
+        return None
+
+    model = f"Kim_Vocal_{version}.onnx"
+
+    logger.info(translations["separator_process"].format(input=input))
+    output_separator = separator_main(audio_file=input, model_filename=model, output_format="wav", output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise)
+
+    for f in output_separator:
+        path = os.path.join(output, f)
+
+        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
+
+        if '_(Instrumental)_' in f: os.rename(path, os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav")
+        elif '_(Vocals)_' in f:
+            rename_file = os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav"
+            os.rename(path, rename_file)
+
+    logger.info(f": {rename_file}")
+    return rename_file
+
+
+def separator_reverb_audio(input, output, segments_size, overlap, denoise, hop_length, batch_size):
+    reverb_models = "Reverb_HQ_By_FoxJoy.onnx"
+    
+    if not os.path.exists(input): 
+        logger.warning(translations["input_not_valid"])
+        return None
+    
+    if not os.path.exists(output): 
+        logger.warning(translations["output_not_valid"])
+        return None
+
+    logger.info(f"{translations['dereverb']}: {input}...")
+    output_dereverb = separator_main(audio_file=input, model_filename=reverb_models, output_format="wav", output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=hop_length, mdx_hop_length=batch_size, mdx_enable_denoise=denoise)
+
+    for f in output_dereverb:
+        path = os.path.join(output, f)
+
+        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
+
+        if '_(Reverb)_' in f: os.rename(path, os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav")
+        elif '_(No Reverb)_' in f:
+            rename_file = os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav"
+            os.rename(path, rename_file)    
+
+    logger.info(f"{translations['dereverb_success']}: {rename_file}")
+    return rename_file
+
+
+def separator_main(audio_file=None, model_filename="Kim_Vocal_1.onnx", output_format="wav", output_dir=".", mdx_segment_size=256, mdx_overlap=0.25, mdx_batch_size=1, mdx_hop_length=1024, mdx_enable_denoise=True):
+    separator = Separator(
+        log_formatter=file_formatter,
+        log_level=logging.INFO,
+        output_dir=output_dir,
+        output_format=output_format,
+        output_bitrate=None,
+        normalization_threshold=0.9,
+        output_single_stem=None,
+        invert_using_spec=False,
+        sample_rate=44100,
+        mdx_params={
+            "hop_length": mdx_hop_length,
+            "segment_size": mdx_segment_size,
+            "overlap": mdx_overlap,
+            "batch_size": mdx_batch_size,
+            "enable_denoise": mdx_enable_denoise,
+        },
+    )
+
+    separator.load_model(model_filename=model_filename)
+    return separator.separate(audio_file)
+
+if __name__ == "__main__": main()

main/inference/create_index.py

@@ -0,0 +1,120 @@
+import os
+import sys
+import faiss
+import logging
+import argparse
+import logging.handlers
+
+import numpy as np
+
+from multiprocessing import cpu_count
+from sklearn.cluster import MiniBatchKMeans
+
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+
+translations = Config().translations
+
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name", type=str, required=True)
+    parser.add_argument("--rvc_version", type=str, default="v2")
+    parser.add_argument("--index_algorithm", type=str, default="Auto")
+
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_arguments()
+
+    exp_dir = os.path.join("assets", "logs", args.model_name)
+    version = args.rvc_version
+    index_algorithm = args.index_algorithm
+
+    log_file = os.path.join(exp_dir, "create_index.log")
+    logger = logging.getLogger(__name__)
+
+
+    if logger.hasHandlers(): logger.handlers.clear()
+    else:  
+        console_handler = logging.StreamHandler()
+        console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+        console_handler.setFormatter(console_formatter)
+        console_handler.setLevel(logging.INFO)
+
+        file_handler = logging.handlers.RotatingFileHandler(log_file, maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
+        file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+        file_handler.setFormatter(file_formatter)
+        file_handler.setLevel(logging.DEBUG)
+
+        logger.addHandler(console_handler)
+        logger.addHandler(file_handler)
+        logger.setLevel(logging.DEBUG)
+
+    logger.debug(f"{translations['modelname']}: {args.model_name}")
+    logger.debug(f"{translations['model_path']}: {exp_dir}")
+    logger.debug(f"{translations['training_version']}: {version}")
+    logger.debug(f"{translations['index_algorithm_info']}: {index_algorithm}")
+
+
+    try:
+        feature_dir = os.path.join(exp_dir, f"{version}_extracted")
+        model_name = os.path.basename(exp_dir)
+
+        npys = []
+        listdir_res = sorted(os.listdir(feature_dir))
+
+        for name in listdir_res:
+            file_path = os.path.join(feature_dir, name)
+            phone = np.load(file_path)
+            npys.append(phone)
+
+        big_npy = np.concatenate(npys, axis=0)
+        big_npy_idx = np.arange(big_npy.shape[0])
+
+        np.random.shuffle(big_npy_idx)
+        
+        big_npy = big_npy[big_npy_idx]
+
+        if big_npy.shape[0] > 2e5 and (index_algorithm == "Auto" or index_algorithm == "KMeans"): big_npy = (MiniBatchKMeans(n_clusters=10000, verbose=True, batch_size=256 * cpu_count(), compute_labels=False, init="random").fit(big_npy).cluster_centers_)
+
+        np.save(os.path.join(exp_dir, "total_fea.npy"), big_npy)
+
+        n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
+
+        index_trained = faiss.index_factory(256 if version == "v1" else 768, f"IVF{n_ivf},Flat")
+
+        index_ivf_trained = faiss.extract_index_ivf(index_trained)
+        index_ivf_trained.nprobe = 1
+
+        index_trained.train(big_npy)
+
+        faiss.write_index(index_trained, os.path.join(exp_dir, f"trained_IVF{n_ivf}_Flat_nprobe_{index_ivf_trained.nprobe}_{model_name}_{version}.index"))
+
+        index_added = faiss.index_factory(256 if version == "v1" else 768, f"IVF{n_ivf},Flat")
+        index_ivf_added = faiss.extract_index_ivf(index_added)
+
+        index_ivf_added.nprobe = 1
+        index_added.train(big_npy)
+
+        batch_size_add = 8192
+        
+        for i in range(0, big_npy.shape[0], batch_size_add):
+            index_added.add(big_npy[i : i + batch_size_add])
+        
+        index_filepath_added = os.path.join(exp_dir, f"added_IVF{n_ivf}_Flat_nprobe_{index_ivf_added.nprobe}_{model_name}_{version}.index")
+
+        faiss.write_index(index_added, index_filepath_added)
+
+        logger.info(f"{translations['save_index']} '{index_filepath_added}'")
+    except Exception as e:
+        logger.error(f"{translations['create_index_error']}: {e}")
+
+if __name__ == "__main__": main()

main/inference/extract.py

@@ -0,0 +1,450 @@
+import os
+import gc
+import sys
+import time
+import tqdm
+import torch
+import shutil
+import codecs
+import pyworld
+import librosa
+import logging
+import argparse
+import warnings
+import subprocess
+import torchcrepe
+import parselmouth
+import logging.handlers
+
+import numpy as np
+import soundfile as sf
+import torch.nn.functional as F
+
+from random import shuffle
+from functools import partial
+from multiprocessing import Pool
+from distutils.util import strtobool
+from fairseq import checkpoint_utils
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+from main.library.predictors.FCPE import FCPE
+from main.library.predictors.RMVPE import RMVPE
+
+logging.getLogger("wget").setLevel(logging.ERROR)
+
+warnings.filterwarnings("ignore", category=FutureWarning)
+warnings.filterwarnings("ignore", category=UserWarning)
+
+logger = logging.getLogger(__name__)
+logger.propagate = False
+
+config = Config()
+translations = config.translations
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name", type=str, required=True)
+    parser.add_argument("--rvc_version", type=str, default="v2")
+    parser.add_argument("--f0_method", type=str, default="rmvpe")
+    parser.add_argument("--pitch_guidance", type=lambda x: bool(strtobool(x)), default=True)
+    parser.add_argument("--hop_length", type=int, default=128)
+    parser.add_argument("--cpu_cores", type=int, default=2)
+    parser.add_argument("--gpu", type=str, default="-")
+    parser.add_argument("--sample_rate", type=int, required=True)
+    parser.add_argument("--embedder_model", type=str, default="contentvec_base")
+
+    args = parser.parse_args()
+    return args
+
+def load_audio(file, sample_rate):
+    try:
+        file = file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        audio, sr = sf.read(file)
+
+        if len(audio.shape) > 1: audio = librosa.to_mono(audio.T)
+        if sr != sample_rate: audio = librosa.resample(audio, orig_sr=sr, target_sr=sample_rate)
+    except Exception as e:
+        raise RuntimeError(f"{translations['errors_loading_audio']}: {e}")
+
+    return audio.flatten()
+
+def check_rmvpe_fcpe(method):
+    if method == "rmvpe" and not os.path.exists(os.path.join("assets", "model", "predictors", "rmvpe.pt")): subprocess.run(["wget", "-q", "--show-progress", "--no-check-certificate", codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Pbyno_EIP_Cebwrpg_2/erfbyir/znva/", "rot13") + "rmvpe.pt", "-P", os.path.join("assets", "model", "predictors")], check=True)
+    elif method == "fcpe" and not os.path.exists(os.path.join("assets", "model", "predictors", "fcpe.pt")): subprocess.run(["wget", "-q", "--show-progress", "--no-check-certificate", codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Pbyno_EIP_Cebwrpg_2/erfbyir/znva/", "rot13") + "fcpe.pt", "-P", os.path.join("assets", "model", "predictors")], check=True)
+
+def check_hubert(hubert):
+    if hubert == "contentvec_base" or hubert == "hubert_base" or hubert == "japanese_hubert_base" or hubert == "korean_hubert_base" or hubert == "chinese_hubert_base":
+        model_path = os.path.join(now_dir, "assets", "model", "embedders", hubert + '.pt')
+
+        if not os.path.exists(model_path): subprocess.run(["wget", "-q", "--show-progress", "--no-check-certificate", codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Pbyno_EIP_Cebwrpg_2/erfbyir/znva/", "rot13") + f"{hubert}.pt", "-P", os.path.join("assets", "model", "embedders")], check=True)
+
+def generate_config(rvc_version, sample_rate, model_path):
+    config_path = os.path.join("main", "configs", rvc_version, f"{sample_rate}.json")
+    config_save_path = os.path.join(model_path, "config.json")
+    if not os.path.exists(config_save_path): shutil.copy(config_path, config_save_path)
+
+
+def generate_filelist(pitch_guidance, model_path, rvc_version, sample_rate):
+    gt_wavs_dir = os.path.join(model_path, "sliced_audios")
+    feature_dir = os.path.join(model_path, f"{rvc_version}_extracted")
+
+    f0_dir, f0nsf_dir = None, None
+    
+    if pitch_guidance:
+        f0_dir = os.path.join(model_path, "f0")
+        f0nsf_dir = os.path.join(model_path, "f0_voiced")
+
+    gt_wavs_files = set(name.split(".")[0] for name in os.listdir(gt_wavs_dir))
+    feature_files = set(name.split(".")[0] for name in os.listdir(feature_dir))
+
+    if pitch_guidance:
+        f0_files = set(name.split(".")[0] for name in os.listdir(f0_dir))
+        f0nsf_files = set(name.split(".")[0] for name in os.listdir(f0nsf_dir))
+        names = gt_wavs_files & feature_files & f0_files & f0nsf_files
+    else: names = gt_wavs_files & feature_files
+
+    options = []
+    mute_base_path = os.path.join(now_dir, "assets", "logs", "mute")
+
+    for name in names:
+        if pitch_guidance: options.append(f"{gt_wavs_dir}/{name}.wav|{feature_dir}/{name}.npy|{f0_dir}/{name}.wav.npy|{f0nsf_dir}/{name}.wav.npy|0")
+        else: options.append(f"{gt_wavs_dir}/{name}.wav|{feature_dir}/{name}.npy|0")
+
+    mute_audio_path = os.path.join(mute_base_path, "sliced_audios", f"mute{sample_rate}.wav")
+    mute_feature_path = os.path.join(mute_base_path, f"{rvc_version}_extracted", "mute.npy")
+
+    for _ in range(2):
+        if pitch_guidance:
+            mute_f0_path = os.path.join(mute_base_path, "f0", "mute.wav.npy")
+            mute_f0nsf_path = os.path.join(mute_base_path, "f0_voiced", "mute.wav.npy")
+            options.append(f"{mute_audio_path}|{mute_feature_path}|{mute_f0_path}|{mute_f0nsf_path}|0")
+        else: options.append(f"{mute_audio_path}|{mute_feature_path}|0")
+
+    shuffle(options)
+
+    with open(os.path.join(model_path, "filelist.txt"), "w") as f:
+        f.write("\n".join(options))
+
+def setup_paths(exp_dir, version = None):
+    wav_path = os.path.join(exp_dir, "sliced_audios_16k")
+    if version:
+        out_path = os.path.join(exp_dir, "v1_extracted" if version == "v1" else "v2_extracted")
+        os.makedirs(out_path, exist_ok=True)
+
+        return wav_path, out_path
+    else:
+        output_root1 = os.path.join(exp_dir, "f0")
+        output_root2 = os.path.join(exp_dir, "f0_voiced")
+
+        os.makedirs(output_root1, exist_ok=True)
+        os.makedirs(output_root2, exist_ok=True)
+
+        return wav_path, output_root1, output_root2
+
+def read_wave(wav_path, normalize = False):
+    wav, sr = sf.read(wav_path)
+    assert sr == 16000, translations["sr_not_16000"]
+
+    feats = torch.from_numpy(wav).float()
+
+    if config.is_half: feats = feats.half()
+    if feats.dim() == 2: feats = feats.mean(-1)
+
+    feats = feats.view(1, -1)
+
+    if normalize: feats = F.layer_norm(feats, feats.shape)
+
+    return feats
+
+def get_device(gpu_index):
+    if gpu_index == "cpu": return "cpu"
+
+    try:
+        index = int(gpu_index)
+        if index < torch.cuda.device_count(): return f"cuda:{index}"
+        else: logger.warning(translations["gpu_not_valid"])
+    except ValueError:
+        logger.warning(translations["gpu_not_valid"])
+
+    return "cpu"
+
+class FeatureInput:
+    def __init__(self, sample_rate=16000, hop_size=160, device="cpu"):
+        self.fs = sample_rate
+        self.hop = hop_size
+        self.f0_bin = 256
+        self.f0_max = 1100.0
+        self.f0_min = 50.0
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+        self.device = device
+        
+    def compute_f0(self, np_arr, f0_method, hop_length):
+        if f0_method == "pm": return self.get_pm(np_arr)
+        elif f0_method == 'dio': return self.get_dio(np_arr)
+        elif f0_method == "crepe": return self.get_crepe(np_arr, int(hop_length))
+        elif f0_method == "crepe-tiny": return self.get_crepe(np_arr, int(hop_length), "tiny")
+        elif f0_method == "fcpe": return self.get_fcpe(np_arr, int(hop_length))
+        elif f0_method == "rmvpe": return self.get_rmvpe(np_arr)
+        elif f0_method == "harvest": return self.get_harvest(np_arr)
+        else: raise ValueError(translations["method_not_valid"])
+
+    def get_pm(self, x):
+        time_step = 160 / 16000 * 1000
+        f0 = (parselmouth.Sound(x, self.fs).to_pitch_ac(time_step=time_step / 1000, voicing_threshold=0.6, pitch_floor=50, pitch_ceiling=1100).selected_array["frequency"])
+        pad_size = ((x.size // self.hop) - len(f0) + 1) // 2
+        if pad_size > 0 or (x.size // self.hop) - len(f0) - pad_size > 0: f0 = np.pad(f0, [[pad_size, (x.size // self.hop) - len(f0) - pad_size]], mode="constant")
+
+        return f0
+    
+    def get_dio(self, x):
+        f0, t = pyworld.dio(x.astype(np.double), fs=self.fs, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=1000 * self.hop / self.fs)
+        f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+        
+        return f0
+    
+    def get_crepe(self, x, hop_length, model="full"):
+        audio = torch.from_numpy(x.astype(np.float32)).to(self.device)
+        audio /= torch.quantile(torch.abs(audio), 0.999)
+        audio = audio.unsqueeze(0)
+
+        pitch = torchcrepe.predict(audio, self.fs, hop_length, self.f0_min, self.f0_max, model=model, batch_size=hop_length * 2, device=self.device, pad=True)
+
+        source = pitch.squeeze(0).cpu().float().numpy()
+        source[source < 0.001] = np.nan
+        target = np.interp(np.arange(0, len(source) * (x.size // self.hop), len(source)) / (x.size // self.hop), np.arange(0, len(source)), source)
+
+        return np.nan_to_num(target)
+    
+
+    def get_fcpe(self, x, hop_length):
+        self.model_fcpe = FCPE(os.path.join("assets", "model", "predictors", "fcpe.pt"), hop_length=int(hop_length), f0_min=self.f0_min, f0_max=self.f0_max, dtype=torch.float32, device=self.device, sample_rate=self.fs, threshold=0.03)
+        f0 = self.model_fcpe.compute_f0(x, p_len=(x.size // self.hop))
+        del self.model_fcpe
+        gc.collect()
+        return f0
+    
+
+    def get_rmvpe(self, x):
+        self.model_rmvpe = RMVPE(os.path.join("assets", "model", "predictors", "rmvpe.pt"), is_half=False, device=self.device)
+        return self.model_rmvpe.infer_from_audio(x, thred=0.03)
+
+
+    def get_harvest(self, x):
+        f0, t = pyworld.harvest(x.astype(np.double), fs=self.fs, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=1000 * self.hop / self.fs)
+        f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+        return f0
+        
+
+    def coarse_f0(self, f0):
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel = np.clip((f0_mel - self.f0_mel_min) * (self.f0_bin - 2) / (self.f0_mel_max - self.f0_mel_min) + 1, 1, self.f0_bin - 1)
+        return np.rint(f0_mel).astype(int)
+
+
+    def process_file(self, file_info, f0_method, hop_length):
+        inp_path, opt_path1, opt_path2, np_arr = file_info
+
+        if os.path.exists(opt_path1 + ".npy") and os.path.exists(opt_path2 + ".npy"): return
+
+        try:
+            feature_pit = self.compute_f0(np_arr, f0_method, hop_length)
+            np.save(opt_path2, feature_pit, allow_pickle=False)
+            coarse_pit = self.coarse_f0(feature_pit)
+            np.save(opt_path1, coarse_pit, allow_pickle=False)
+        except Exception as e:
+            raise RuntimeError(f"{translations['extract_file_error']} {inp_path}: {e}")
+
+
+    def process_files(self, files, f0_method, hop_length, pbar):
+        for file_info in files:
+            self.process_file(file_info, f0_method, hop_length)
+            pbar.update()
+
+def run_pitch_extraction(exp_dir, f0_method, hop_length, num_processes, gpus):
+    input_root, *output_roots = setup_paths(exp_dir)
+
+    if len(output_roots) == 2: output_root1, output_root2 = output_roots
+    else:
+        output_root1 = output_roots[0]
+        output_root2 = None
+
+    paths = [
+        (
+            os.path.join(input_root, name),
+            os.path.join(output_root1, name) if output_root1 else None,
+            os.path.join(output_root2, name) if output_root2 else None,
+            load_audio(os.path.join(input_root, name), 16000),
+        )
+        for name in sorted(os.listdir(input_root))
+        if "spec" not in name
+    ]
+
+    logger.info(translations["extract_f0_method"].format(num_processes=num_processes, f0_method=f0_method))
+    start_time = time.time()
+
+    if gpus != "-":
+        gpus = gpus.split("-")
+        num_gpus = len(gpus)
+
+        process_partials = []
+
+        pbar = tqdm.tqdm(total=len(paths), desc=translations["extract_f0"])
+
+        for idx, gpu in enumerate(gpus):
+            device = get_device(gpu)
+            feature_input = FeatureInput(device=device)
+
+            part_paths = paths[idx::num_gpus]
+            process_partials.append((feature_input, part_paths))
+
+        with ThreadPoolExecutor() as executor:
+            futures = [executor.submit(FeatureInput.process_files, feature_input, part_paths, f0_method, hop_length, pbar) for feature_input, part_paths in process_partials]
+
+            for future in as_completed(futures):
+                pbar.update(1)
+                future.result()
+
+        pbar.close()
+    else:
+        feature_input = FeatureInput(device="cpu")
+        
+        with tqdm.tqdm(total=len(paths), desc=translations["extract_f0"]) as pbar:
+            with Pool(processes=num_processes) as pool:
+                process_file_partial = partial(feature_input.process_file, f0_method=f0_method, hop_length=hop_length)
+
+                for _ in pool.imap_unordered(process_file_partial, paths):
+                    pbar.update(1)
+
+
+    elapsed_time = time.time() - start_time
+    logger.info(translations["extract_f0_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
+
+def process_file_embedding(file, wav_path, out_path, model, device, version, saved_cfg):
+    wav_file_path = os.path.join(wav_path, file)
+    out_file_path = os.path.join(out_path, file.replace("wav", "npy"))
+
+    if os.path.exists(out_file_path): return
+
+    feats = read_wave(wav_file_path, normalize=saved_cfg.task.normalize)
+    dtype = torch.float16 if device.startswith("cuda") else torch.float32
+    feats = feats.to(dtype).to(device)
+
+    padding_mask = torch.BoolTensor(feats.shape).fill_(False).to(dtype).to(device)
+
+    inputs = {
+        "source": feats,
+        "padding_mask": padding_mask,
+        "output_layer": 9 if version == "v1" else 12,
+    }
+
+    with torch.no_grad():
+        model = model.to(device).to(dtype)
+        logits = model.extract_features(**inputs)
+        feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+
+    feats = feats.squeeze(0).float().cpu().numpy()
+
+    if not np.isnan(feats).any(): np.save(out_file_path, feats, allow_pickle=False)
+    else: logger.warning(f"{file} {translations['NaN']}")
+
+def run_embedding_extraction(exp_dir, version, gpus, embedder_model):
+    wav_path, out_path = setup_paths(exp_dir, version)
+
+    logger.info(translations["start_extract_hubert"])
+    start_time = time.time()
+
+    try:
+        models, saved_cfg, _ = checkpoint_utils.load_model_ensemble_and_task([os.path.join(now_dir, "assets", "model", "embedders", embedder_model + '.pt')], suffix="")
+    except Exception as e:
+        raise ImportError(translations["read_model_error"].format(e=e))
+
+    model = models[0]
+    devices = [get_device(gpu) for gpu in (gpus.split("-") if gpus != "-" else ["cpu"])]
+
+    paths = sorted([file for file in os.listdir(wav_path) if file.endswith(".wav")])
+
+    if not paths:
+        logger.warning(translations["not_found_audio_file"])
+        sys.exit(1)
+
+    pbar = tqdm.tqdm(total=len(paths) * len(devices), desc=translations["extract_hubert"])
+
+    tasks = [(file, wav_path, out_path, model, device, version, saved_cfg) for file in paths for device in devices]
+
+    for task in tasks:
+        try:
+            process_file_embedding(*task)
+        except Exception as e:
+            raise RuntimeError(f"{translations['process_error']} {task[0]}: {e}")
+
+        pbar.update(1)
+
+    pbar.close()
+
+    elapsed_time = time.time() - start_time
+    logger.info(translations["extract_hubert_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
+
+if __name__ == "__main__":
+    args = parse_arguments()
+
+    exp_dir = os.path.join("assets", "logs", args.model_name)
+    f0_method = args.f0_method
+    hop_length = args.hop_length
+    num_processes = args.cpu_cores
+    gpus = args.gpu
+    version = args.rvc_version
+    pitch_guidance = args.pitch_guidance
+    sample_rate = args.sample_rate
+    embedder_model = args.embedder_model
+
+    check_rmvpe_fcpe(f0_method)
+    check_hubert(embedder_model)
+
+    if len([f for f in os.listdir(os.path.join(exp_dir, "sliced_audios")) if os.path.isfile(os.path.join(exp_dir, "sliced_audios", f))]) < 1 or len([f for f in os.listdir(os.path.join(exp_dir, "sliced_audios_16k")) if os.path.isfile(os.path.join(exp_dir, "sliced_audios_16k", f))]) < 1: raise FileNotFoundError("Không tìm thấy dữ liệu được xử lý, vui lòng xử lý lại âm thanh")
+
+    log_file = os.path.join(exp_dir, "extract.log")
+
+    if logger.hasHandlers(): logger.handlers.clear()
+    else:
+        console_handler = logging.StreamHandler()
+        console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+        console_handler.setFormatter(console_formatter)
+        console_handler.setLevel(logging.INFO)
+
+        file_handler = logging.handlers.RotatingFileHandler(log_file, maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
+        file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+        file_handler.setFormatter(file_formatter)
+        file_handler.setLevel(logging.DEBUG)
+
+        logger.addHandler(console_handler)
+        logger.addHandler(file_handler)
+        logger.setLevel(logging.DEBUG)
+
+    logger.debug(f"{translations['modelname']}: {args.model_name}")
+    logger.debug(f"{translations['export_process']}: {exp_dir}")
+    logger.debug(f"{translations['f0_method']}: {f0_method}")
+    logger.debug(f"{translations['pretrain_sr']}: {sample_rate}")
+    logger.debug(f"{translations['cpu_core']}: {num_processes}")
+    logger.debug(f"Gpu: {gpus}")
+    if f0_method == "crepe" or f0_method == "crepe-tiny" or f0_method == "fcpe": logger.debug(f"Hop length: {hop_length}")
+    logger.debug(f"{translations['training_version']}: {version}")
+    logger.debug(f"{translations['extract_f0']}: {pitch_guidance}")
+    logger.debug(f"{translations['hubert_model']}: {embedder_model}")
+
+    try:
+        run_pitch_extraction(exp_dir, f0_method, hop_length, num_processes, gpus)
+        run_embedding_extraction(exp_dir, version, gpus, embedder_model)
+        
+        generate_config(version, sample_rate, exp_dir)
+        generate_filelist(pitch_guidance, exp_dir, version, sample_rate)
+    except Exception as e:
+        logger.error(f"{translations['extract_error']}: {e}")
+
+    logger.info(f"{translations['extract_success']} {args.model_name}.")

main/inference/preprocess.py

@@ -0,0 +1,360 @@
+import os
+import sys
+import time
+import torch
+import logging
+import librosa
+import argparse
+import logging.handlers
+
+import numpy as np
+import soundfile as sf
+import multiprocessing
+import noisereduce as nr
+
+from tqdm import tqdm
+from scipy import signal
+from scipy.io import wavfile
+from distutils.util import strtobool
+from concurrent.futures import ProcessPoolExecutor, as_completed
+
+now_directory = os.getcwd()
+sys.path.append(now_directory)
+
+from main.configs.config import Config
+
+logger = logging.getLogger(__name__)
+
+
+logging.getLogger("numba.core.byteflow").setLevel(logging.ERROR)
+logging.getLogger("numba.core.ssa").setLevel(logging.ERROR)
+logging.getLogger("numba.core.interpreter").setLevel(logging.ERROR)
+
+OVERLAP = 0.3
+MAX_AMPLITUDE = 0.9
+ALPHA = 0.75
+HIGH_PASS_CUTOFF = 48
+SAMPLE_RATE_16K = 16000
+
+
+config = Config()
+per = 3.0 if config.is_half else 3.7
+translations = config.translations
+
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name", type=str, required=True)
+    parser.add_argument("--dataset_path", type=str, default="./dataset")
+    parser.add_argument("--sample_rate", type=int, required=True)
+    parser.add_argument("--cpu_cores", type=int, default=2)
+    parser.add_argument("--cut_preprocess", type=lambda x: bool(strtobool(x)), default=True)
+    parser.add_argument("--process_effects", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--clean_dataset", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--clean_strength", type=float, default=0.7)
+
+    args = parser.parse_args()
+    return args
+
+
+def load_audio(file, sample_rate):
+    try:
+        file = file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        audio, sr = sf.read(file)
+
+        if len(audio.shape) > 1: audio = librosa.to_mono(audio.T)
+        if sr != sample_rate: audio = librosa.resample(audio, orig_sr=sr, target_sr=sample_rate)
+    except Exception as e:
+        raise RuntimeError(f"{translations['errors_loading_audio']}: {e}")
+
+    return audio.flatten()
+
+class Slicer:
+    def __init__(self, sr, threshold = -40.0, min_length = 5000, min_interval = 300, hop_size = 20, max_sil_kept = 5000):
+        if not min_length >= min_interval >= hop_size: raise ValueError(translations["min_length>=min_interval>=hop_size"])
+        if not max_sil_kept >= hop_size: raise ValueError(translations["max_sil_kept>=hop_size"])
+
+        min_interval = sr * min_interval / 1000
+        self.threshold = 10 ** (threshold / 20.0)
+        self.hop_size = round(sr * hop_size / 1000)
+        self.win_size = min(round(min_interval), 4 * self.hop_size)
+        self.min_length = round(sr * min_length / 1000 / self.hop_size)
+        self.min_interval = round(min_interval / self.hop_size)
+        self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+
+    def _apply_slice(self, waveform, begin, end):
+        start_idx = begin * self.hop_size
+
+        if len(waveform.shape) > 1:
+            end_idx = min(waveform.shape[1], end * self.hop_size)
+            return waveform[:, start_idx:end_idx]
+        else:
+            end_idx = min(waveform.shape[0], end * self.hop_size)
+            return waveform[start_idx:end_idx]
+
+    def slice(self, waveform):
+        samples = waveform.mean(axis=0) if len(waveform.shape) > 1 else waveform
+        if samples.shape[0] <= self.min_length: return [waveform]
+
+        rms_list = get_rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
+
+        sil_tags = []
+        silence_start, clip_start = None, 0
+
+        for i, rms in enumerate(rms_list):
+            if rms < self.threshold:
+                if silence_start is None: silence_start = i
+                continue
+
+            if silence_start is None: continue
+
+            is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+            need_slice_middle = (i - silence_start >= self.min_interval and i - clip_start >= self.min_length)
+
+            if not is_leading_silence and not need_slice_middle:
+                silence_start = None
+                continue
+
+            if i - silence_start <= self.max_sil_kept:
+                pos = rms_list[silence_start : i + 1].argmin() + silence_start
+                if silence_start == 0: sil_tags.append((0, pos))
+                else: sil_tags.append((pos, pos))
+
+                clip_start = pos
+
+            elif i - silence_start <= self.max_sil_kept * 2:
+                pos = rms_list[i - self.max_sil_kept : silence_start + self.max_sil_kept + 1].argmin()
+
+                pos += i - self.max_sil_kept
+
+                pos_l = (rms_list[silence_start : silence_start + self.max_sil_kept + 1].argmin() + silence_start)
+                pos_r = (rms_list[i - self.max_sil_kept : i + 1].argmin() + i - self.max_sil_kept)
+
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                    clip_start = pos_r
+                else:
+                    sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+                    clip_start = max(pos_r, pos)
+            else:
+                pos_l = (rms_list[silence_start : silence_start + self.max_sil_kept + 1].argmin() + silence_start)
+                pos_r = (rms_list[i - self.max_sil_kept : i + 1].argmin() + i - self.max_sil_kept)
+
+                if silence_start == 0: sil_tags.append((0, pos_r))
+                else: sil_tags.append((pos_l, pos_r))
+
+                clip_start = pos_r
+            silence_start = None
+
+        total_frames = rms_list.shape[0]
+
+        if (silence_start is not None and total_frames - silence_start >= self.min_interval):
+            silence_end = min(total_frames, silence_start + self.max_sil_kept)
+            pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start
+            sil_tags.append((pos, total_frames + 1))
+
+        if not sil_tags: return [waveform]
+        else:
+            chunks = []
+
+            if sil_tags[0][0] > 0: chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))
+
+            for i in range(len(sil_tags) - 1):
+                chunks.append(self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0]))
+
+            if sil_tags[-1][1] < total_frames: chunks.append(self._apply_slice(waveform, sil_tags[-1][1], total_frames))
+
+            return chunks
+
+
+def get_rms(y, frame_length=2048, hop_length=512, pad_mode="constant"):
+    padding = (int(frame_length // 2), int(frame_length // 2))
+    y = np.pad(y, padding, mode=pad_mode)
+
+    axis = -1
+    out_strides = y.strides + tuple([y.strides[axis]])
+    x_shape_trimmed = list(y.shape)
+    x_shape_trimmed[axis] -= frame_length - 1
+    out_shape = tuple(x_shape_trimmed) + tuple([frame_length])
+    xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)
+
+    target_axis = axis - 1 if axis < 0 else axis + 1
+
+
+    xw = np.moveaxis(xw, -1, target_axis)
+    slices = [slice(None)] * xw.ndim
+    slices[axis] = slice(0, None, hop_length)
+    x = xw[tuple(slices)]
+
+    power = np.mean(np.abs(x) ** 2, axis=-2, keepdims=True)
+    return np.sqrt(power)
+
+
+class PreProcess:
+    def __init__(self, sr, exp_dir, per):
+        self.slicer = Slicer(sr=sr, threshold=-42, min_length=1500, min_interval=400, hop_size=15, max_sil_kept=500)
+        self.sr = sr
+        self.b_high, self.a_high = signal.butter(N=5, Wn=HIGH_PASS_CUTOFF, btype="high", fs=self.sr)
+        self.per = per
+        self.exp_dir = exp_dir
+        self.device = "cpu"
+        self.gt_wavs_dir = os.path.join(exp_dir, "sliced_audios")
+        self.wavs16k_dir = os.path.join(exp_dir, "sliced_audios_16k")
+
+        os.makedirs(self.gt_wavs_dir, exist_ok=True)
+        os.makedirs(self.wavs16k_dir, exist_ok=True)
+
+    def _normalize_audio(self, audio: torch.Tensor):
+        tmp_max = torch.abs(audio).max()
+        if tmp_max > 2.5: return None
+        
+        return (audio / tmp_max * (MAX_AMPLITUDE * ALPHA)) + (1 - ALPHA) * audio
+
+    def process_audio_segment(self, normalized_audio: np.ndarray, sid, idx0, idx1):
+        if normalized_audio is None:
+            logs(f"{sid}-{idx0}-{idx1}-filtered")
+            return
+        
+        wavfile.write(os.path.join(self.gt_wavs_dir, f"{sid}_{idx0}_{idx1}.wav"), self.sr, normalized_audio.astype(np.float32))
+        audio_16k = librosa.resample(normalized_audio, orig_sr=self.sr, target_sr=SAMPLE_RATE_16K)
+        wavfile.write(os.path.join(self.wavs16k_dir, f"{sid}_{idx0}_{idx1}.wav"), SAMPLE_RATE_16K, audio_16k.astype(np.float32))
+
+    def process_audio(self, path, idx0, sid, cut_preprocess, process_effects, clean_dataset, clean_strength):
+        try:
+            audio = load_audio(path, self.sr)
+
+            if process_effects: 
+                audio = signal.lfilter(self.b_high, self.a_high, audio)
+                audio = self._normalize_audio(audio)
+
+            if clean_dataset: audio = nr.reduce_noise(y=audio, sr=self.sr, prop_decrease=clean_strength)
+
+            idx1 = 0
+
+            if cut_preprocess:
+                for audio_segment in self.slicer.slice(audio):
+                    i = 0
+
+                    while 1:
+                        start = int(self.sr * (self.per - OVERLAP) * i)
+                        i += 1
+
+                        if len(audio_segment[start:]) > (self.per + OVERLAP) * self.sr:
+                            tmp_audio = audio_segment[start : start + int(self.per * self.sr)]
+                            self.process_audio_segment(tmp_audio, sid, idx0, idx1)
+                            idx1 += 1
+                        else:
+                            tmp_audio = audio_segment[start:]
+                            self.process_audio_segment(tmp_audio, sid, idx0, idx1)
+                            idx1 += 1
+                            break
+                        
+            else: self.process_audio_segment(audio, sid, idx0, idx1)
+        except Exception as e:
+            raise RuntimeError(f"{translations['process_audio_error']}: {e}")
+
+def process_file(args):
+    pp, file, cut_preprocess, process_effects, clean_dataset, clean_strength = (args)
+    file_path, idx0, sid = file
+
+    pp.process_audio(file_path, idx0, sid, cut_preprocess, process_effects, clean_dataset, clean_strength)
+
+def preprocess_training_set(input_root, sr, num_processes, exp_dir, per, cut_preprocess, process_effects, clean_dataset, clean_strength):
+    start_time = time.time()
+
+    pp = PreProcess(sr, exp_dir, per)
+    logger.info(translations["start_preprocess"].format(num_processes=num_processes))
+
+    files = []
+    idx = 0
+
+    for root, _, filenames in os.walk(input_root):
+        try:
+            sid = 0 if root == input_root else int(os.path.basename(root))
+            for f in filenames:
+                if f.lower().endswith((".wav", ".mp3", ".flac", ".ogg")):
+                    files.append((os.path.join(root, f), idx, sid))
+                    idx += 1
+        except ValueError:
+            raise ValueError(f"{translations['not_integer']} '{os.path.basename(root)}'.")
+
+    with tqdm(total=len(files), desc=translations["preprocess"]) as pbar:
+        with ProcessPoolExecutor(max_workers=num_processes) as executor:
+            futures = [
+                executor.submit(
+                    process_file,
+                    (
+                        pp,
+                        file,
+                        cut_preprocess,
+                        process_effects,
+                        clean_dataset,
+                        clean_strength,
+                    ),
+                )
+                for file in files
+            ]
+            for future in as_completed(futures):
+                try:
+                    future.result() 
+                except Exception as e:
+                    raise RuntimeError(f"{translations['process_error']}: {e}")
+
+                pbar.update(1)
+
+    elapsed_time = time.time() - start_time
+    logger.info(translations["preprocess_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
+
+if __name__ == "__main__":
+    args = parse_arguments()
+
+    experiment_directory = os.path.join("assets", "logs", args.model_name)
+    num_processes = args.cpu_cores
+    num_processes = multiprocessing.cpu_count() if num_processes is None else int(num_processes)
+    dataset = args.dataset_path
+    sample_rate = args.sample_rate
+    cut_preprocess = args.cut_preprocess
+    preprocess_effects = args.process_effects
+    clean_dataset = args.clean_dataset
+    clean_strength = args.clean_strength
+
+    os.makedirs(experiment_directory, exist_ok=True)
+
+    if len([f for f in os.listdir(os.path.join(dataset)) if os.path.isfile(os.path.join(dataset, f)) and f.lower().endswith((".wav", ".mp3", ".flac", ".ogg"))]) < 1: raise FileNotFoundError("Không tìm thấy dữ liệu")
+
+    log_file = os.path.join(experiment_directory, "preprocess.log")
+
+    if logger.hasHandlers(): logger.handlers.clear()
+    else:
+        console_handler = logging.StreamHandler()
+        console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+        console_handler.setFormatter(console_formatter)
+        console_handler.setLevel(logging.INFO)
+
+        file_handler = logging.handlers.RotatingFileHandler(log_file, maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
+        file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+        file_handler.setFormatter(file_formatter)
+        file_handler.setLevel(logging.DEBUG)
+
+        logger.addHandler(console_handler)
+        logger.addHandler(file_handler)
+        logger.setLevel(logging.DEBUG)
+
+    logger.debug(f"{translations['modelname']}: {args.model_name}")
+    logger.debug(f"{translations['export_process']}: {experiment_directory}")
+    logger.debug(f"{translations['dataset_folder']}: {dataset}")
+    logger.debug(f"{translations['pretrain_sr']}: {sample_rate}")
+    logger.debug(f"{translations['cpu_core']}: {num_processes}")
+    logger.debug(f"{translations['split_audio']}: {cut_preprocess}")
+    logger.debug(f"{translations['preprocess_effect']}: {preprocess_effects}")
+    logger.debug(f"{translations['clear_audio']}: {clean_dataset}")
+    if clean_dataset: logger.debug(f"{translations['clean_strength']}: {clean_strength}")
+    
+    try:
+        preprocess_training_set(dataset, sample_rate, num_processes, experiment_directory, per, cut_preprocess, preprocess_effects, clean_dataset, clean_strength)
+    except Exception as e:
+        logger.error(f"{translations['process_audio_error']} {e}")
+
+    logger.info(f"{translations['preprocess_model_success']} {args.model_name}")

main/inference/separator_music.py

@@ -0,0 +1,400 @@
+import os
+import sys
+import time
+import logging
+import argparse
+import logging.handlers
+
+import soundfile as sf
+import noisereduce as nr
+
+from pydub import AudioSegment
+from distutils.util import strtobool
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+from main.library.algorithm.separator import Separator
+
+translations = Config().translations
+
+log_file = os.path.join("assets", "logs", "separator.log")
+logger = logging.getLogger(__name__)
+
+if logger.hasHandlers(): logger.handlers.clear()
+else:
+    console_handler = logging.StreamHandler()
+    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    console_handler.setFormatter(console_formatter)
+    console_handler.setLevel(logging.INFO)
+
+    file_handler = logging.handlers.RotatingFileHandler(log_file, maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
+    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    file_handler.setFormatter(file_formatter)
+    file_handler.setLevel(logging.DEBUG)
+
+    logger.addHandler(console_handler)
+    logger.addHandler(file_handler)
+    logger.setLevel(logging.DEBUG)
+
+demucs_models = {
+    "HT-Tuned": "htdemucs_ft.yaml",
+    "HT-Normal": "htdemucs.yaml",
+    "HD_MMI": "hdemucs_mmi.yaml",
+    "HT_6S": "htdemucs_6s.yaml"
+}
+
+mdx_models = {
+    "Main_340": "UVR-MDX-NET_Main_340.onnx",
+    "Main_390": "UVR-MDX-NET_Main_390.onnx",
+    "Main_406": "UVR-MDX-NET_Main_406.onnx",
+    "Main_427": "UVR-MDX-NET_Main_427.onnx",
+    "Main_438": "UVR-MDX-NET_Main_438.onnx",
+    "Inst_full_292": "UVR-MDX-NET-Inst_full_292.onnx",
+    "Inst_HQ_1": "UVR-MDX-NET_Inst_HQ_1.onnx",
+    "Inst_HQ_2": "UVR-MDX-NET_Inst_HQ_2.onnx",
+    "Inst_HQ_3": "UVR-MDX-NET_Inst_HQ_3.onnx",
+    "Inst_HQ_4": "UVR-MDX-NET-Inst_HQ_4.onnx",
+    "Kim_Vocal_1": "Kim_Vocal_1.onnx",
+    "Kim_Vocal_2": "Kim_Vocal_2.onnx",
+    "Kim_Inst": "Kim_Inst.onnx",
+    "Inst_187_beta": "UVR-MDX-NET_Inst_187_beta.onnx",
+    "Inst_82_beta": "UVR-MDX-NET_Inst_82_beta.onnx",
+    "Inst_90_beta": "UVR-MDX-NET_Inst_90_beta.onnx",
+    "Voc_FT": "UVR-MDX-NET-Voc_FT.onnx",
+    "Crowd_HQ": "UVR-MDX-NET_Crowd_HQ_1.onnx",
+    "MDXNET_9482": "UVR_MDXNET_9482.onnx",
+    "Inst_1": "UVR-MDX-NET-Inst_1.onnx",
+    "Inst_2": "UVR-MDX-NET-Inst_2.onnx",
+    "Inst_3": "UVR-MDX-NET-Inst_3.onnx",
+    "MDXNET_1_9703": "UVR_MDXNET_1_9703.onnx",
+    "MDXNET_2_9682": "UVR_MDXNET_2_9682.onnx",
+    "MDXNET_3_9662": "UVR_MDXNET_3_9662.onnx",
+    "Inst_Main": "UVR-MDX-NET-Inst_Main.onnx",
+    "MDXNET_Main": "UVR_MDXNET_Main.onnx"
+}
+
+kara_models = {
+    "Version-1": "UVR_MDXNET_KARA.onnx",
+    "Version-2": "UVR_MDXNET_KARA_2.onnx"
+}
+
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input_path", type=str, required=True)
+    parser.add_argument("--output_path", type=str, default="./audios")
+    parser.add_argument("--format", type=str, default="wav")
+    parser.add_argument("--shifts", type=int, default=10)
+    parser.add_argument("--segments_size", type=int, default=256)
+    parser.add_argument("--overlap", type=float, default=0.25)
+    parser.add_argument("--mdx_hop_length", type=int, default=1024)
+    parser.add_argument("--mdx_batch_size", type=int, default=1)
+    parser.add_argument("--clean_audio", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--clean_strength", type=float, default=0.7)
+    parser.add_argument("--backing_denoise", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--demucs_model", type=str, default="HT-Normal")
+    parser.add_argument("--kara_model", type=str, default="Version-1")
+    parser.add_argument("--mdx_model", type=str, default="Main_340")
+    parser.add_argument("--backing", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--mdx", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--mdx_denoise", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--reverb", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--reverb_denoise", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--backing_reverb", type=lambda x: bool(strtobool(x)), default=False)
+
+    args = parser.parse_args()
+    return args
+
+def main():
+    start_time = time.time()
+
+    try:
+        args = parse_arguments()
+
+        input_path = args.input_path
+        output_path = args.output_path
+        export_format = args.format
+        shifts = args.shifts
+        segments_size = args.segments_size
+        overlap = args.overlap
+        hop_length = args.mdx_hop_length
+        batch_size = args.mdx_batch_size
+        clean_audio = args.clean_audio
+        clean_strength = args.clean_strength
+        backing_denoise = args.backing_denoise
+        demucs_model = args.demucs_model
+        kara_model = args.kara_model
+        backing = args.backing
+        mdx = args.mdx
+        mdx_model = args.mdx_model
+        mdx_denoise = args.mdx_denoise
+        reverb = args.reverb
+        reverb_denoise = args.reverb_denoise
+        backing_reverb = args.backing_reverb
+
+        if backing_reverb and not reverb: 
+            logger.warning(translations["turn_on_dereverb"])
+            sys.exit(1)
+
+        if backing_reverb and not backing: 
+            logger.warning(translations["turn_on_separator_backing"])
+            sys.exit(1)
+
+        logger.debug(f"{translations['audio_path']}: {input_path}")
+        logger.debug(f"{translations['output_path']}: {output_path}")
+        logger.debug(f"{translations['export_format']}: {export_format}")
+        if not mdx: logger.debug(f"{translations['shift']}: {shifts}") 
+        logger.debug(f"{translations['segments_size']}: {segments_size}")
+        logger.debug(f"{translations['overlap']}: {overlap}")
+        if clean_audio: logger.debug(f"{translations['clear_audio']}: {clean_audio}")
+        if clean_audio: logger.debug(f"{translations['clean_strength']}: {clean_strength}")
+        if not mdx: logger.debug(f"{translations['demucs_model']}: {demucs_model}")
+        if backing: logger.debug(f"{translations['denoise_backing']}: {backing_denoise}")
+        if backing: logger.debug(f"{translations['backing_model_ver']}: {kara_model}")
+        if backing: logger.debug(f"{translations['separator_backing']}: {backing}")
+        if mdx: logger.debug(f"{translations['use_mdx']}: {mdx}")
+        if mdx: logger.debug(f"{translations['mdx_model']}: {mdx_model}")
+        if mdx: logger.debug(f"{translations['denoise_mdx']}: {mdx_denoise}")
+        if mdx or backing or reverb: logger.debug(f"Hop length: {hop_length}")
+        if mdx or backing or reverb: logger.debug(f"{translations['batch_size']}: {batch_size}")
+        if reverb: logger.debug(f"{translations['dereveb_audio']}: {reverb}")
+        if reverb: logger.debug(f"{translations['denoise_dereveb']}: {reverb_denoise}")
+        if reverb: logger.debug(f"{translations['dereveb_backing']}: {backing_reverb}")
+
+        if not mdx: vocals, instruments = separator_music_demucs(input_path, output_path, export_format, shifts, overlap, segments_size, demucs_model)
+        else: vocals, instruments = separator_music_mdx(input_path, output_path, export_format, segments_size, overlap, mdx_denoise, mdx_model, hop_length, batch_size)
+
+        if backing: main_vocals, backing_vocals = separator_backing(vocals, output_path, export_format, segments_size, overlap, backing_denoise, kara_model, hop_length, batch_size)
+        if reverb: vocals_no_reverb, main_vocals_no_reverb, backing_vocals_no_reverb = separator_reverb(output_path, export_format, segments_size, overlap, reverb_denoise, reverb, backing, backing_reverb, hop_length, batch_size)
+
+        original_output = os.path.join(output_path, f"Original_Vocals_No_Reverb.{export_format}") if reverb else os.path.join(output_path, f"Original_Vocals.{export_format}")
+        main_output = os.path.join(output_path, f"Main_Vocals_No_Reverb.{export_format}") if reverb and backing else os.path.join(output_path, f"Main_Vocals.{export_format}")
+        backing_output = os.path.join(output_path, f"Backing_Vocals_No_Reverb.{export_format}") if reverb and backing_reverb else os.path.join(output_path, f"Backing_Vocals.{export_format}")
+        
+        if clean_audio:
+            logger.info(f"{translations['clear_audio']}...")
+            vocal_data, vocal_sr = sf.read(vocals_no_reverb if reverb else vocals)
+            main_data, main_sr = sf.read(main_vocals_no_reverb if reverb and backing else main_vocals)
+            backing_data, backing_sr = sf.read(backing_vocals_no_reverb if reverb and backing_reverb else backing_vocals)
+
+            vocals_clean = nr.reduce_noise(y=vocal_data, prop_decrease=clean_strength)
+            
+            sf.write(original_output, vocals_clean, vocal_sr, format=export_format)
+
+            if backing:
+                mains_clean = nr.reduce_noise(y=main_data, sr=main_sr, prop_decrease=clean_strength)
+                backing_clean = nr.reduce_noise(y=backing_data, sr=backing_sr, prop_decrease=clean_strength)
+                sf.write(main_output, mains_clean, main_sr, format=export_format)
+                sf.write(backing_output, backing_clean, backing_sr, format=export_format)          
+
+            logger.info(translations["clean_audio_success"])
+            return original_output, instruments, main_output, backing_output
+    except Exception as e:
+        logger.error(f"{translations['separator_error']}: {e}")
+        return None, None, None, None
+    
+    elapsed_time = time.time() - start_time
+    logger.info(translations["separator_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
+    
+
+def separator_music_demucs(input, output, format, shifts, overlap, segments_size, demucs_model):
+    if not os.path.exists(input): 
+        logger.warning(translations["input_not_valid"])
+        sys.exit(1)
+    
+    if not os.path.exists(output): 
+        logger.warning(translations["output_not_valid"])
+        sys.exit(1)
+    
+    for i in [f"Original_Vocals.{format}", f"Instruments.{format}"]:
+        if os.path.exists(os.path.join(output, i)): os.remove(os.path.join(output, i))
+
+    model = demucs_models.get(demucs_model)
+
+    segment_size = segments_size / 2
+
+    logger.info(f"{translations['separator_process_2']}...")
+
+    demucs_output = separator_main(audio_file=input, model_filename=model, output_format=format, output_dir=output, demucs_segment_size=segment_size, demucs_shifts=shifts, demucs_overlap=overlap)
+    
+    for f in demucs_output:
+        path = os.path.join(output, f)
+
+        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
+
+        if '_(Drums)_' in f: drums = path
+        elif '_(Bass)_' in f: bass = path
+        elif '_(Other)_' in f: other = path
+        elif '_(Vocals)_' in f: os.rename(path, os.path.join(output, f"Original_Vocals.{format}"))
+
+    AudioSegment.from_file(drums).overlay(AudioSegment.from_file(bass)).overlay(AudioSegment.from_file(other)).export(os.path.join(output, f"Instruments.{format}"), format=format)
+
+    for f in [drums, bass, other]:
+        if os.path.exists(f): os.remove(f)
+    
+    logger.info(translations["separator_success_2"])
+    return os.path.join(output, f"Original_Vocals.{format}"), os.path.join(output, f"Instruments.{format}")
+
+def separator_backing(input, output, format, segments_size, overlap, denoise, kara_model, hop_length, batch_size):
+    if not os.path.exists(input): 
+        logger.warning(translations["input_not_valid"])
+        sys.exit(1)
+    
+    if not os.path.exists(output): 
+        logger.warning(translations["output_not_valid"])
+        sys.exit(1)
+    
+    for f in [f"Main_Vocals.{format}", f"Backing_Vocals.{format}"]:
+        if os.path.exists(os.path.join(output, f)): os.remove(os.path.join(output, f))
+
+    model_2 = kara_models.get(kara_model)
+
+    logger.info(f"{translations['separator_process_backing']}...")
+    backing_outputs = separator_main(audio_file=input, model_filename=model_2, output_format=format, output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise)
+
+    main_output = os.path.join(output, f"Main_Vocals.{format}")
+    backing_output = os.path.join(output, f"Backing_Vocals.{format}")
+
+    for f in backing_outputs:
+        path = os.path.join(output, f)
+
+        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
+
+        if '_(Instrumental)_' in f: os.rename(path, backing_output)
+        elif '_(Vocals)_' in f: os.rename(path, main_output)
+
+    logger.info(translations["separator_process_backing_success"])
+    return main_output, backing_output
+
+def separator_music_mdx(input, output, format, segments_size, overlap, denoise, mdx_model, hop_length, batch_size):
+    if not os.path.exists(input): 
+        logger.warning(translations["input_not_valid"])
+        sys.exit(1)
+    
+    if not os.path.exists(output): 
+        logger.warning(translations["output_not_valid"])
+        sys.exit(1)
+
+    for i in [f"Original_Vocals.{format}", f"Instruments.{format}"]:
+        if os.path.exists(os.path.join(output, i)): os.remove(os.path.join(output, i))
+    
+    model_3 = mdx_models.get(mdx_model)
+
+    logger.info(f"{translations['separator_process_2']}...")
+    output_music = separator_main(audio_file=input, model_filename=model_3, output_format=format, output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise)
+    
+    original_output = os.path.join(output, f"Original_Vocals.{format}")
+    instruments_output = os.path.join(output, f"Instruments.{format}")
+
+    for f in output_music:
+        path = os.path.join(output, f)
+
+        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
+
+        if '_(Instrumental)_' in f: os.rename(path, instruments_output)
+        elif '_(Vocals)_' in f: os.rename(path, original_output)
+
+    logger.info(translations["separator_process_backing_success"])
+    return original_output, instruments_output
+
+def separator_reverb(output, format, segments_size, overlap, denoise, original, main, backing_reverb, hop_length, batch_size):
+    if not os.path.exists(output): 
+        logger.warning(translations["output_not_valid"])
+        sys.exit(1)
+    
+    for i in [f"Original_Vocals_Reverb.{format}", f"Main_Vocals_Reverb.{format}", f"Original_Vocals_No_Reverb.{format}", f"Main_Vocals_No_Reverb.{format}"]:
+        if os.path.exists(os.path.join(output, i)): os.remove(os.path.join(output, i))
+
+    dereveb_path = []
+
+    if original: 
+        try:
+            dereveb_path.append(os.path.join(output, [f for f in os.listdir(output) if 'Original_Vocals' in f][0]))
+        except IndexError:
+            logger.warning(translations["not_found_original_vocal"])
+            sys.exit(1)
+        
+    if main:
+        try:
+            dereveb_path.append(os.path.join(output, [f for f in os.listdir(output) if 'Main_Vocals' in f][0]))
+        except IndexError:
+            logger.warning(translations["not_found_main_vocal"])
+            sys.exit(1)
+    
+    if backing_reverb:
+        try:
+            dereveb_path.append(os.path.join(output, [f for f in os.listdir(output) if 'Backing_Vocals' in f][0]))
+        except IndexError:
+            logger.warning(translations["not_found_backing_vocal"])
+            sys.exit(1)
+        
+    for path in dereveb_path:
+        if not os.path.exists(path): 
+            logger.warning(translations["not_found"].format(name=path))
+            sys.exit(1)
+        
+        if "Original_Vocals" in path: 
+            reverb_path = os.path.join(output, f"Original_Vocals_Reverb.{format}")
+            no_reverb_path = os.path.join(output, f"Original_Vocals_No_Reverb.{format}")
+            start_title = translations["process_original"]
+            end_title = translations["process_original_success"]
+        elif "Main_Vocals" in path:
+            reverb_path = os.path.join(output, f"Main_Vocals_Reverb.{format}")
+            no_reverb_path = os.path.join(output, f"Main_Vocals_No_Reverb.{format}")
+            start_title = translations["process_main"]
+            end_title = translations["process_main_success"]
+        elif "Backing_Vocals" in path:
+            reverb_path = os.path.join(output, f"Backing_Vocals_Reverb.{format}")
+            no_reverb_path = os.path.join(output, f"Backing_Vocals_No_Reverb.{format}")
+            start_title = translations["process_backing"]
+            end_title = translations["process_backing_success"]
+
+        logger.info(start_title)
+        output_dereveb = separator_main(audio_file=path, model_filename="Reverb_HQ_By_FoxJoy.onnx", output_format=format, output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise)
+
+        for f in output_dereveb:
+            path = os.path.join(output, f)
+
+            if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
+
+            if '_(Reverb)_' in f: os.rename(path, reverb_path)
+            elif '_(No Reverb)_' in f: os.rename(path, no_reverb_path)
+
+        logger.info(end_title)
+
+    return (os.path.join(output, f"Original_Vocals_No_Reverb.{format}") if original else None), (os.path.join(output, f"Main_Vocals_No_Reverb.{format}") if main else None), (os.path.join(output, f"Backing_Vocals_No_Reverb.{format}") if backing_reverb else None)
+
+def separator_main(audio_file=None, model_filename="UVR-MDX-NET_Main_340.onnx", output_format="wav", output_dir=".", mdx_segment_size=256, mdx_overlap=0.25, mdx_batch_size=1, mdx_hop_length=1024, mdx_enable_denoise=True, demucs_segment_size=256, demucs_shifts=2, demucs_overlap=0.25):
+    separator = Separator(
+        log_formatter=file_formatter,
+        log_level=logging.INFO,
+        output_dir=output_dir,
+        output_format=output_format,
+        output_bitrate=None,
+        normalization_threshold=0.9,
+        output_single_stem=None,
+        invert_using_spec=False,
+        sample_rate=44100,
+        mdx_params={
+            "hop_length": mdx_hop_length,
+            "segment_size": mdx_segment_size,
+            "overlap": mdx_overlap,
+            "batch_size": mdx_batch_size,
+            "enable_denoise": mdx_enable_denoise,
+        },
+        demucs_params={
+            "segment_size": demucs_segment_size,
+            "shifts": demucs_shifts,
+            "overlap": demucs_overlap,
+            "segments_enabled": True,
+        }
+    )
+
+    separator.load_model(model_filename=model_filename)
+    return separator.separate(audio_file)
+
+if __name__ == "__main__": main()

main/inference/train.py

@@ -0,0 +1,1600 @@
+import os
+import re
+import sys
+import glob
+import json
+import torch
+import logging
+import hashlib
+import argparse
+import datetime
+import warnings
+import logging.handlers
+
+import numpy as np
+import torch.utils.data
+import matplotlib.pyplot as plt
+import torch.distributed as dist
+import torch.multiprocessing as mp
+
+from tqdm import tqdm
+from time import time as ttime
+from scipy.io.wavfile import read
+from collections import OrderedDict
+from random import randint, shuffle
+ 
+from torch.nn import functional as F
+from distutils.util import strtobool
+from torch.utils.data import DataLoader
+from torch.cuda.amp import GradScaler, autocast
+from torch.utils.tensorboard import SummaryWriter
+from librosa.filters import mel as librosa_mel_fn
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils.parametrizations import spectral_norm, weight_norm
+
+current_dir = os.getcwd()
+sys.path.append(current_dir)
+
+from main.configs.config import Config
+from main.library.algorithm.residuals import LRELU_SLOPE
+from main.library.algorithm.synthesizers import Synthesizer
+from main.library.algorithm.commons import get_padding, slice_segments, clip_grad_value
+
+warnings.filterwarnings("ignore", category=FutureWarning)
+warnings.filterwarnings("ignore", category=UserWarning)
+
+logging.getLogger("torch").setLevel(logging.ERROR)
+
+MATPLOTLIB_FLAG = False
+
+translations = Config().translations
+
+class HParams:
+    def __init__(self, **kwargs):
+        for k, v in kwargs.items():
+            self[k] = HParams(**v) if isinstance(v, dict) else v
+
+
+    def keys(self):
+        return self.__dict__.keys()
+    
+
+    def items(self):
+        return self.__dict__.items()
+    
+
+    def values(self):
+        return self.__dict__.values()
+    
+
+    def __len__(self):
+        return len(self.__dict__)
+    
+
+    def __getitem__(self, key):
+        return self.__dict__[key]
+    
+
+    def __setitem__(self, key, value):
+        self.__dict__[key] = value
+
+
+    def __contains__(self, key):
+        return key in self.__dict__
+    
+
+    def __repr__(self):
+        return repr(self.__dict__)
+
+def parse_arguments() -> tuple:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_name", type=str, required=True)
+    parser.add_argument("--rvc_version", type=str, default="v2")
+    parser.add_argument("--save_every_epoch", type=int, required=True)
+    parser.add_argument("--save_only_latest", type=lambda x: bool(strtobool(x)), default=True)
+    parser.add_argument("--save_every_weights", type=lambda x: bool(strtobool(x)), default=True)
+    parser.add_argument("--total_epoch", type=int, default=300)
+    parser.add_argument("--sample_rate", type=int, required=True)
+    parser.add_argument("--batch_size", type=int, default=8)
+    parser.add_argument("--gpu", type=str, default="0")
+    parser.add_argument("--pitch_guidance", type=lambda x: bool(strtobool(x)), default=True)
+    parser.add_argument("--g_pretrained_path", type=str, default="")
+    parser.add_argument("--d_pretrained_path", type=str, default="")
+    parser.add_argument("--overtraining_detector", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--overtraining_threshold", type=int, default=50)
+    parser.add_argument("--sync_graph", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--cache_data_in_gpu", type=lambda x: bool(strtobool(x)), default=False)
+    parser.add_argument("--model_author", type=str)
+
+    args = parser.parse_args()
+    return args
+
+
+args = parse_arguments()
+
+model_name = args.model_name
+save_every_epoch = args.save_every_epoch
+total_epoch = args.total_epoch
+pretrainG = args.g_pretrained_path
+pretrainD = args.d_pretrained_path
+version = args.rvc_version
+gpus = args.gpu
+batch_size = args.batch_size
+sample_rate = args.sample_rate
+pitch_guidance = args.pitch_guidance
+save_only_latest = args.save_only_latest
+save_every_weights = args.save_every_weights
+cache_data_in_gpu = args.cache_data_in_gpu
+overtraining_detector = args.overtraining_detector
+overtraining_threshold = args.overtraining_threshold
+sync_graph = args.sync_graph
+model_author = args.model_author
+
+experiment_dir = os.path.join(current_dir, "assets", "logs", model_name)
+config_save_path = os.path.join(experiment_dir, "config.json")
+
+os.environ["CUDA_VISIBLE_DEVICES"] = gpus.replace("-", ",")
+n_gpus = len(gpus.split("-"))
+
+torch.backends.cudnn.deterministic = False
+torch.backends.cudnn.benchmark = False
+
+global_step = 0
+last_loss_gen_all = 0
+overtrain_save_epoch = 0
+
+loss_gen_history = []
+smoothed_loss_gen_history = []
+loss_disc_history = []
+smoothed_loss_disc_history = []
+
+lowest_value = {"step": 0, "value": float("inf"), "epoch": 0}
+training_file_path = os.path.join(experiment_dir, "training_data.json")
+
+with open(config_save_path, "r") as f:
+    config = json.load(f)
+
+config = HParams(**config)
+config.data.training_files = os.path.join(experiment_dir, "filelist.txt")
+
+log_file = os.path.join(experiment_dir, "train.log")
+
+logger = logging.getLogger(__name__)
+
+if logger.hasHandlers(): logger.handlers.clear()
+else:  
+    console_handler = logging.StreamHandler()
+    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    console_handler.setFormatter(console_formatter)
+    console_handler.setLevel(logging.INFO)
+
+    file_handler = logging.handlers.RotatingFileHandler(log_file, maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
+    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
+
+    file_handler.setFormatter(file_formatter)
+    file_handler.setLevel(logging.DEBUG)
+
+    logger.addHandler(console_handler)
+    logger.addHandler(file_handler)
+    logger.setLevel(logging.DEBUG)
+
+logger.debug(f"{translations['modelname']}: {model_name}")
+logger.debug(translations["save_every_epoch"].format(save_every_epoch=save_every_epoch))
+logger.debug(translations["total_e"].format(total_epoch=total_epoch))
+logger.debug(translations["dorg"].format(pretrainG=pretrainG, pretrainD=pretrainD))
+logger.debug(f"{translations['training_version']}: {version}")
+logger.debug(f"Gpu: {gpus}")
+logger.debug(f"{translations['batch_size']}: {batch_size}")
+logger.debug(f"{translations['pretrain_sr']}: {sample_rate}")
+logger.debug(f"{translations['training_f0']}: {pitch_guidance}")
+logger.debug(f"{translations['save_only_latest']}: {save_only_latest}")
+logger.debug(f"{translations['save_every_weights']}: {save_every_weights}")
+logger.debug(f"{translations['cache_in_gpu']}: {cache_data_in_gpu}")
+logger.debug(f"{translations['overtraining_detector']}: {overtraining_detector}")
+logger.debug(f"{translations['threshold']}: {overtraining_threshold}")
+logger.debug(f"{translations['sync_graph']}: {sync_graph}")
+if not model_author: logger.debug(translations["model_author"].format(model_author=model_author))
+
+def main():
+    global training_file_path, last_loss_gen_all, smoothed_loss_gen_history, loss_gen_history, loss_disc_history, smoothed_loss_disc_history, overtrain_save_epoch, model_author
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(randint(20000, 55555))
+
+    if torch.cuda.is_available():
+        device = torch.device("cuda")
+        n_gpus = torch.cuda.device_count()
+    elif torch.backends.mps.is_available():
+        device = torch.device("mps")
+        n_gpus = 1
+    else:
+        device = torch.device("cpu")
+        n_gpus = 1
+
+    def start():
+        children = []
+
+        for i in range(n_gpus):
+            subproc = mp.Process(target=run, args=(i, n_gpus, experiment_dir, pretrainG, pretrainD, pitch_guidance, custom_total_epoch, custom_save_every_weights, config, device, model_author))
+            children.append(subproc)
+            subproc.start()
+
+        for i in range(n_gpus):
+            children[i].join()
+
+    def load_from_json(file_path):
+        if os.path.exists(file_path):
+            with open(file_path, "r") as f:
+                data = json.load(f)
+
+                return (
+                    data.get("loss_disc_history", []),
+                    data.get("smoothed_loss_disc_history", []),
+                    data.get("loss_gen_history", []),
+                    data.get("smoothed_loss_gen_history", []),
+                )
+            
+        return [], [], [], []
+
+    def continue_overtrain_detector(training_file_path):
+        if overtraining_detector:
+            if os.path.exists(training_file_path):
+                (
+                    loss_disc_history,
+                    smoothed_loss_disc_history,
+                    loss_gen_history,
+                    smoothed_loss_gen_history,
+                ) = load_from_json(training_file_path)
+
+
+    n_gpus = torch.cuda.device_count()
+
+    if not torch.cuda.is_available() and torch.backends.mps.is_available(): n_gpus = 1
+
+    if n_gpus < 1:
+        logger.warning(translations["not_gpu"])
+        n_gpus = 1
+
+    if sync_graph:
+        logger.debug(translations["sync"])
+        custom_total_epoch = 1
+        custom_save_every_weights = True
+
+        start()
+
+        model_config_file = os.path.join(experiment_dir, "config.json")
+        rvc_config_file = os.path.join(current_dir, "main", "configs", version, str(sample_rate) + ".json")
+
+        if not os.path.exists(rvc_config_file): rvc_config_file = os.path.join(current_dir, "main", "configs", "v1", str(sample_rate) + ".json")
+
+        pattern = rf"{os.path.basename(model_name)}_(\d+)e_(\d+)s\.pth"
+
+        for filename in os.listdir(os.path.join("assets", "weights")):
+            match = re.match(pattern, filename)
+
+            if match: steps = int(match.group(2))
+
+        def edit_config(config_file):
+            with open(config_file, "r", encoding="utf8") as json_file:
+                config_data = json.load(json_file)
+
+            config_data["train"]["log_interval"] = steps
+
+            with open(config_file, "w", encoding="utf8") as json_file:
+                json.dump(config_data, json_file, indent=2, separators=(",", ": "), ensure_ascii=False)
+
+        edit_config(model_config_file)
+        edit_config(rvc_config_file)
+
+        for root, dirs, files in os.walk(experiment_dir, topdown=False):
+            for name in files:
+                file_path = os.path.join(root, name)
+                _, file_extension = os.path.splitext(name)
+
+                if file_extension == ".0": os.remove(file_path)
+                elif ("D" in name or "G" in name) and file_extension == ".pth": os.remove(file_path)
+                elif ("added" in name or "trained" in name) and file_extension == ".index": os.remove(file_path)
+
+            for name in dirs:
+                if name == "eval":
+                    folder_path = os.path.join(root, name)
+
+                    for item in os.listdir(folder_path):
+                        item_path = os.path.join(folder_path, item)
+                        if os.path.isfile(item_path): os.remove(item_path)
+
+                    os.rmdir(folder_path)
+
+        logger.info(translations["sync_success"])
+        custom_total_epoch = total_epoch
+        custom_save_every_weights = save_every_weights
+
+        continue_overtrain_detector(training_file_path)
+        start()
+    else:
+        custom_total_epoch = total_epoch
+        custom_save_every_weights = save_every_weights
+
+        continue_overtrain_detector(training_file_path)
+        start()
+
+
+def plot_spectrogram_to_numpy(spectrogram):
+    global MATPLOTLIB_FLAG
+
+    if not MATPLOTLIB_FLAG:
+        plt.switch_backend("Agg")
+        MATPLOTLIB_FLAG = True
+
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close(fig)
+    return data
+
+
+def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sample_rate=22050):
+    for k, v in scalars.items():
+        writer.add_scalar(k, v, global_step)
+
+    for k, v in histograms.items():
+        writer.add_histogram(k, v, global_step)
+
+    for k, v in images.items():
+        writer.add_image(k, v, global_step, dataformats="HWC")
+
+    for k, v in audios.items():
+        writer.add_audio(k, v, global_step, audio_sample_rate)
+
+
+def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
+    assert os.path.isfile(checkpoint_path), translations["not_found_checkpoint"].format(checkpoint_path=checkpoint_path)
+
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
+    checkpoint_dict = replace_keys_in_dict(replace_keys_in_dict(checkpoint_dict, ".weight_v", ".parametrizations.weight.original1"), ".weight_g", ".parametrizations.weight.original0")
+
+    model_state_dict = (model.module.state_dict() if hasattr(model, "module") else model.state_dict())
+    new_state_dict = {k: checkpoint_dict["model"].get(k, v) for k, v in model_state_dict.items()}
+
+    if hasattr(model, "module"): model.module.load_state_dict(new_state_dict, strict=False)
+    else: model.load_state_dict(new_state_dict, strict=False)
+
+    if optimizer and load_opt == 1: optimizer.load_state_dict(checkpoint_dict.get("optimizer", {}))
+
+    logger.debug(translations["save_checkpoint"].format(checkpoint_path=checkpoint_path, checkpoint_dict=checkpoint_dict['iteration']))
+
+    return (
+        model,
+        optimizer,
+        checkpoint_dict.get("learning_rate", 0),
+        checkpoint_dict["iteration"],
+    )
+
+
+def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
+    state_dict = (model.module.state_dict() if hasattr(model, "module") else model.state_dict())
+
+    checkpoint_data = {
+        "model": state_dict,
+        "iteration": iteration,
+        "optimizer": optimizer.state_dict(),
+        "learning_rate": learning_rate,
+    }
+
+    torch.save(checkpoint_data, checkpoint_path)
+
+    old_version_path = checkpoint_path.replace(".pth", "_old_version.pth")
+    checkpoint_data = replace_keys_in_dict(replace_keys_in_dict(checkpoint_data, ".parametrizations.weight.original1", ".weight_v"), ".parametrizations.weight.original0", ".weight_g")
+
+    torch.save(checkpoint_data, old_version_path)
+
+    os.replace(old_version_path, checkpoint_path)
+    logger.info(translations["save_model"].format(checkpoint_path=checkpoint_path, iteration=iteration))
+
+
+def latest_checkpoint_path(dir_path, regex="G_*.pth"):
+    checkpoints = sorted(glob.glob(os.path.join(dir_path, regex)), key=lambda f: int("".join(filter(str.isdigit, f))))
+
+    return checkpoints[-1] if checkpoints else None
+
+
+def load_wav_to_torch(full_path):
+    sample_rate, data = read(full_path)
+    
+    return torch.FloatTensor(data.astype(np.float32)), sample_rate
+
+
+def load_filepaths_and_text(filename, split="|"):
+    with open(filename, encoding="utf-8") as f:
+        return [line.strip().split(split) for line in f]
+
+
+def feature_loss(fmap_r, fmap_g):
+    loss = 0
+
+    for dr, dg in zip(fmap_r, fmap_g):
+        for rl, gl in zip(dr, dg):
+            rl = rl.float().detach()
+            gl = gl.float()
+            loss += torch.mean(torch.abs(rl - gl))
+
+    return loss * 2
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+    loss = 0
+    r_losses = []
+    g_losses = []
+
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        dr = dr.float()
+        dg = dg.float()
+        r_loss = torch.mean((1 - dr) ** 2)
+        g_loss = torch.mean(dg**2)
+        loss += r_loss + g_loss
+        r_losses.append(r_loss.item())
+        g_losses.append(g_loss.item())
+
+    return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+    loss = 0
+    gen_losses = []
+
+    for dg in disc_outputs:
+        dg = dg.float()
+        l = torch.mean((1 - dg) ** 2)
+        gen_losses.append(l)
+        loss += l
+
+    return loss, gen_losses
+
+
+def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
+    z_p = z_p.float()
+    logs_q = logs_q.float()
+    m_p = m_p.float()
+    logs_p = logs_p.float()
+    z_mask = z_mask.float()
+
+    kl = logs_p - logs_q - 0.5
+    kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)
+    kl = torch.sum(kl * z_mask)
+    l = kl / torch.sum(z_mask)
+
+    return l
+
+
+class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
+    def __init__(self, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(hparams.training_files)
+        self.max_wav_value = hparams.max_wav_value
+        self.sample_rate = hparams.sample_rate
+        self.filter_length = hparams.filter_length
+        self.hop_length = hparams.hop_length
+        self.win_length = hparams.win_length
+        self.sample_rate = hparams.sample_rate
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 5000)
+        self._filter()
+
+    def _filter(self):
+        audiopaths_and_text_new = []
+        lengths = []
+
+        for audiopath, text, pitch, pitchf, dv in self.audiopaths_and_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_and_text_new.append([audiopath, text, pitch, pitchf, dv])
+                lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_sid(self, sid):
+        try:
+            sid = torch.LongTensor([int(sid)])
+        except ValueError as e:
+            logger.error(translations["sid_error"].format(sid=sid, e=e))
+            sid = torch.LongTensor([0])
+
+        return sid
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        file = audiopath_and_text[0]
+        phone = audiopath_and_text[1]
+        pitch = audiopath_and_text[2]
+        pitchf = audiopath_and_text[3]
+        dv = audiopath_and_text[4]
+
+        phone, pitch, pitchf = self.get_labels(phone, pitch, pitchf)
+        spec, wav = self.get_audio(file)
+        dv = self.get_sid(dv)
+
+        len_phone = phone.size()[0]
+        len_spec = spec.size()[-1]
+        if len_phone != len_spec:
+            len_min = min(len_phone, len_spec)
+            len_wav = len_min * self.hop_length
+
+            spec = spec[:, :len_min]
+            wav = wav[:, :len_wav]
+
+            phone = phone[:len_min, :]
+            pitch = pitch[:len_min]
+            pitchf = pitchf[:len_min]
+
+        return (spec, wav, phone, pitch, pitchf, dv)
+
+    def get_labels(self, phone, pitch, pitchf):
+        phone = np.load(phone)
+        phone = np.repeat(phone, 2, axis=0)
+
+        pitch = np.load(pitch)
+        pitchf = np.load(pitchf)
+
+        n_num = min(phone.shape[0], 900)
+        phone = phone[:n_num, :]
+
+        pitch = pitch[:n_num]
+        pitchf = pitchf[:n_num]
+
+        phone = torch.FloatTensor(phone)
+
+        pitch = torch.LongTensor(pitch)
+        pitchf = torch.FloatTensor(pitchf)
+
+        return phone, pitch, pitchf
+
+    def get_audio(self, filename):
+        audio, sample_rate = load_wav_to_torch(filename)
+
+        if sample_rate != self.sample_rate: raise ValueError(translations["sr_does_not_match"].format(sample_rate=sample_rate, sample_rate2=self.sample_rate))
+
+        audio_norm = audio
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+
+        if os.path.exists(spec_filename):
+            try:
+                spec = torch.load(spec_filename)
+            except Exception as e:
+                logger.error(translations["spec_error"].format(spec_filename=spec_filename, e=e))
+                spec = spectrogram_torch(
+                    audio_norm,
+                    self.filter_length,
+                    self.hop_length,
+                    self.win_length,
+                    center=False,
+                )
+                spec = torch.squeeze(spec, 0)
+
+                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        else:
+            spec = spectrogram_torch(
+                audio_norm,
+                self.filter_length,
+                self.hop_length,
+                self.win_length,
+                center=False,
+            )
+            spec = torch.squeeze(spec, 0)
+
+            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        return spec, audio_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollateMultiNSFsid:
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        _, ids_sorted_decreasing = torch.sort(torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True)
+
+        max_spec_len = max([x[0].size(1) for x in batch])
+        max_wave_len = max([x[1].size(1) for x in batch])
+
+        spec_lengths = torch.LongTensor(len(batch))
+        wave_lengths = torch.LongTensor(len(batch))
+        spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+        wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+
+        spec_padded.zero_()
+        wave_padded.zero_()
+
+        max_phone_len = max([x[2].size(0) for x in batch])
+
+        phone_lengths = torch.LongTensor(len(batch))
+        phone_padded = torch.FloatTensor(len(batch), max_phone_len, batch[0][2].shape[1])
+        pitch_padded = torch.LongTensor(len(batch), max_phone_len)
+        pitchf_padded = torch.FloatTensor(len(batch), max_phone_len)
+
+        phone_padded.zero_()
+        pitch_padded.zero_()
+        pitchf_padded.zero_()
+        sid = torch.LongTensor(len(batch))
+
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            spec = row[0]
+            spec_padded[i, :, : spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wave = row[1]
+            wave_padded[i, :, : wave.size(1)] = wave
+            wave_lengths[i] = wave.size(1)
+
+            phone = row[2]
+            phone_padded[i, : phone.size(0), :] = phone
+            phone_lengths[i] = phone.size(0)
+
+            pitch = row[3]
+            pitch_padded[i, : pitch.size(0)] = pitch
+            pitchf = row[4]
+            pitchf_padded[i, : pitchf.size(0)] = pitchf
+
+            sid[i] = row[5]
+
+        return (
+            phone_padded,
+            phone_lengths,
+            pitch_padded,
+            pitchf_padded,
+            spec_padded,
+            spec_lengths,
+            wave_padded,
+            wave_lengths,
+            sid,
+        )
+
+
+class TextAudioLoader(torch.utils.data.Dataset):
+    def __init__(self, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(hparams.training_files)
+        self.max_wav_value = hparams.max_wav_value
+        self.sample_rate = hparams.sample_rate
+        self.filter_length = hparams.filter_length
+        self.hop_length = hparams.hop_length
+        self.win_length = hparams.win_length
+        self.sample_rate = hparams.sample_rate
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 5000)
+        self._filter()
+
+    def _filter(self):
+        audiopaths_and_text_new = []
+        lengths = []
+
+        for entry in self.audiopaths_and_text:
+            if len(entry) >= 3:
+                audiopath, text, dv = entry[:3]
+
+                if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                    audiopaths_and_text_new.append([audiopath, text, dv])
+                    lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_sid(self, sid):
+        try:
+            sid = torch.LongTensor([int(sid)])
+        except ValueError as e:
+            logger.error(translations["sid_error"].format(sid=sid, e=e))
+            sid = torch.LongTensor([0])
+
+        return sid
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        file = audiopath_and_text[0]
+        phone = audiopath_and_text[1]
+        dv = audiopath_and_text[2]
+
+        phone = self.get_labels(phone)
+        spec, wav = self.get_audio(file)
+        dv = self.get_sid(dv)
+
+        len_phone = phone.size()[0]
+        len_spec = spec.size()[-1]
+
+        if len_phone != len_spec:
+            len_min = min(len_phone, len_spec)
+            len_wav = len_min * self.hop_length
+            spec = spec[:, :len_min]
+            wav = wav[:, :len_wav]
+            phone = phone[:len_min, :]
+
+        return (spec, wav, phone, dv)
+
+    def get_labels(self, phone):
+        phone = np.load(phone)
+        phone = np.repeat(phone, 2, axis=0)
+        n_num = min(phone.shape[0], 900)
+        phone = phone[:n_num, :]
+        phone = torch.FloatTensor(phone)
+        return phone
+
+    def get_audio(self, filename):
+        audio, sample_rate = load_wav_to_torch(filename)
+
+        if sample_rate != self.sample_rate: raise ValueError(translations["sr_does_not_match"].format(sample_rate=sample_rate, sample_rate2=self.sample_rate))
+
+        audio_norm = audio
+        audio_norm = audio_norm.unsqueeze(0)
+
+        spec_filename = filename.replace(".wav", ".spec.pt")
+
+        if os.path.exists(spec_filename):
+            try:
+                spec = torch.load(spec_filename)
+            except Exception as e:
+                logger.error(translations["spec_error"].format(spec_filename=spec_filename, e=e))
+                spec = spectrogram_torch(
+                    audio_norm,
+                    self.filter_length,
+                    self.hop_length,
+                    self.win_length,
+                    center=False,
+                )
+                spec = torch.squeeze(spec, 0)
+
+                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        else:
+            spec = spectrogram_torch(
+                audio_norm,
+                self.filter_length,
+                self.hop_length,
+                self.win_length,
+                center=False,
+            )
+            spec = torch.squeeze(spec, 0)
+
+            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        return spec, audio_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollate:
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        _, ids_sorted_decreasing = torch.sort(torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True)
+
+        max_spec_len = max([x[0].size(1) for x in batch])
+        max_wave_len = max([x[1].size(1) for x in batch])
+
+        spec_lengths = torch.LongTensor(len(batch))
+        wave_lengths = torch.LongTensor(len(batch))
+
+        spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+        wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+
+        spec_padded.zero_()
+        wave_padded.zero_()
+
+        max_phone_len = max([x[2].size(0) for x in batch])
+
+        phone_lengths = torch.LongTensor(len(batch))
+        phone_padded = torch.FloatTensor(len(batch), max_phone_len, batch[0][2].shape[1])
+
+        phone_padded.zero_()
+        sid = torch.LongTensor(len(batch))
+
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            spec = row[0]
+            spec_padded[i, :, : spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wave = row[1]
+            wave_padded[i, :, : wave.size(1)] = wave
+            wave_lengths[i] = wave.size(1)
+
+            phone = row[2]
+            phone_padded[i, : phone.size(0), :] = phone
+            phone_lengths[i] = phone.size(0)
+
+            sid[i] = row[3]
+
+        return (
+            phone_padded,
+            phone_lengths,
+            spec_padded,
+            spec_lengths,
+            wave_padded,
+            wave_lengths,
+            sid,
+        )
+
+
+class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
+    def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+        self.lengths = dataset.lengths
+        self.batch_size = batch_size
+        self.boundaries = boundaries
+        self.buckets, self.num_samples_per_bucket = self._create_buckets()
+        self.total_size = sum(self.num_samples_per_bucket)
+        self.num_samples = self.total_size // self.num_replicas
+
+    def _create_buckets(self):
+        buckets = [[] for _ in range(len(self.boundaries) - 1)]
+
+        for i in range(len(self.lengths)):
+            length = self.lengths[i]
+            idx_bucket = self._bisect(length)
+            if idx_bucket != -1: buckets[idx_bucket].append(i)
+
+        for i in range(len(buckets) - 1, -1, -1):  
+            if len(buckets[i]) == 0:
+                buckets.pop(i)
+                self.boundaries.pop(i + 1)
+
+        num_samples_per_bucket = []
+
+        for i in range(len(buckets)):
+            len_bucket = len(buckets[i])
+            total_batch_size = self.num_replicas * self.batch_size
+            rem = (total_batch_size - (len_bucket % total_batch_size)) % total_batch_size
+            num_samples_per_bucket.append(len_bucket + rem)
+
+        return buckets, num_samples_per_bucket
+
+    def __iter__(self):
+        g = torch.Generator()
+        g.manual_seed(self.epoch)
+
+        indices = []
+
+        if self.shuffle:
+            for bucket in self.buckets:
+                indices.append(torch.randperm(len(bucket), generator=g).tolist())
+        else:
+            for bucket in self.buckets:
+                indices.append(list(range(len(bucket))))
+
+        batches = []
+
+        for i in range(len(self.buckets)):
+            bucket = self.buckets[i]
+            len_bucket = len(bucket)
+            ids_bucket = indices[i]
+            num_samples_bucket = self.num_samples_per_bucket[i]
+
+            rem = num_samples_bucket - len_bucket
+
+            ids_bucket = (ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[: (rem % len_bucket)])
+            ids_bucket = ids_bucket[self.rank :: self.num_replicas]
+
+            for j in range(len(ids_bucket) // self.batch_size):
+                batch = [bucket[idx] for idx in ids_bucket[j * self.batch_size : (j + 1) * self.batch_size]]
+                batches.append(batch)
+
+        if self.shuffle:
+            batch_ids = torch.randperm(len(batches), generator=g).tolist()
+            batches = [batches[i] for i in batch_ids]
+
+        self.batches = batches
+
+        assert len(self.batches) * self.batch_size == self.num_samples
+        return iter(self.batches)
+
+    def _bisect(self, x, lo=0, hi=None):
+        if hi is None: hi = len(self.boundaries) - 1
+
+        if hi > lo:
+            mid = (hi + lo) // 2
+
+            if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]: return mid
+            elif x <= self.boundaries[mid]: return self._bisect(x, lo, mid)
+            else: return self._bisect(x, mid + 1, hi)
+
+        else: return -1
+
+    def __len__(self):
+        return self.num_samples // self.batch_size
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17]
+        self.discriminators = torch.nn.ModuleList([DiscriminatorS(use_spectral_norm=use_spectral_norm)] + [DiscriminatorP(p, use_spectral_norm=use_spectral_norm) for p in periods])
+
+    def forward(self, y, y_hat):
+        y_d_rs, y_d_gs, fmap_rs, fmap_gs = [], [], [], []
+
+        for d in self.discriminators:
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiPeriodDiscriminatorV2(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminatorV2, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17, 23, 37]
+        self.discriminators = torch.nn.ModuleList([DiscriminatorS(use_spectral_norm=use_spectral_norm)] + [DiscriminatorP(p, use_spectral_norm=use_spectral_norm) for p in periods])
+
+    def forward(self, y, y_hat):
+        y_d_rs, y_d_gs, fmap_rs, fmap_gs = [], [], [], []
+
+        for d in self.discriminators:
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = spectral_norm if use_spectral_norm else weight_norm
+
+        self.convs = torch.nn.ModuleList([norm_f(torch.nn.Conv1d(1, 16, 15, 1, padding=7)), norm_f(torch.nn.Conv1d(16, 64, 41, 4, groups=4, padding=20)), norm_f(torch.nn.Conv1d(64, 256, 41, 4, groups=16, padding=20)), norm_f(torch.nn.Conv1d(256, 1024, 41, 4, groups=64, padding=20)), norm_f(torch.nn.Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), norm_f(torch.nn.Conv1d(1024, 1024, 5, 1, padding=2))])
+        self.conv_post = norm_f(torch.nn.Conv1d(1024, 1, 3, 1, padding=1))
+        self.lrelu = torch.nn.LeakyReLU(LRELU_SLOPE)
+
+    def forward(self, x):
+        fmap = []
+
+        for conv in self.convs:
+            x = self.lrelu(conv(x))
+            fmap.append(x)
+
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        norm_f = spectral_norm if use_spectral_norm else weight_norm
+
+        in_channels = [1, 32, 128, 512, 1024]
+        out_channels = [32, 128, 512, 1024, 1024]
+
+        self.convs = torch.nn.ModuleList(
+            [
+                norm_f(
+                    torch.nn.Conv2d(
+                        in_ch,
+                        out_ch,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                )
+                for in_ch, out_ch in zip(in_channels, out_channels)
+            ]
+        )
+
+        self.conv_post = norm_f(torch.nn.Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+        self.lrelu = torch.nn.LeakyReLU(LRELU_SLOPE)
+
+    def forward(self, x):
+        fmap = []
+        b, c, t = x.shape
+
+        if t % self.period != 0:
+            n_pad = self.period - (t % self.period)
+            x = torch.nn.functional.pad(x, (0, n_pad), "reflect")
+
+        x = x.view(b, c, -1, self.period)
+
+        for conv in self.convs:
+            x = self.lrelu(conv(x))
+            fmap.append(x)
+
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+        return x, fmap
+
+
+class EpochRecorder:
+    def __init__(self):
+        self.last_time = ttime()
+
+    def record(self):
+        now_time = ttime()
+        elapsed_time = now_time - self.last_time
+        self.last_time = now_time
+        elapsed_time = round(elapsed_time, 1)
+        elapsed_time_str = str(datetime.timedelta(seconds=int(elapsed_time)))
+        current_time = datetime.datetime.now().strftime("%H:%M:%S")
+        return translations["time_or_speed_training"].format(current_time=current_time, elapsed_time_str=elapsed_time_str)
+    
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    return dynamic_range_compression_torch(magnitudes)
+
+
+def spectral_de_normalize_torch(magnitudes):
+    return dynamic_range_decompression_torch(magnitudes)
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def spectrogram_torch(y, n_fft, hop_size, win_size, center=False):
+    global hann_window
+    dtype_device = str(y.dtype) + "_" + str(y.device)
+    wnsize_dtype_device = str(win_size) + "_" + dtype_device
+    if wnsize_dtype_device not in hann_window: hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
+
+    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect")
+
+    y = y.squeeze(1)
+
+    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device], center=center, pad_mode="reflect", normalized=False, onesided=True, return_complex=True)
+
+    spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-6)
+    return spec
+
+
+def spec_to_mel_torch(spec, n_fft, num_mels, sample_rate, fmin, fmax):
+    global mel_basis
+    dtype_device = str(spec.dtype) + "_" + str(spec.device)
+    fmax_dtype_device = str(fmax) + "_" + dtype_device
+    
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(sr=sample_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
+        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
+
+    melspec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+    melspec = spectral_normalize_torch(melspec)
+    return melspec
+
+
+def mel_spectrogram_torch(y, n_fft, num_mels, sample_rate, hop_size, win_size, fmin, fmax, center=False):
+    spec = spectrogram_torch(y, n_fft, hop_size, win_size, center)
+
+    melspec = spec_to_mel_torch(spec, n_fft, num_mels, sample_rate, fmin, fmax)
+
+    return melspec
+
+
+def replace_keys_in_dict(d, old_key_part, new_key_part):
+    updated_dict = OrderedDict() if isinstance(d, OrderedDict) else {}
+
+    for key, value in d.items():
+        new_key = (key.replace(old_key_part, new_key_part) if isinstance(key, str) else key)
+        updated_dict[new_key] = (replace_keys_in_dict(value, old_key_part, new_key_part) if isinstance(value, dict) else value)
+        
+    return updated_dict
+
+
+def extract_model(ckpt, sr, pitch_guidance, name, model_dir, epoch, step, version, hps, model_author):
+    try:
+        logger.info(translations["savemodel"].format(model_dir=model_dir, epoch=epoch, step=step))
+
+        model_dir_path = os.path.join("assets", "weights")
+
+        if "best_epoch" in model_dir: pth_file = f"{name}_{epoch}e_{step}s_best_epoch.pth"
+        else: pth_file = f"{name}_{epoch}e_{step}s.pth"
+
+        pth_file_old_version_path = os.path.join(model_dir_path, f"{pth_file}_old_version.pth")
+
+        opt = OrderedDict(weight={key: value.half() for key, value in ckpt.items() if "enc_q" not in key})
+
+        opt["config"] = [
+            hps.data.filter_length // 2 + 1,
+            32,
+            hps.model.inter_channels,
+            hps.model.hidden_channels,
+            hps.model.filter_channels,
+            hps.model.n_heads,
+            hps.model.n_layers,
+            hps.model.kernel_size,
+            hps.model.p_dropout,
+            hps.model.resblock,
+            hps.model.resblock_kernel_sizes,
+            hps.model.resblock_dilation_sizes,
+            hps.model.upsample_rates,
+            hps.model.upsample_initial_channel,
+            hps.model.upsample_kernel_sizes,
+            hps.model.spk_embed_dim,
+            hps.model.gin_channels,
+            hps.data.sample_rate,
+        ]
+
+        opt["epoch"] = f"{epoch}epoch"
+        opt["step"] = step
+        opt["sr"] = sr
+        opt["f0"] = int(pitch_guidance)
+        opt["version"] = version
+        opt["creation_date"] = datetime.datetime.now().isoformat()
+
+        hash_input = f"{str(ckpt)} {epoch} {step} {datetime.datetime.now().isoformat()}"
+        model_hash = hashlib.sha256(hash_input.encode()).hexdigest()
+        opt["model_hash"] = model_hash
+        opt["model_name"] = name
+        opt["author"] = model_author
+
+        torch.save(opt, os.path.join(model_dir_path, pth_file))
+
+        model = torch.load(model_dir, map_location=torch.device("cpu"))
+        torch.save(replace_keys_in_dict(replace_keys_in_dict(model, ".parametrizations.weight.original1", ".weight_v"), ".parametrizations.weight.original0", ".weight_g"), pth_file_old_version_path)
+
+        os.remove(model_dir)
+        os.rename(pth_file_old_version_path, model_dir)
+
+    except Exception as e:
+        logger.error(f"{translations['extract_model_error']}: {e}")
+
+
+def run(rank, n_gpus, experiment_dir, pretrainG, pretrainD, pitch_guidance, custom_total_epoch, custom_save_every_weights, config, device, model_author):
+    global global_step
+
+    if rank == 0:
+        writer = SummaryWriter(log_dir=experiment_dir)
+        writer_eval = SummaryWriter(log_dir=os.path.join(experiment_dir, "eval"))
+
+    dist.init_process_group(backend="gloo", init_method="env://", world_size=n_gpus, rank=rank)
+    torch.manual_seed(config.train.seed)
+
+    if torch.cuda.is_available(): torch.cuda.set_device(rank)
+
+    train_dataset = TextAudioLoaderMultiNSFsid(config.data)
+
+    train_sampler = DistributedBucketSampler(train_dataset, batch_size * n_gpus, [100, 200, 300, 400, 500, 600, 700, 800, 900], num_replicas=n_gpus, rank=rank, shuffle=True)
+
+    collate_fn = TextAudioCollateMultiNSFsid()
+
+    train_loader = DataLoader(train_dataset, num_workers=4, shuffle=False, pin_memory=True, collate_fn=collate_fn, batch_sampler=train_sampler, persistent_workers=True, prefetch_factor=8)
+
+    net_g = Synthesizer(config.data.filter_length // 2 + 1, config.train.segment_size // config.data.hop_length, **config.model, use_f0=pitch_guidance == True, is_half=config.train.fp16_run and device.type == "cuda", sr=sample_rate).to(device)
+
+    if torch.cuda.is_available(): net_g = net_g.cuda(rank)
+
+    if version == "v1": net_d = MultiPeriodDiscriminator(config.model.use_spectral_norm)
+    else: net_d = MultiPeriodDiscriminatorV2(config.model.use_spectral_norm)
+
+    if torch.cuda.is_available(): net_d = net_d.cuda(rank)
+
+    optim_g = torch.optim.AdamW(net_g.parameters(), config.train.learning_rate, betas=config.train.betas, eps=config.train.eps)
+    optim_d = torch.optim.AdamW(net_d.parameters(), config.train.learning_rate, betas=config.train.betas, eps=config.train.eps)
+
+    if torch.cuda.is_available():
+        net_g = DDP(net_g, device_ids=[rank])
+        net_d = DDP(net_d, device_ids=[rank])
+    else:
+        net_g = DDP(net_g)
+        net_d = DDP(net_d)
+
+    try:
+        logger.info(translations["start_training"])
+
+        _, _, _, epoch_str = load_checkpoint(latest_checkpoint_path(experiment_dir, "D_*.pth"), net_d, optim_d)
+        _, _, _, epoch_str = load_checkpoint(latest_checkpoint_path(experiment_dir, "G_*.pth"), net_g, optim_g)
+
+        epoch_str += 1
+        global_step = (epoch_str - 1) * len(train_loader)
+
+    except:
+        epoch_str = 1
+        global_step = 0
+
+        if pretrainG != "":
+            if rank == 0: logger.info(translations["import_pretrain"].format(dg="G", pretrain=pretrainG))
+
+            if hasattr(net_g, "module"): net_g.module.load_state_dict(torch.load(pretrainG, map_location="cpu")["model"])
+            else: net_g.load_state_dict(torch.load(pretrainG, map_location="cpu")["model"])
+        else: logger.warning(translations["not_using_pretrain"].format(dg="G"))
+
+        if pretrainD != "":
+            if rank == 0: logger.info(translations["import_pretrain"].format(dg="D", pretrain=pretrainD))
+
+            if hasattr(net_d, "module"): net_d.module.load_state_dict(torch.load(pretrainD, map_location="cpu")["model"])
+            else: net_d.load_state_dict(torch.load(pretrainD, map_location="cpu")["model"])
+        else: logger.warning(translations["not_using_pretrain"].format(dg="D"))
+
+    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=config.train.lr_decay, last_epoch=epoch_str - 2)
+    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=config.train.lr_decay, last_epoch=epoch_str - 2)
+
+    optim_d.step()
+    optim_g.step()
+
+    scaler = GradScaler(enabled=config.train.fp16_run)
+
+    cache = []
+
+    for info in train_loader:
+        phone, phone_lengths, pitch, pitchf, _, _, _, _, sid = info
+        reference = (
+            phone.to(device),
+            phone_lengths.to(device),
+            pitch.to(device) if pitch_guidance else None,
+            pitchf.to(device) if pitch_guidance else None,
+            sid.to(device),
+        )
+        break
+
+    for epoch in range(epoch_str, total_epoch + 1):
+        if rank == 0: train_and_evaluate(rank, epoch, config, [net_g, net_d], [optim_g, optim_d], scaler, [train_loader, None], [writer, writer_eval], cache, custom_save_every_weights, custom_total_epoch, device, reference, model_author)
+        else: train_and_evaluate(rank, epoch, config, [net_g, net_d], [optim_g, optim_d], scaler, [train_loader, None], None, cache, custom_save_every_weights, custom_total_epoch, device, reference, model_author)
+
+        scheduler_g.step()
+        scheduler_d.step()
+
+
+def train_and_evaluate(rank, epoch, hps, nets, optims, scaler, loaders, writers, cache, custom_save_every_weights, custom_total_epoch, device, reference, model_author):
+    global global_step, lowest_value, loss_disc, consecutive_increases_gen, consecutive_increases_disc
+
+    if epoch == 1:
+        lowest_value = {"step": 0, "value": float("inf"), "epoch": 0}
+        last_loss_gen_all = 0.0
+        consecutive_increases_gen = 0
+        consecutive_increases_disc = 0
+
+    net_g, net_d = nets
+    optim_g, optim_d = optims
+    train_loader = loaders[0] if loaders is not None else None
+
+    if writers is not None: writer = writers[0]
+
+    train_loader.batch_sampler.set_epoch(epoch)
+
+    net_g.train()
+    net_d.train()
+
+    if device.type == "cuda" and cache_data_in_gpu:
+        data_iterator = cache
+
+        if cache == []:
+            for batch_idx, info in enumerate(train_loader):
+                (
+                    phone,
+                    phone_lengths,
+                    pitch,
+                    pitchf,
+                    spec,
+                    spec_lengths,
+                    wave,
+                    wave_lengths,
+                    sid,
+                ) = info
+                cache.append(
+                    (batch_idx, (
+                        phone.cuda(rank, non_blocking=True),
+                        phone_lengths.cuda(rank, non_blocking=True),
+                        (pitch.cuda(rank, non_blocking=True) if pitch_guidance else None),
+                        (pitchf.cuda(rank, non_blocking=True) if pitch_guidance else None),
+                        spec.cuda(rank, non_blocking=True),
+                        spec_lengths.cuda(rank, non_blocking=True),
+                        wave.cuda(rank, non_blocking=True),
+                        wave_lengths.cuda(rank, non_blocking=True),
+                        sid.cuda(rank, non_blocking=True),
+                        ),
+                    ))
+        else: shuffle(cache)
+    else: data_iterator = enumerate(train_loader)
+
+    epoch_recorder = EpochRecorder()
+
+    with tqdm(total=len(train_loader), leave=False) as pbar:
+        for batch_idx, info in data_iterator:
+            (
+                phone,
+                phone_lengths,
+                pitch,
+                pitchf,
+                spec,
+                spec_lengths,
+                wave,
+                wave_lengths,
+                sid,
+            ) = info
+            if device.type == "cuda" and not cache_data_in_gpu:
+                phone = phone.cuda(rank, non_blocking=True)
+                phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
+                pitch = pitch.cuda(rank, non_blocking=True) if pitch_guidance else None
+                pitchf = (pitchf.cuda(rank, non_blocking=True) if pitch_guidance else None)
+                sid = sid.cuda(rank, non_blocking=True)
+                spec = spec.cuda(rank, non_blocking=True)
+                spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
+                wave = wave.cuda(rank, non_blocking=True)
+                wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
+            else:
+                phone = phone.to(device)
+                phone_lengths = phone_lengths.to(device)
+                pitch = pitch.to(device) if pitch_guidance else None
+                pitchf = pitchf.to(device) if pitch_guidance else None
+                sid = sid.to(device)
+                spec = spec.to(device)
+                spec_lengths = spec_lengths.to(device)
+                wave = wave.to(device)
+                wave_lengths = wave_lengths.to(device)
+
+            use_amp = config.train.fp16_run and device.type == "cuda"
+
+            with autocast(enabled=use_amp):
+                (
+                    y_hat,
+                    ids_slice,
+                    x_mask,
+                    z_mask,
+                    (z, z_p, m_p, logs_p, m_q, logs_q),
+                ) = net_g(phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid)
+                mel = spec_to_mel_torch(
+                    spec,
+                    config.data.filter_length,
+                    config.data.n_mel_channels,
+                    config.data.sample_rate,
+                    config.data.mel_fmin,
+                    config.data.mel_fmax,
+                )
+                y_mel = slice_segments(mel, ids_slice, config.train.segment_size // config.data.hop_length, dim=3)
+                with autocast(enabled=False):
+                    y_hat_mel = mel_spectrogram_torch(
+                        y_hat.float().squeeze(1),
+                        config.data.filter_length,
+                        config.data.n_mel_channels,
+                        config.data.sample_rate,
+                        config.data.hop_length,
+                        config.data.win_length,
+                        config.data.mel_fmin,
+                        config.data.mel_fmax,
+                    )
+                if use_amp: y_hat_mel = y_hat_mel.half()
+
+                wave = slice_segments(wave, ids_slice * config.data.hop_length, config.train.segment_size, dim=3)
+                y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
+
+                with autocast(enabled=False):
+                    loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
+
+            optim_d.zero_grad()
+            scaler.scale(loss_disc).backward()
+            scaler.unscale_(optim_d)
+            grad_norm_d = clip_grad_value(net_d.parameters(), None)
+            scaler.step(optim_d)
+
+            with autocast(enabled=use_amp):
+                y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
+                with autocast(enabled=False):
+                    loss_mel = F.l1_loss(y_mel, y_hat_mel) * config.train.c_mel
+                    loss_kl = (kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * config.train.c_kl)
+                    loss_fm = feature_loss(fmap_r, fmap_g)
+                    loss_gen, losses_gen = generator_loss(y_d_hat_g)
+                    loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
+
+                    if loss_gen_all < lowest_value["value"]:
+                        lowest_value["value"] = loss_gen_all
+                        lowest_value["step"] = global_step
+                        lowest_value["epoch"] = epoch
+                       
+                        if epoch > lowest_value["epoch"]: logger.warning(translations["training_warning"])
+
+            optim_g.zero_grad()
+            scaler.scale(loss_gen_all).backward()
+            scaler.unscale_(optim_g)
+            grad_norm_g = clip_grad_value(net_g.parameters(), None)
+            scaler.step(optim_g)
+            scaler.update()
+
+            if rank == 0:
+                if global_step % config.train.log_interval == 0:
+                    lr = optim_g.param_groups[0]["lr"]
+
+                    if loss_mel > 75: loss_mel = 75
+                    if loss_kl > 9: loss_kl = 9
+
+                    scalar_dict = {
+                        "loss/g/total": loss_gen_all,
+                        "loss/d/total": loss_disc,
+                        "learning_rate": lr,
+                        "grad_norm_d": grad_norm_d,
+                        "grad_norm_g": grad_norm_g,
+                    }
+                    scalar_dict.update(
+                        {
+                            "loss/g/fm": loss_fm,
+                            "loss/g/mel": loss_mel,
+                            "loss/g/kl": loss_kl,
+                        }
+                    )
+                    scalar_dict.update(
+                        {f"loss/g/{i}": v for i, v in enumerate(losses_gen)}
+                    )
+                    scalar_dict.update(
+                        {f"loss/d_r/{i}": v for i, v in enumerate(losses_disc_r)}
+                    )
+                    scalar_dict.update(
+                        {f"loss/d_g/{i}": v for i, v in enumerate(losses_disc_g)}
+                    )
+                    image_dict = {
+                        "slice/mel_org": plot_spectrogram_to_numpy(
+                            y_mel[0].data.cpu().numpy()
+                        ),
+                        "slice/mel_gen": plot_spectrogram_to_numpy(
+                            y_hat_mel[0].data.cpu().numpy()
+                        ),
+                        "all/mel": plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
+                    }
+
+                    with torch.no_grad():
+                        if hasattr(net_g, "module"): o, *_ = net_g.module.infer(*reference)
+                        else: o, *_ = net_g.infer(*reference)
+
+
+                    audio_dict = {f"gen/audio_{global_step:07d}": o[0, :, :]}
+
+                    summarize(
+                        writer=writer,
+                        global_step=global_step,
+                        images=image_dict,
+                        scalars=scalar_dict,
+                        audios=audio_dict,
+                        audio_sample_rate=config.data.sample_rate,
+                    )
+
+            global_step += 1
+            pbar.update(1)
+
+    def check_overtraining(smoothed_loss_history, threshold, epsilon=0.004):
+        if len(smoothed_loss_history) < threshold + 1: return False
+
+        for i in range(-threshold, -1):
+            if smoothed_loss_history[i + 1] > smoothed_loss_history[i]: return True
+            if abs(smoothed_loss_history[i + 1] - smoothed_loss_history[i]) >= epsilon: return False
+
+        return True
+
+    def update_exponential_moving_average(smoothed_loss_history, new_value, smoothing=0.987):
+        smoothed_value = new_value if not smoothed_loss_history else (smoothing * smoothed_loss_history[-1] + (1 - smoothing) * new_value)      
+        smoothed_loss_history.append(smoothed_value)
+        return smoothed_value
+
+    def save_to_json(file_path, loss_disc_history, smoothed_loss_disc_history, loss_gen_history, smoothed_loss_gen_history):
+        data = {
+            "loss_disc_history": loss_disc_history,
+            "smoothed_loss_disc_history": smoothed_loss_disc_history,
+            "loss_gen_history": loss_gen_history,
+            "smoothed_loss_gen_history": smoothed_loss_gen_history,
+        }
+
+        with open(file_path, "w") as f:
+            json.dump(data, f)
+    
+    model_add = []
+    model_del = []
+    done = False
+    
+    if rank == 0:
+        if epoch % save_every_epoch == False:
+            checkpoint_suffix = f"{2333333 if save_only_latest else global_step}.pth"
+
+            save_checkpoint(net_g, optim_g, config.train.learning_rate, epoch, os.path.join(experiment_dir, "G_" + checkpoint_suffix))
+            save_checkpoint(net_d, optim_d, config.train.learning_rate, epoch, os.path.join(experiment_dir, "D_" + checkpoint_suffix))
+
+            if custom_save_every_weights: model_add.append(os.path.join("assets", "weights", f"{model_name}_{epoch}e_{global_step}s.pth"))
+
+        if overtraining_detector and epoch > 1:
+            current_loss_disc = float(loss_disc)
+            loss_disc_history.append(current_loss_disc)
+
+            smoothed_value_disc = update_exponential_moving_average(smoothed_loss_disc_history, current_loss_disc)
+            is_overtraining_disc = check_overtraining(smoothed_loss_disc_history, overtraining_threshold * 2)
+
+            if is_overtraining_disc: consecutive_increases_disc += 1
+            else: consecutive_increases_disc = 0
+
+            current_loss_gen = float(lowest_value["value"])
+            loss_gen_history.append(current_loss_gen)
+
+            smoothed_value_gen = update_exponential_moving_average(smoothed_loss_gen_history, current_loss_gen)
+            is_overtraining_gen = check_overtraining(smoothed_loss_gen_history, overtraining_threshold, 0.01)
+
+            if is_overtraining_gen: consecutive_increases_gen += 1
+            else: consecutive_increases_gen = 0
+
+            if epoch % save_every_epoch == 0: save_to_json(training_file_path, loss_disc_history, smoothed_loss_disc_history, loss_gen_history, smoothed_loss_gen_history)
+
+            if (is_overtraining_gen and consecutive_increases_gen == overtraining_threshold or is_overtraining_disc and consecutive_increases_disc == (overtraining_threshold * 2)):
+                logger.info(translations["overtraining_find"].format(epoch=epoch, smoothed_value_gen=f"{smoothed_value_gen:.3f}", smoothed_value_disc=f"{smoothed_value_disc:.3f}"))
+                done = True
+            else:
+                logger.info(translations["best_epoch"].format(epoch=epoch, smoothed_value_gen=f"{smoothed_value_gen:.3f}", smoothed_value_disc=f"{smoothed_value_disc:.3f}"))
+
+                old_model_files = glob.glob(os.path.join("assets", "weights", f"{model_name}_*e_*s_best_epoch.pth"))
+
+                for file in old_model_files:
+                    model_del.append(file)
+
+                model_add.append(os.path.join("assets", "weights", f"{model_name}_{epoch}e_{global_step}s_best_epoch.pth"))
+        
+        if epoch >= custom_total_epoch:
+            lowest_value_rounded = float(lowest_value["value"])
+            lowest_value_rounded = round(lowest_value_rounded, 3)
+
+            logger.info(translations["success_training"].format(epoch=epoch, global_step=global_step, loss_gen_all=round(loss_gen_all.item(), 3)))
+            logger.info(translations["training_info"].format(lowest_value_rounded=lowest_value_rounded, lowest_value_epoch=lowest_value['epoch'], lowest_value_step=lowest_value['step']))
+
+            pid_file_path = os.path.join(experiment_dir, "config.json")
+
+            with open(pid_file_path, "r") as pid_file:
+                pid_data = json.load(pid_file)
+            with open(pid_file_path, "w") as pid_file:
+                pid_data.pop("process_pids", None)
+                json.dump(pid_data, pid_file, indent=4)
+
+            model_add.append(os.path.join("assets", "weights", f"{model_name}_{epoch}e_{global_step}s.pth"))
+            done = True
+            
+        if model_add:
+            ckpt = (net_g.module.state_dict() if hasattr(net_g, "module") else net_g.state_dict())
+
+            for m in model_add:
+                if not os.path.exists(m): extract_model(ckpt=ckpt, sr=sample_rate, pitch_guidance=pitch_guidance == True, name=model_name, model_dir=m, epoch=epoch, step=global_step, version=version, hps=hps, model_author=model_author)
+
+        for m in model_del:
+            os.remove(m)
+
+        lowest_value_rounded = float(lowest_value["value"])
+        lowest_value_rounded = round(lowest_value_rounded, 3)
+
+        if epoch > 1 and overtraining_detector:
+            remaining_epochs_gen = overtraining_threshold - consecutive_increases_gen
+            remaining_epochs_disc = (overtraining_threshold * 2) - consecutive_increases_disc
+
+            logger.info(translations["model_training_info"].format(model_name=model_name, epoch=epoch, global_step=global_step, epoch_recorder=epoch_recorder.record(), lowest_value_rounded=lowest_value_rounded, lowest_value_epoch=lowest_value['epoch'], lowest_value_step=lowest_value['step'], remaining_epochs_gen=remaining_epochs_gen, remaining_epochs_disc=remaining_epochs_disc, smoothed_value_gen=f"{smoothed_value_gen:.3f}", smoothed_value_disc=f"{smoothed_value_disc:.3f}"))
+        elif epoch > 1 and overtraining_detector == False: logger.info(translations["model_training_info_2"].format(model_name=model_name, epoch=epoch, global_step=global_step, epoch_recorder=epoch_recorder.record(), lowest_value_rounded=lowest_value_rounded, lowest_value_epoch=lowest_value['epoch'], lowest_value_step=lowest_value['step']))
+        else: logger.info(translations["model_training_info_3"].format(model_name=model_name, epoch=epoch, global_step=global_step, epoch_recorder=epoch_recorder.record()))
+
+        last_loss_gen_all = loss_gen_all
+
+        if done: os._exit(2333333)
+
+if __name__ == "__main__":
+    torch.multiprocessing.set_start_method("spawn")
+
+    try:
+        main()
+    except Exception as e:
+        logger.error(f"{translations['training_error']} {e}")

main/library/algorithm/commons.py

@@ -0,0 +1,100 @@
+import math
+import torch
+
+from typing import List, Optional
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1: m.weight.data.normal_(mean, std)
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+def convert_pad_shape(pad_shape):
+    l = pad_shape[::-1]
+    pad_shape = [item for sublist in l for item in sublist]
+    return pad_shape
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+    kl = (logs_q - logs_p) - 0.5
+    kl += (0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q))
+    return kl
+
+def slice_segments(x: torch.Tensor, ids_str: torch.Tensor, segment_size: int = 4, dim: int = 2):
+    if dim == 2: ret = torch.zeros_like(x[:, :segment_size])
+    elif dim == 3: ret = torch.zeros_like(x[:, :, :segment_size])
+
+    for i in range(x.size(0)):
+        idx_str = ids_str[i].item()
+        idx_end = idx_str + segment_size
+
+        if dim == 2: ret[i] = x[i, idx_str:idx_end]
+        else: ret[i] = x[i, :, idx_str:idx_end]
+
+    return ret
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+    b, d, t = x.size()
+
+    if x_lengths is None: x_lengths = t
+
+    ids_str_max = x_lengths - segment_size + 1
+    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+    ret = slice_segments(x, ids_str, segment_size, dim=3)
+    return ret, ids_str
+
+def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
+    position = torch.arange(length, dtype=torch.float)
+    num_timescales = channels // 2
+    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (num_timescales - 1)
+    inv_timescales = min_timescale * torch.exp(torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
+    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+    signal = torch.nn.functional.pad(signal, [0, 0, 0, channels % 2])
+    signal = signal.view(1, channels, length)
+    return signal
+
+
+def subsequent_mask(length):
+    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+    return mask
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+def convert_pad_shape(pad_shape: List[List[int]]) -> List[int]:
+    return torch.tensor(pad_shape).flip(0).reshape(-1).int().tolist()
+
+
+def sequence_mask(length: torch.Tensor, max_length: Optional[int] = None):
+    if max_length is None: max_length = length.max()
+    
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def clip_grad_value(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor): parameters = [parameters]
+
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+
+    if clip_value is not None: clip_value = float(clip_value)
+
+    total_norm = 0
+
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+        if clip_value is not None: p.grad.data.clamp_(min=-clip_value, max=clip_value)
+
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm

main/library/algorithm/modules.py

@@ -0,0 +1,80 @@
+import os
+import sys
+import torch
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from .commons import fused_add_tanh_sigmoid_multiply
+
+class WaveNet(torch.nn.Module):
+    def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
+        super(WaveNet, self).__init__()
+        assert kernel_size % 2 == 1
+
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = p_dropout
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = torch.nn.Dropout(p_dropout)
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(gin_channels, 2 * hidden_channels * n_layers, 1)
+            self.cond_layer = torch.nn.utils.parametrizations.weight_norm(cond_layer, name="weight")
+
+        dilations = [dilation_rate**i for i in range(n_layers)]
+        paddings = [(kernel_size * d - d) // 2 for d in dilations]
+
+        for i in range(n_layers):
+            in_layer = torch.nn.Conv1d(hidden_channels, 2 * hidden_channels, kernel_size, dilation=dilations[i], padding=paddings[i])
+            in_layer = torch.nn.utils.parametrizations.weight_norm(in_layer, name="weight")
+
+            self.in_layers.append(in_layer)
+
+            res_skip_channels = (hidden_channels if i == n_layers - 1 else 2 * hidden_channels)
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.parametrizations.weight_norm(res_skip_layer, name="weight")
+            
+            self.res_skip_layers.append(res_skip_layer)
+
+    def forward(self, x, x_mask, g=None, **kwargs):
+        output = torch.zeros_like(x)
+        n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+        if g is not None: g = self.cond_layer(g)
+
+        for i in range(self.n_layers):
+            x_in = self.in_layers[i](x)
+
+            if g is not None:
+                cond_offset = i * 2 * self.hidden_channels
+                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+            else: g_l = torch.zeros_like(x_in)
+
+            acts = fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+            acts = self.drop(acts)
+
+            res_skip_acts = self.res_skip_layers[i](acts)
+
+            if i < self.n_layers - 1:
+                res_acts = res_skip_acts[:, : self.hidden_channels, :]
+                x = (x + res_acts) * x_mask
+                output = output + res_skip_acts[:, self.hidden_channels :, :]
+            else: output = output + res_skip_acts
+
+        return output * x_mask
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0: torch.nn.utils.remove_weight_norm(self.cond_layer)
+
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)

main/library/algorithm/residuals.py

@@ -0,0 +1,170 @@
+import os
+import sys
+import torch
+
+from typing import Optional
+
+from torch.nn.utils import remove_weight_norm
+from torch.nn.utils.parametrizations import weight_norm
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from .modules import WaveNet
+from .commons import get_padding, init_weights
+
+
+LRELU_SLOPE = 0.1
+
+def create_conv1d_layer(channels, kernel_size, dilation):
+    return weight_norm(torch.nn.Conv1d(channels, channels, kernel_size, 1, dilation=dilation, padding=get_padding(kernel_size, dilation)))
+
+def apply_mask(tensor, mask):
+    return tensor * mask if mask is not None else tensor
+
+class ResBlockBase(torch.nn.Module):
+    def __init__(self, channels, kernel_size, dilations):
+        super(ResBlockBase, self).__init__()
+        
+        self.convs1 = torch.nn.ModuleList([create_conv1d_layer(channels, kernel_size, d) for d in dilations])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = torch.nn.ModuleList([create_conv1d_layer(channels, kernel_size, 1) for _ in dilations])
+        self.convs2.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = torch.nn.functional.leaky_relu(x, LRELU_SLOPE)
+            xt = apply_mask(xt, x_mask)
+            xt = torch.nn.functional.leaky_relu(c1(xt), LRELU_SLOPE)
+            xt = apply_mask(xt, x_mask)
+            xt = c2(xt)
+            x = xt + x
+        return apply_mask(x, x_mask)
+
+    def remove_weight_norm(self):
+        for conv in self.convs1 + self.convs2:
+            remove_weight_norm(conv)
+
+class ResBlock1(ResBlockBase):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__(channels, kernel_size, dilation)
+
+class ResBlock2(ResBlockBase):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__(channels, kernel_size, dilation)
+
+class Log(torch.nn.Module):
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+class Flip(torch.nn.Module):
+    def forward(self, x, *args, reverse=False, **kwargs):
+        x = torch.flip(x, [1])
+        if not reverse:
+            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+            return x, logdet
+        else: return x
+
+class ElementwiseAffine(torch.nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.channels = channels
+        self.m = torch.nn.Parameter(torch.zeros(channels, 1))
+        self.logs = torch.nn.Parameter(torch.zeros(channels, 1))
+
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class ResidualCouplingBlock(torch.nn.Module):
+    def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, n_flows=4, gin_channels=0):
+        super(ResidualCouplingBlock, self).__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = torch.nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
+            self.flows.append(Flip())
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor, g: Optional[torch.Tensor] = None, reverse = False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow.forward(x, x_mask, g=g, reverse=reverse)
+
+        return x
+
+    def remove_weight_norm(self):
+        for i in range(self.n_flows):
+            self.flows[i * 2].remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for i in range(self.n_flows):
+            for hook in self.flows[i * 2]._forward_pre_hooks.values():
+                if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(self.flows[i * 2])
+        return self
+
+
+class ResidualCouplingLayer(torch.nn.Module):
+    def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=0, gin_channels=0, mean_only=False):
+        assert channels % 2 == 0, "Channels/2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = torch.nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = WaveNet(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
+        self.post = torch.nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+
+        if not self.mean_only: m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()

main/library/algorithm/separator.py

@@ -0,0 +1,420 @@
+import os
+import sys
+import time
+import json
+import yaml
+import torch
+import codecs
+import hashlib
+import logging
+import platform
+import warnings
+import requests
+import subprocess
+
+import onnxruntime as ort
+
+from tqdm import tqdm
+from importlib import metadata, import_module
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+translations = Config().translations
+
+class Separator:
+    def __init__(self, log_level=logging.INFO, log_formatter=None, model_file_dir="assets/model/uvr5", output_dir=None, output_format="wav", output_bitrate=None, normalization_threshold=0.9, output_single_stem=None, invert_using_spec=False, sample_rate=44100, mdx_params={"hop_length": 1024, "segment_size": 256, "overlap": 0.25, "batch_size": 1, "enable_denoise": False}, demucs_params={"segment_size": "Default", "shifts": 2, "overlap": 0.25, "segments_enabled": True}):
+        self.logger = logging.getLogger(__name__)
+        self.logger.setLevel(log_level)
+        self.log_level = log_level
+        self.log_formatter = log_formatter
+        self.log_handler = logging.StreamHandler()
+
+        if self.log_formatter is None: self.log_formatter = logging.Formatter("%(asctime)s - %(levelname)s - %(module)s - %(message)s")
+
+        self.log_handler.setFormatter(self.log_formatter)
+
+        if not self.logger.hasHandlers(): self.logger.addHandler(self.log_handler)
+        if log_level > logging.DEBUG: warnings.filterwarnings("ignore")
+
+        self.logger.info(translations["separator_info"].format(output_dir=output_dir, output_format=output_format))
+
+        self.model_file_dir = model_file_dir
+
+        if output_dir is None:
+            output_dir = os.getcwd()
+            self.logger.info(translations["output_dir_is_none"])
+
+        self.output_dir = output_dir
+
+        os.makedirs(self.model_file_dir, exist_ok=True)
+        os.makedirs(self.output_dir, exist_ok=True)
+
+        self.output_format = output_format
+        self.output_bitrate = output_bitrate
+
+        if self.output_format is None: self.output_format = "wav"
+
+        self.normalization_threshold = normalization_threshold
+
+        if normalization_threshold <= 0 or normalization_threshold > 1: raise ValueError(translations[">0or=1"])
+
+        self.output_single_stem = output_single_stem
+
+        if output_single_stem is not None: self.logger.debug(translations["output_single"].format(output_single_stem=output_single_stem))
+
+        self.invert_using_spec = invert_using_spec
+        if self.invert_using_spec: self.logger.debug(translations["step2"])
+
+        try:
+            self.sample_rate = int(sample_rate)
+            
+            if self.sample_rate <= 0: raise ValueError(translations["other_than_zero"].format(sample_rate=self.sample_rate))
+            if self.sample_rate > 12800000: raise ValueError(translations["too_high"].format(sample_rate=self.sample_rate))
+        except ValueError: 
+            raise ValueError(translations["sr_not_valid"])
+
+        self.arch_specific_params = {"MDX": mdx_params, "Demucs": demucs_params}
+        self.torch_device = None
+        self.torch_device_cpu = None
+        self.torch_device_mps = None
+        self.onnx_execution_provider = None
+        self.model_instance = None
+        self.model_is_uvr_vip = False
+        self.model_friendly_name = None
+
+        self.setup_accelerated_inferencing_device()
+
+    def setup_accelerated_inferencing_device(self):
+        system_info = self.get_system_info()
+        self.check_ffmpeg_installed()
+        self.log_onnxruntime_packages()
+        self.setup_torch_device(system_info)
+
+    def get_system_info(self):
+        os_name = platform.system()
+        os_version = platform.version()
+
+        self.logger.info(f"{translations['os']}: {os_name} {os_version}")
+
+        system_info = platform.uname()
+        self.logger.info(translations["platform_info"].format(system_info=system_info, node=system_info.node, release=system_info.release, machine=system_info.machine, processor=system_info.processor))
+
+        python_version = platform.python_version()
+        self.logger.info(f"{translations['name_ver'].format(name='python')}: {python_version}")
+
+        pytorch_version = torch.__version__
+        self.logger.info(f"{translations['name_ver'].format(name='pytorch')}: {pytorch_version}")
+
+        return system_info
+
+    def check_ffmpeg_installed(self):
+        try:
+            ffmpeg_version_output = subprocess.check_output(["ffmpeg", "-version"], text=True)
+            first_line = ffmpeg_version_output.splitlines()[0]
+            self.logger.info(f"{translations['install_ffmpeg']}: {first_line}")
+        except FileNotFoundError:
+            self.logger.error(translations["none_ffmpeg"])
+            if "PYTEST_CURRENT_TEST" not in os.environ: raise
+
+    def log_onnxruntime_packages(self):
+        onnxruntime_gpu_package = self.get_package_distribution("onnxruntime-gpu")
+        onnxruntime_cpu_package = self.get_package_distribution("onnxruntime")
+
+        if onnxruntime_gpu_package is not None: self.logger.info(f"{translations['install_onnx'].format(pu='GPU')}: {onnxruntime_gpu_package.version}")
+        if onnxruntime_cpu_package is not None: self.logger.info(f"{translations['install_onnx'].format(pu='CPU')}: {onnxruntime_cpu_package.version}")
+
+    def setup_torch_device(self, system_info):
+        hardware_acceleration_enabled = False
+        ort_providers = ort.get_available_providers()
+
+        self.torch_device_cpu = torch.device("cpu")
+
+        if torch.cuda.is_available():
+            self.configure_cuda(ort_providers)
+            hardware_acceleration_enabled = True
+        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available() and system_info.processor == "arm":
+            self.configure_mps(ort_providers)
+            hardware_acceleration_enabled = True
+
+        if not hardware_acceleration_enabled:
+            self.logger.info(translations["running_in_cpu"])
+            self.torch_device = self.torch_device_cpu
+            self.onnx_execution_provider = ["CPUExecutionProvider"]
+
+    def configure_cuda(self, ort_providers):
+        self.logger.info(translations["running_in_cuda"])
+        self.torch_device = torch.device("cuda")
+
+        if "CUDAExecutionProvider" in ort_providers:
+            self.logger.info(translations["onnx_have"].format(have='CUDAExecutionProvider'))
+            self.onnx_execution_provider = ["CUDAExecutionProvider"]
+        else: self.logger.warning(translations["onnx_not_have"].format(have='CUDAExecutionProvider'))
+
+    def configure_mps(self, ort_providers):
+        self.logger.info("Cài đặt thiết bị Torch thành MPS")
+        self.torch_device_mps = torch.device("mps")
+        self.torch_device = self.torch_device_mps
+
+        if "CoreMLExecutionProvider" in ort_providers:
+            self.logger.info(translations["onnx_have"].format(have='CoreMLExecutionProvider'))
+            self.onnx_execution_provider = ["CoreMLExecutionProvider"]
+        else: self.logger.warning(translations["onnx_not_have"].format(have='CoreMLExecutionProvider'))
+
+    def get_package_distribution(self, package_name):
+        try:
+            return metadata.distribution(package_name)
+        except metadata.PackageNotFoundError:
+            self.logger.debug(translations["python_not_install"].format(package_name=package_name))
+            return None
+
+    def get_model_hash(self, model_path):
+        self.logger.debug(translations["hash"].format(model_path=model_path))
+
+        try:
+            with open(model_path, "rb") as f:
+                f.seek(-10000 * 1024, 2)
+                return hashlib.md5(f.read()).hexdigest()
+        except IOError as e:
+            self.logger.error(translations["ioerror"].format(e=e))
+
+            return hashlib.md5(open(model_path, "rb").read()).hexdigest()
+
+    def download_file_if_not_exists(self, url, output_path):
+        if os.path.isfile(output_path):
+            self.logger.debug(translations["cancel_download"].format(output_path=output_path))
+            return
+
+        self.logger.debug(translations["download_model"].format(url=url, output_path=output_path))
+        response = requests.get(url, stream=True, timeout=300)
+
+        if response.status_code == 200:
+            total_size_in_bytes = int(response.headers.get("content-length", 0))
+            progress_bar = tqdm(total=total_size_in_bytes)
+
+            with open(output_path, "wb") as f:
+                for chunk in response.iter_content(chunk_size=8192):
+                    progress_bar.update(len(chunk))
+                    f.write(chunk)
+
+            progress_bar.close()
+        else: raise RuntimeError(translations["download_error"].format(url=url, status_code=response.status_code))
+
+    def print_uvr_vip_message(self):
+        if self.model_is_uvr_vip:
+            self.logger.warning(translations["vip_model"].format(model_friendly_name=self.model_friendly_name))
+            self.logger.warning(translations["vip_print"])
+
+    def list_supported_model_files(self):
+        download_checks_path = os.path.join(self.model_file_dir, "download_checks.json")
+
+        model_downloads_list = json.load(open(download_checks_path, encoding="utf-8"))
+        self.logger.debug(translations["load_download_json"])
+
+        filtered_demucs_v4 = {key: value for key, value in model_downloads_list["demucs_download_list"].items() if key.startswith("Demucs v4")}
+
+        model_files_grouped_by_type = {"MDX": {**model_downloads_list["mdx_download_list"], **model_downloads_list["mdx_download_vip_list"]}, "Demucs": filtered_demucs_v4}
+        return model_files_grouped_by_type
+    
+    def download_model_files(self, model_filename):
+        model_path = os.path.join(self.model_file_dir, f"{model_filename}")
+
+        supported_model_files_grouped = self.list_supported_model_files()
+        public_model_repo_url_prefix = codecs.decode("uggcf://tvguho.pbz/GEiyie/zbqry_ercb/eryrnfrf/qbjaybnq/nyy_choyvp_hie_zbqryf", "rot13")
+        vip_model_repo_url_prefix = codecs.decode("uggcf://tvguho.pbz/Nawbx0109/nv_zntvp/eryrnfrf/qbjaybnq/i5", "rot13")
+
+        audio_separator_models_repo_url_prefix = codecs.decode("uggcf://tvguho.pbz/abznqxnenbxr/clguba-nhqvb-frcnengbe/eryrnfrf/qbjaybnq/zbqry-pbasvtf", "rot13")
+
+        yaml_config_filename = None
+
+        self.logger.debug(translations["search_model"].format(model_filename=model_filename))
+
+        for model_type, model_list in supported_model_files_grouped.items():
+            for model_friendly_name, model_download_list in model_list.items():
+                self.model_is_uvr_vip = "VIP" in model_friendly_name
+                model_repo_url_prefix = vip_model_repo_url_prefix if self.model_is_uvr_vip else public_model_repo_url_prefix
+
+                if isinstance(model_download_list, str) and model_download_list == model_filename:
+                    self.logger.debug(translations["single_model"].format(model_friendly_name=model_friendly_name))
+                    self.model_friendly_name = model_friendly_name
+
+                    try:
+                        self.download_file_if_not_exists(f"{model_repo_url_prefix}/{model_filename}", model_path)
+                    except RuntimeError:
+                        self.logger.debug(translations["not_found_model"])
+                        self.download_file_if_not_exists(f"{audio_separator_models_repo_url_prefix}/{model_filename}", model_path)
+
+                    self.print_uvr_vip_message()
+
+                    self.logger.debug(translations["single_model_path"].format(model_path=model_path))
+
+                    return model_filename, model_type, model_friendly_name, model_path, yaml_config_filename
+                elif isinstance(model_download_list, dict):
+                    this_model_matches_input_filename = False
+
+                    for file_name, file_url in model_download_list.items():
+                        if file_name == model_filename or file_url == model_filename:
+                            self.logger.debug(translations["find_model"].format(model_filename=model_filename, model_friendly_name=model_friendly_name))
+                            this_model_matches_input_filename = True
+
+                    if this_model_matches_input_filename:
+                        self.logger.debug(translations["find_models"].format(model_friendly_name=model_friendly_name))
+                        self.model_friendly_name = model_friendly_name
+                        self.print_uvr_vip_message()
+
+                        for config_key, config_value in model_download_list.items():
+                            self.logger.debug(f"{translations['find_path']}: {config_key} -> {config_value}")
+
+                            if config_value.startswith("http"): self.download_file_if_not_exists(config_value, os.path.join(self.model_file_dir, config_key))
+                            elif config_key.endswith(".ckpt"):
+                                try:
+                                    download_url = f"{model_repo_url_prefix}/{config_key}"
+                                    self.download_file_if_not_exists(download_url, os.path.join(self.model_file_dir, config_key))
+                                except RuntimeError:
+                                    self.logger.debug(translations["not_found_model_warehouse"])
+                                    download_url = f"{audio_separator_models_repo_url_prefix}/{config_key}"
+                                    self.download_file_if_not_exists(download_url, os.path.join(self.model_file_dir, config_key))
+
+                                if model_filename.endswith(".yaml"):
+                                    self.logger.warning(translations["yaml_warning"].format(model_filename=model_filename))
+                                    self.logger.warning(translations["yaml_warning_2"].format(config_key=config_key))
+                                    self.logger.warning(translations["yaml_warning_3"])
+
+                                    model_filename = config_key
+                                    model_path = os.path.join(self.model_file_dir, f"{model_filename}")
+
+                                yaml_config_filename = config_value
+                                yaml_config_filepath = os.path.join(self.model_file_dir, yaml_config_filename)
+
+                                try:
+                                    url = codecs.decode("uggcf://enj.tvguhohfrepbagrag.pbz/GEiyie/nccyvpngvba_qngn/znva/zqk_zbqry_qngn/zqk_p_pbasvtf", "rot13")
+                                    yaml_config_url = f"{url}/{yaml_config_filename}"
+                                    self.download_file_if_not_exists(f"{yaml_config_url}", yaml_config_filepath)
+                                except RuntimeError:
+                                    self.logger.debug(translations["yaml_debug"])
+                                    yaml_config_url = f"{audio_separator_models_repo_url_prefix}/{yaml_config_filename}"
+                                    self.download_file_if_not_exists(f"{yaml_config_url}", yaml_config_filepath)
+                            else:
+                                download_url = f"{model_repo_url_prefix}/{config_value}"
+                                self.download_file_if_not_exists(download_url, os.path.join(self.model_file_dir, config_value))
+
+                        self.logger.debug(translations["download_model_friendly"].format(model_friendly_name=model_friendly_name, model_path=model_path))
+
+                        return model_filename, model_type, model_friendly_name, model_path, yaml_config_filename
+
+        raise ValueError(translations["not_found_model_2"].format(model_filename=model_filename))
+
+    def load_model_data_from_yaml(self, yaml_config_filename):
+        model_data_yaml_filepath = os.path.join(self.model_file_dir, yaml_config_filename) if not os.path.exists(yaml_config_filename) else yaml_config_filename
+
+        self.logger.debug(translations["load_yaml"].format(model_data_yaml_filepath=model_data_yaml_filepath))
+
+        model_data = yaml.load(open(model_data_yaml_filepath, encoding="utf-8"), Loader=yaml.FullLoader)
+        self.logger.debug(translations["load_yaml_2"].format(model_data=model_data))
+
+        if "roformer" in model_data_yaml_filepath: model_data["is_roformer"] = True
+
+        return model_data
+
+    def load_model_data_using_hash(self, model_path):
+        mdx_model_data_url = codecs.decode("uggcf://enj.tvguhohfrepbagrag.pbz/GEiyie/nccyvpngvba_qngn/znva/zqk_zbqry_qngn/zbqry_qngn_arj.wfba", "rot13")
+
+        self.logger.debug(translations["hash_md5"])
+        model_hash = self.get_model_hash(model_path)
+        self.logger.debug(translations["model_hash"].format(model_path=model_path, model_hash=model_hash))
+
+        mdx_model_data_path = os.path.join(self.model_file_dir, "mdx_model_data.json")
+        self.logger.debug(translations["mdx_data"].format(mdx_model_data_path=mdx_model_data_path))
+        self.download_file_if_not_exists(mdx_model_data_url, mdx_model_data_path)
+
+        self.logger.debug(translations["load_mdx"])
+        mdx_model_data_object = json.load(open(mdx_model_data_path, encoding="utf-8"))
+
+        if model_hash in mdx_model_data_object: model_data = mdx_model_data_object[model_hash]
+        else: raise ValueError(translations["model_not_support"].format(model_hash=model_hash))
+
+        self.logger.debug(translations["uvr_json"].format(model_hash=model_hash, model_data=model_data))
+
+        return model_data
+
+    def load_model(self, model_filename):
+        self.logger.info(translations["loading_model"].format(model_filename=model_filename))
+
+        load_model_start_time = time.perf_counter()
+
+        model_filename, model_type, model_friendly_name, model_path, yaml_config_filename = self.download_model_files(model_filename)
+        model_name = model_filename.split(".")[0]
+        self.logger.debug(translations["download_model_friendly_2"].format(model_friendly_name=model_friendly_name, model_path=model_path))
+
+        if model_path.lower().endswith(".yaml"): yaml_config_filename = model_path
+
+        model_data = self.load_model_data_from_yaml(yaml_config_filename) if yaml_config_filename is not None else self.load_model_data_using_hash(model_path)
+
+        common_params = {
+            "logger": self.logger,
+            "log_level": self.log_level,
+            "torch_device": self.torch_device,
+            "torch_device_cpu": self.torch_device_cpu,
+            "torch_device_mps": self.torch_device_mps,
+            "onnx_execution_provider": self.onnx_execution_provider,
+            "model_name": model_name,
+            "model_path": model_path,
+            "model_data": model_data,
+            "output_format": self.output_format,
+            "output_bitrate": self.output_bitrate,
+            "output_dir": self.output_dir,
+            "normalization_threshold": self.normalization_threshold,
+            "output_single_stem": self.output_single_stem,
+            "invert_using_spec": self.invert_using_spec,
+            "sample_rate": self.sample_rate,
+        }
+
+        separator_classes = {"MDX": "mdx_separator.MDXSeparator", "Demucs": "demucs_separator.DemucsSeparator"}
+
+        if model_type not in self.arch_specific_params or model_type not in separator_classes: raise ValueError(translations["model_type_not_support"].format(model_type=model_type))
+        if model_type == "Demucs" and sys.version_info < (3, 10): raise Exception(translations["demucs_not_support_python<3.10"])
+
+        self.logger.debug(f"{translations['import_module']} {model_type}: {separator_classes[model_type]}")
+
+        module_name, class_name = separator_classes[model_type].split(".")
+        module = import_module(f"main.library.architectures.{module_name}")
+        separator_class = getattr(module, class_name)
+
+        self.logger.debug(f"{translations['initialization']} {model_type}: {separator_class}")
+        self.model_instance = separator_class(common_config=common_params, arch_config=self.arch_specific_params[model_type])
+
+        self.logger.debug(translations["loading_model_success"])
+        self.logger.info(f"{translations['loading_model_duration']}: {time.strftime('%H:%M:%S', time.gmtime(int(time.perf_counter() - load_model_start_time)))}")
+
+    def separate(self, audio_file_path):
+        self.logger.info(f"{translations['starting_separator']}: {audio_file_path}")
+        separate_start_time = time.perf_counter()
+
+        self.logger.debug(translations["normalization"].format(normalization_threshold=self.normalization_threshold))
+
+        output_files = self.model_instance.separate(audio_file_path)
+
+        self.model_instance.clear_gpu_cache()
+
+        self.model_instance.clear_file_specific_paths()
+
+        self.print_uvr_vip_message()
+
+        self.logger.debug(translations["separator_success_3"])
+        self.logger.info(f"{translations['separator_duration']}: {time.strftime('%H:%M:%S', time.gmtime(int(time.perf_counter() - separate_start_time)))}")
+
+        return output_files
+
+    def download_model_and_data(self, model_filename):
+        self.logger.info(translations["loading_separator_model"].format(model_filename=model_filename))
+
+        model_filename, model_type, model_friendly_name, model_path, yaml_config_filename = self.download_model_files(model_filename)
+
+        if model_path.lower().endswith(".yaml"): yaml_config_filename = model_path
+
+        model_data = self.load_model_data_from_yaml(yaml_config_filename) if yaml_config_filename is not None else self.load_model_data_using_hash(model_path)
+
+        model_data_dict_size = len(model_data)
+
+        self.logger.info(translations["downloading_model"].format(model_type=model_type, model_friendly_name=model_friendly_name, model_path=model_path, model_data_dict_size=model_data_dict_size))

main/library/algorithm/synthesizers.py

@@ -0,0 +1,590 @@
+import os
+import sys
+import math
+import torch
+
+from typing import Optional
+from torch.nn.utils import remove_weight_norm
+from torch.nn.utils.parametrizations import weight_norm
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from .modules import WaveNet
+from .residuals import ResidualCouplingBlock, ResBlock1, ResBlock2, LRELU_SLOPE
+from .commons import init_weights, slice_segments, rand_slice_segments, sequence_mask, convert_pad_shape
+
+class Generator(torch.nn.Module):
+    def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
+        super(Generator, self).__init__()
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = torch.nn.Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
+        resblock = ResBlock1 if resblock == "1" else ResBlock2
+
+        self.ups_and_resblocks = torch.nn.ModuleList()
+
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups_and_resblocks.append(weight_norm(torch.nn.ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), k, u, padding=(k - u) // 2)))
+
+            ch = upsample_initial_channel // (2 ** (i + 1))
+
+            for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+                self.ups_and_resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = torch.nn.Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups_and_resblocks.apply(init_weights)
+
+        if gin_channels != 0: self.cond = torch.nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        def forward(self, x: torch.Tensor, g: Optional[torch.Tensor] = None):
+            x = self.conv_pre(x)
+            if g is not None: x = x + self.cond(g)
+
+            resblock_idx = 0
+
+            for _ in range(self.num_upsamples):
+                x = torch.nn.functional.leaky_relu(x, LRELU_SLOPE)
+                x = self.ups_and_resblocks[resblock_idx](x)
+                resblock_idx += 1
+                xs = 0
+
+                for _ in range(self.num_kernels):
+                    xs += self.ups_and_resblocks[resblock_idx](x)
+                    resblock_idx += 1
+
+                x = xs / self.num_kernels
+
+            x = torch.nn.functional.leaky_relu(x)
+            x = self.conv_post(x)
+            x = torch.tanh(x)
+
+            return x
+
+    def __prepare_scriptable__(self):
+        for l in self.ups_and_resblocks:
+            for hook in l._forward_pre_hooks.values():
+                if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(l)
+
+        return self
+    def remove_weight_norm(self):
+        for l in self.ups_and_resblocks:
+            remove_weight_norm(l)
+
+class SineGen(torch.nn.Module):
+    def __init__(self, samp_rate, harmonic_num=0, sine_amp=0.1, noise_std=0.003, voiced_threshold=0, flag_for_pulse=False):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sample_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        uv = torch.ones_like(f0)
+        uv = uv * (f0 > self.voiced_threshold)
+        return uv
+
+    def forward(self, f0: torch.Tensor, upp: int):
+        with torch.no_grad():
+            f0 = f0[:, None].transpose(1, 2)
+            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+            f0_buf[:, :, 0] = f0[:, :, 0]
+            f0_buf[:, :, 1:] = (f0_buf[:, :, 0:1] * torch.arange(2, self.harmonic_num + 2, device=f0.device)[None, None, :])
+            rad_values = (f0_buf / float(self.sample_rate)) % 1
+            rand_ini = torch.rand(f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device)
+            rand_ini[:, 0] = 0
+            rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+            tmp_over_one = torch.cumsum(rad_values, 1)
+            tmp_over_one *= upp
+            tmp_over_one = torch.nn.functional.interpolate(tmp_over_one.transpose(2, 1), scale_factor=float(upp), mode="linear", align_corners=True).transpose(2, 1)
+            rad_values = torch.nn.functional.interpolate(rad_values.transpose(2, 1), scale_factor=float(upp), mode="nearest").transpose(2, 1)
+            tmp_over_one %= 1
+            tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
+            cumsum_shift = torch.zeros_like(rad_values)
+            cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+            sine_waves = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * torch.pi)
+            sine_waves = sine_waves * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = torch.nn.functional.interpolate(uv.transpose(2, 1), scale_factor=float(upp), mode="nearest").transpose(2, 1)
+            noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+            noise = noise_amp * torch.randn_like(sine_waves)
+            sine_waves = sine_waves * uv + noise
+            
+        return sine_waves, uv, noise
+
+class SourceModuleHnNSF(torch.nn.Module):
+    def __init__(self, sample_rate, harmonic_num=0, sine_amp=0.1, add_noise_std=0.003, voiced_threshod=0, is_half=True):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+        self.is_half = is_half
+
+        self.l_sin_gen = SineGen(sample_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod)
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x: torch.Tensor, upsample_factor: int = 1):
+        sine_wavs, uv, _ = self.l_sin_gen(x, upsample_factor)
+        sine_wavs = sine_wavs.to(dtype=self.l_linear.weight.dtype)
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+        return sine_merge, None, None
+
+
+class GeneratorNSF(torch.nn.Module):
+    def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels, sr, is_half=False):
+        super(GeneratorNSF, self).__init__()
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=math.prod(upsample_rates))
+        self.m_source = SourceModuleHnNSF(sample_rate=sr, harmonic_num=0, is_half=is_half)
+
+        self.conv_pre = torch.nn.Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
+        resblock_cls = ResBlock1 if resblock == "1" else ResBlock2
+
+        self.ups = torch.nn.ModuleList()
+        self.noise_convs = torch.nn.ModuleList()
+
+        channels = [upsample_initial_channel // (2 ** (i + 1)) for i in range(len(upsample_rates))]
+        stride_f0s = [math.prod(upsample_rates[i + 1 :]) if i + 1 < len(upsample_rates) else 1 for i in range(len(upsample_rates))]
+
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(weight_norm(torch.nn.ConvTranspose1d(upsample_initial_channel // (2**i), channels[i], k, u, padding=(k - u) // 2)))
+            self.noise_convs.append(torch.nn.Conv1d(1, channels[i], kernel_size=(stride_f0s[i] * 2 if stride_f0s[i] > 1 else 1), stride=stride_f0s[i], padding=(stride_f0s[i] // 2 if stride_f0s[i] > 1 else 0)))
+
+        self.resblocks = torch.nn.ModuleList([resblock_cls(channels[i], k, d) for i in range(len(self.ups)) for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes)])
+
+        self.conv_post = torch.nn.Conv1d(channels[-1], 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0: self.cond = torch.nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        self.upp = math.prod(upsample_rates)
+        self.lrelu_slope = LRELU_SLOPE
+
+    def forward(self, x, f0, g: Optional[torch.Tensor] = None):
+        har_source, _, _ = self.m_source(f0, self.upp)
+        har_source = har_source.transpose(1, 2)
+        x = self.conv_pre(x)
+
+        if g is not None: x = x + self.cond(g)
+
+        for i, (ups, noise_convs) in enumerate(zip(self.ups, self.noise_convs)):
+            x = torch.nn.functional.leaky_relu(x, self.lrelu_slope)
+            x = ups(x)
+            x = x + noise_convs(har_source)
+
+            xs = sum([resblock(x) for j, resblock in enumerate(self.resblocks) if j in range(i * self.num_kernels, (i + 1) * self.num_kernels)])
+            x = xs / self.num_kernels
+
+        x = torch.nn.functional.leaky_relu(x)
+        x = torch.tanh(self.conv_post(x))
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for l in self.ups:
+            for hook in l._forward_pre_hooks.values():
+                if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm"  and hook.__class__.__name__ == "WeightNorm"): remove_weight_norm(l)
+
+        for l in self.resblocks:
+            for hook in l._forward_pre_hooks.values():
+                if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): remove_weight_norm(l)
+        return self
+
+class LayerNorm(torch.nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.eps = eps
+        self.gamma = torch.nn.Parameter(torch.ones(channels))
+        self.beta = torch.nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = torch.nn.functional.layer_norm(x, (x.size(-1),), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+class MultiHeadAttention(torch.nn.Module):
+    def __init__(self, channels, out_channels, n_heads, p_dropout=0.0, window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
+        super().__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+        self.k_channels = channels // n_heads
+        self.conv_q = torch.nn.Conv1d(channels, channels, 1)
+        self.conv_k = torch.nn.Conv1d(channels, channels, 1)
+        self.conv_v = torch.nn.Conv1d(channels, channels, 1)
+        self.conv_o = torch.nn.Conv1d(channels, out_channels, 1)
+        self.drop = torch.nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+
+            self.emb_rel_k = torch.nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
+            self.emb_rel_v = torch.nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
+
+        torch.nn.init.xavier_uniform_(self.conv_q.weight)
+        torch.nn.init.xavier_uniform_(self.conv_k.weight)
+        torch.nn.init.xavier_uniform_(self.conv_v.weight)
+
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(self, x, c, attn_mask=None):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(self, query, key, value, mask=None):
+        b, d, t_s, t_t = (*key.size(), query.size(2))
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            assert (t_s == t_t), "(t_s == t_t)"
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(query / math.sqrt(self.k_channels), key_relative_embeddings)
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+
+        if self.proximal_bias:
+            assert t_s == t_t, "t_s == t_t"
+            scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
+
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (t_s == t_t), "(t_s == t_t)"
+
+                block_mask = (torch.ones_like(scores).triu(-self.block_length).tril(self.block_length))
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+
+        p_attn = torch.nn.functional.softmax(scores, dim=-1)  
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
+            output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
+
+        output = (output.transpose(2, 3).contiguous().view(b, d, t_t)) 
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        pad_length = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+
+        if pad_length > 0: padded_relative_embeddings = torch.nn.functional.pad(relative_embeddings, convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
+        else: padded_relative_embeddings = relative_embeddings
+
+        used_relative_embeddings = padded_relative_embeddings[:, slice_start_position:slice_end_position]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        batch, heads, length, _ = x.size()
+
+        x = torch.nn.functional.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
+
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = torch.nn.functional.pad(x_flat, convert_pad_shape([[0, 0], [0, 0], [0, length - 1]]))
+
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[:, :, :length, length - 1 :]
+
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        batch, heads, length, _ = x.size()
+        x = torch.nn.functional.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]]))
+        x_flat = x.view([batch, heads, length**2 + length * (length - 1)])
+        x_flat = torch.nn.functional.pad(x_flat, convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length):
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(torch.nn.Module):
+    def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0.0, activation=None, causal=False):
+        super().__init__()
+        
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+
+        if causal: self.padding = self._causal_padding
+        else: self.padding = self._same_padding
+
+        self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = torch.nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = torch.nn.Dropout(p_dropout)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(self.padding(x * x_mask))
+
+        if self.activation == "gelu": x = x * torch.sigmoid(1.702 * x)
+        else: x = torch.relu(x)
+
+        x = self.drop(x)
+        x = self.conv_2(self.padding(x * x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1: return x
+
+        pad_l = self.kernel_size - 1
+        pad_r = 0
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = torch.nn.functional.pad(x, convert_pad_shape(padding))
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1: return x
+        
+        pad_l = (self.kernel_size - 1) // 2
+        pad_r = self.kernel_size // 2
+
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = torch.nn.functional.pad(x, convert_pad_shape(padding))
+
+        return x
+
+class Encoder(torch.nn.Module):
+    def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0.0, window_size=10, **kwargs):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.drop = torch.nn.Dropout(p_dropout)
+        self.attn_layers = torch.nn.ModuleList()
+        self.norm_layers_1 = torch.nn.ModuleList()
+        self.ffn_layers = torch.nn.ModuleList()
+        self.norm_layers_2 = torch.nn.ModuleList()
+
+        for _ in range(self.n_layers):
+            self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+
+        for i in range(self.n_layers):
+            y = self.attn_layers[i](x, x, attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+
+        x = x * x_mask
+        return x
+
+
+class TextEncoder(torch.nn.Module):
+    def __init__(self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, embedding_dim, f0=True):
+        super(TextEncoder, self).__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = float(p_dropout)
+        self.emb_phone = torch.nn.Linear(embedding_dim, hidden_channels)
+        self.lrelu = torch.nn.LeakyReLU(0.1, inplace=True)
+
+        if f0: self.emb_pitch = torch.nn.Embedding(256, hidden_channels)
+
+        self.encoder = Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, float(p_dropout))
+        self.proj = torch.nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone: torch.Tensor, pitch: Optional[torch.Tensor], lengths: torch.Tensor):
+        if pitch is None: x = self.emb_phone(phone)
+        else: x = self.emb_phone(phone) + self.emb_pitch(pitch)
+
+        x = x * math.sqrt(self.hidden_channels)  
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1) 
+        x_mask = torch.unsqueeze(sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class PosteriorEncoder(torch.nn.Module):
+    def __init__(self, in_channels, out_channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0):
+        super(PosteriorEncoder, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = torch.nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = WaveNet(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
+        self.proj = torch.nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x: torch.Tensor, x_lengths: torch.Tensor, g: Optional[torch.Tensor] = None):
+        x_mask = torch.unsqueeze(sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for hook in self.enc._forward_pre_hooks.values():
+            if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(self.enc)
+
+        return self
+
+class Synthesizer(torch.nn.Module):
+    def __init__(self, spec_channels, segment_size, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, spk_embed_dim, gin_channels, sr, use_f0, text_enc_hidden_dim=768, **kwargs):
+        super(Synthesizer, self).__init__()
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = float(p_dropout)
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        self.spk_embed_dim = spk_embed_dim
+        self.use_f0 = use_f0
+
+        self.enc_p = TextEncoder(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, float(p_dropout), text_enc_hidden_dim, f0=use_f0)
+
+        if use_f0: self.dec = GeneratorNSF(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels, sr=sr, is_half=kwargs["is_half"])
+        else: self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
+
+        self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
+        self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels)
+        self.emb_g = torch.nn.Embedding(self.spk_embed_dim, gin_channels)
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for hook in self.dec._forward_pre_hooks.values():
+            if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(self.dec)
+
+        for hook in self.flow._forward_pre_hooks.values():
+            if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(self.flow)
+
+        if hasattr(self, "enc_q"):
+            for hook in self.enc_q._forward_pre_hooks.values():
+                if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(self.enc_q)
+
+        return self
+
+    @torch.jit.ignore
+    def forward(self, phone: torch.Tensor, phone_lengths: torch.Tensor, pitch: Optional[torch.Tensor] = None, pitchf: Optional[torch.Tensor] = None, y: torch.Tensor = None, y_lengths: torch.Tensor = None, ds: Optional[torch.Tensor] = None):
+        g = self.emb_g(ds).unsqueeze(-1)
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+
+        if y is not None:
+            z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+            z_p = self.flow(z, y_mask, g=g)
+            z_slice, ids_slice = rand_slice_segments(z, y_lengths, self.segment_size)
+
+            if self.use_f0:
+                pitchf = slice_segments(pitchf, ids_slice, self.segment_size, 2)
+                o = self.dec(z_slice, pitchf, g=g)
+            else: o = self.dec(z_slice, g=g)
+
+            return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+        else: return None, None, x_mask, None, (None, None, m_p, logs_p, None, None)
+
+    @torch.jit.export
+    def infer(self, phone: torch.Tensor, phone_lengths: torch.Tensor, pitch: Optional[torch.Tensor] = None, nsff0: Optional[torch.Tensor] = None, sid: torch.Tensor = None, rate: Optional[torch.Tensor] = None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+
+        if rate is not None:
+            assert isinstance(rate, torch.Tensor)
+            head = int(z_p.shape[2] * (1.0 - rate.item()))
+            z_p = z_p[:, :, head:]
+            x_mask = x_mask[:, :, head:]
+
+            if self.use_f0: nsff0 = nsff0[:, head:]
+
+        if self.use_f0:
+            z = self.flow(z_p, x_mask, g=g, reverse=True)
+            o = self.dec(z * x_mask, nsff0, g=g)
+        else:
+            z = self.flow(z_p, x_mask, g=g, reverse=True)
+            o = self.dec(z * x_mask, g=g)
+
+        return o, x_mask, (z, z_p, m_p, logs_p)

main/library/architectures/demucs_separator.py

@@ -0,0 +1,340 @@
+import os
+import sys
+import yaml
+import torch
+
+import numpy as np
+import typing as tp
+
+from pathlib import Path
+from hashlib import sha256
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+from main.library.uvr5_separator import spec_utils
+from main.library.uvr5_separator.demucs.hdemucs import HDemucs
+from main.library.uvr5_separator.demucs.states import load_model
+from main.library.uvr5_separator.demucs.apply import BagOfModels, Model
+from main.library.uvr5_separator.common_separator import CommonSeparator
+from main.library.uvr5_separator.demucs.apply import apply_model, demucs_segments
+
+
+translations = Config().translations
+
+DEMUCS_4_SOURCE = ["drums", "bass", "other", "vocals"]
+
+DEMUCS_2_SOURCE_MAPPER = {
+    CommonSeparator.INST_STEM: 0, 
+    CommonSeparator.VOCAL_STEM: 1
+}
+
+DEMUCS_4_SOURCE_MAPPER = {
+    CommonSeparator.BASS_STEM: 0, 
+    CommonSeparator.DRUM_STEM: 1, 
+    CommonSeparator.OTHER_STEM: 2, 
+    CommonSeparator.VOCAL_STEM: 3
+}
+
+DEMUCS_6_SOURCE_MAPPER = {
+    CommonSeparator.BASS_STEM: 0,
+    CommonSeparator.DRUM_STEM: 1,
+    CommonSeparator.OTHER_STEM: 2,
+    CommonSeparator.VOCAL_STEM: 3,
+    CommonSeparator.GUITAR_STEM: 4,
+    CommonSeparator.PIANO_STEM: 5,
+}
+
+
+REMOTE_ROOT = Path(__file__).parent / "remote"
+
+PRETRAINED_MODELS = {
+    "demucs": "e07c671f",
+    "demucs48_hq": "28a1282c",
+    "demucs_extra": "3646af93",
+    "demucs_quantized": "07afea75",
+    "tasnet": "beb46fac",
+    "tasnet_extra": "df3777b2",
+    "demucs_unittest": "09ebc15f",
+}
+
+
+sys.path.insert(0, os.path.join(os.getcwd(), "main", "library", "uvr5_separator"))
+
+AnyModel = tp.Union[Model, BagOfModels]
+
+
+class DemucsSeparator(CommonSeparator):
+    def __init__(self, common_config, arch_config):
+        super().__init__(config=common_config)
+
+        self.segment_size = arch_config.get("segment_size", "Default")
+        self.shifts = arch_config.get("shifts", 2)
+        self.overlap = arch_config.get("overlap", 0.25)
+        self.segments_enabled = arch_config.get("segments_enabled", True)
+
+        self.logger.debug(translations["demucs_info"].format(segment_size=self.segment_size, segments_enabled=self.segments_enabled))
+        self.logger.debug(translations["demucs_info_2"].format(shifts=self.shifts, overlap=self.overlap))
+
+        self.demucs_source_map = DEMUCS_4_SOURCE_MAPPER
+
+        self.audio_file_path = None
+        self.audio_file_base = None
+        self.demucs_model_instance = None
+
+        self.logger.info(translations["start_demucs"])
+
+    def separate(self, audio_file_path):
+        self.logger.debug(translations["start_separator"])
+        
+        source = None
+        stem_source = None
+
+        inst_source = {}
+
+        self.audio_file_path = audio_file_path
+        self.audio_file_base = os.path.splitext(os.path.basename(audio_file_path))[0]
+
+        self.logger.debug(translations["prepare_mix"])
+        mix = self.prepare_mix(self.audio_file_path)
+
+        self.logger.debug(translations["demix"].format(shape=mix.shape))
+
+        self.logger.debug(translations["cancel_mix"])
+
+        self.demucs_model_instance = HDemucs(sources=DEMUCS_4_SOURCE)
+        self.demucs_model_instance = get_demucs_model(name=os.path.splitext(os.path.basename(self.model_path))[0], repo=Path(os.path.dirname(self.model_path)))
+        self.demucs_model_instance = demucs_segments(self.segment_size, self.demucs_model_instance)
+        self.demucs_model_instance.to(self.torch_device)
+        self.demucs_model_instance.eval()
+
+        self.logger.debug(translations["model_review"])
+
+        source = self.demix_demucs(mix)
+
+        del self.demucs_model_instance
+        self.clear_gpu_cache()
+        self.logger.debug(translations["del_gpu_cache_after_demix"])
+
+        output_files = []
+        self.logger.debug(translations["process_output_file"])
+
+        if isinstance(inst_source, np.ndarray):
+            self.logger.debug(translations["process_ver"])
+            source_reshape = spec_utils.reshape_sources(inst_source[self.demucs_source_map[CommonSeparator.VOCAL_STEM]], source[self.demucs_source_map[CommonSeparator.VOCAL_STEM]])
+            inst_source[self.demucs_source_map[CommonSeparator.VOCAL_STEM]] = source_reshape
+            source = inst_source
+
+        if isinstance(source, np.ndarray):
+            source_length = len(source)
+            self.logger.debug(translations["source_length"].format(source_length=source_length))
+
+            match source_length:
+                case 2:
+                    self.logger.debug(translations["set_map"].format(part="2"))
+                    self.demucs_source_map = DEMUCS_2_SOURCE_MAPPER
+                case 6:
+                    self.logger.debug(translations["set_map"].format(part="6"))
+                    self.demucs_source_map = DEMUCS_6_SOURCE_MAPPER
+                case _:
+                    self.logger.debug(translations["set_map"].format(part="2"))
+                    self.demucs_source_map = DEMUCS_4_SOURCE_MAPPER
+
+        self.logger.debug(translations["process_all_part"])
+
+        for stem_name, stem_value in self.demucs_source_map.items():
+            if self.output_single_stem is not None:
+                if stem_name.lower() != self.output_single_stem.lower():
+                    self.logger.debug(translations["skip_part"].format(stem_name=stem_name, output_single_stem=self.output_single_stem))
+                    continue
+
+            stem_path = os.path.join(f"{self.audio_file_base}_({stem_name})_{self.model_name}.{self.output_format.lower()}")
+            stem_source = source[stem_value].T
+
+            self.final_process(stem_path, stem_source, stem_name)
+            output_files.append(stem_path)
+
+        return output_files
+
+    def demix_demucs(self, mix):
+        self.logger.debug(translations["starting_demix_demucs"])
+
+        processed = {}
+        mix = torch.tensor(mix, dtype=torch.float32)
+        ref = mix.mean(0)
+        mix = (mix - ref.mean()) / ref.std()
+        mix_infer = mix
+
+        with torch.no_grad():
+            self.logger.debug(translations["model_infer"])
+            sources = apply_model(model=self.demucs_model_instance, mix=mix_infer[None], shifts=self.shifts, split=self.segments_enabled, overlap=self.overlap, static_shifts=1 if self.shifts == 0 else self.shifts, set_progress_bar=None, device=self.torch_device, progress=True)[0]
+
+        sources = (sources * ref.std() + ref.mean()).cpu().numpy()
+        sources[[0, 1]] = sources[[1, 0]]
+
+        processed[mix] = sources[:, :, 0:None].copy()
+
+        sources = list(processed.values())
+        sources = [s[:, :, 0:None] for s in sources]
+        sources = np.concatenate(sources, axis=-1)
+        return sources
+
+
+class ModelOnlyRepo:
+    def has_model(self, sig: str) -> bool:
+        raise NotImplementedError()
+
+    def get_model(self, sig: str) -> Model:
+        raise NotImplementedError()
+
+
+class RemoteRepo(ModelOnlyRepo):
+    def __init__(self, models: tp.Dict[str, str]):
+        self._models = models
+
+    def has_model(self, sig: str) -> bool:
+        return sig in self._models
+
+    def get_model(self, sig: str) -> Model:
+        try:
+            url = self._models[sig]
+        except KeyError:
+            raise RuntimeError(translations["not_found_model_signature"].format(sig=sig))
+        
+        pkg = torch.hub.load_state_dict_from_url(url, map_location="cpu", check_hash=True)
+        return load_model(pkg)
+
+
+class LocalRepo(ModelOnlyRepo):
+    def __init__(self, root: Path):
+        self.root = root
+        self.scan()
+
+    def scan(self):
+        self._models = {}
+        self._checksums = {}
+
+        for file in self.root.iterdir():
+            if file.suffix == ".th":
+                if "-" in file.stem:
+                    xp_sig, checksum = file.stem.split("-")
+                    self._checksums[xp_sig] = checksum
+                else: xp_sig = file.stem
+
+                if xp_sig in self._models: raise RuntimeError(translations["del_all_but_one"].format(xp_sig=xp_sig))
+                
+                self._models[xp_sig] = file
+
+    def has_model(self, sig: str) -> bool:
+        return sig in self._models
+
+    def get_model(self, sig: str) -> Model:
+        try:
+            file = self._models[sig]
+        except KeyError:
+            raise RuntimeError(translations["not_found_model_signature"].format(sig=sig))
+        
+        if sig in self._checksums: check_checksum(file, self._checksums[sig])
+
+        return load_model(file)
+
+
+class BagOnlyRepo:
+    def __init__(self, root: Path, model_repo: ModelOnlyRepo):
+        self.root = root
+        self.model_repo = model_repo
+        self.scan()
+
+    def scan(self):
+        self._bags = {}
+
+        for file in self.root.iterdir():
+            if file.suffix == ".yaml": self._bags[file.stem] = file
+
+    def has_model(self, name: str) -> bool:
+        return name in self._bags
+
+    def get_model(self, name: str) -> BagOfModels:
+        try:
+            yaml_file = self._bags[name]
+        except KeyError:
+            raise RuntimeError(translations["name_not_pretrained"].format(name=name))
+        
+        bag = yaml.safe_load(open(yaml_file))
+        signatures = bag["models"]
+        models = [self.model_repo.get_model(sig) for sig in signatures]
+
+        weights = bag.get("weights")
+        segment = bag.get("segment")
+
+        return BagOfModels(models, weights, segment)
+
+
+class AnyModelRepo:
+    def __init__(self, model_repo: ModelOnlyRepo, bag_repo: BagOnlyRepo):
+        self.model_repo = model_repo
+        self.bag_repo = bag_repo
+
+    def has_model(self, name_or_sig: str) -> bool:
+        return self.model_repo.has_model(name_or_sig) or self.bag_repo.has_model(name_or_sig)
+
+    def get_model(self, name_or_sig: str) -> AnyModel:
+        if self.model_repo.has_model(name_or_sig): return self.model_repo.get_model(name_or_sig)
+        else: return self.bag_repo.get_model(name_or_sig)
+
+
+def check_checksum(path: Path, checksum: str):
+    sha = sha256()
+
+    with open(path, "rb") as file:
+        while 1:
+            buf = file.read(2**20)
+            if not buf: break
+
+            sha.update(buf)
+
+    actual_checksum = sha.hexdigest()[: len(checksum)]
+
+    if actual_checksum != checksum: raise RuntimeError(translations["invalid_checksum"].format(path=path, checksum=checksum, actual_checksum=actual_checksum))
+
+
+def _parse_remote_files(remote_file_list) -> tp.Dict[str, str]:
+    root: str = ""
+    models: tp.Dict[str, str] = {}
+
+    for line in remote_file_list.read_text().split("\n"):
+        line = line.strip()
+
+        if line.startswith("#"): continue
+        elif line.startswith("root:"): root = line.split(":", 1)[1].strip()
+        else:
+            sig = line.split("-", 1)[0]
+            assert sig not in models
+
+            models[sig] = "https://dl.fbaipublicfiles.com/demucs/mdx_final/" + root + line
+
+    return models
+
+
+def get_demucs_model(name: str, repo: tp.Optional[Path] = None):
+    if name == "demucs_unittest": return HDemucs(channels=4, sources=DEMUCS_4_SOURCE)
+
+    model_repo: ModelOnlyRepo
+
+    if repo is None:
+        models = _parse_remote_files(REMOTE_ROOT / "files.txt")
+        model_repo = RemoteRepo(models)
+        bag_repo = BagOnlyRepo(REMOTE_ROOT, model_repo)
+    else:
+        if not repo.is_dir(): print(translations["repo_must_be_folder"].format(repo=repo))
+
+        model_repo = LocalRepo(repo)
+        bag_repo = BagOnlyRepo(repo, model_repo)
+
+    any_repo = AnyModelRepo(model_repo, bag_repo)
+
+    model = any_repo.get_model(name)
+    model.eval()
+
+    return model

main/library/architectures/mdx_separator.py

@@ -0,0 +1,370 @@
+import os
+import sys
+import onnx
+import torch
+import platform
+import onnx2torch
+
+import numpy as np
+import onnxruntime as ort
+
+from tqdm import tqdm
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+from main.configs.config import Config
+from main.library.uvr5_separator import spec_utils
+from main.library.uvr5_separator.common_separator import CommonSeparator
+
+
+translations = Config().translations
+
+class MDXSeparator(CommonSeparator):
+    def __init__(self, common_config, arch_config):
+        super().__init__(config=common_config)
+
+        self.segment_size = arch_config.get("segment_size")
+        self.overlap = arch_config.get("overlap")
+        self.batch_size = arch_config.get("batch_size", 1)
+        self.hop_length = arch_config.get("hop_length")
+        self.enable_denoise = arch_config.get("enable_denoise")
+        self.logger.debug(translations["mdx_info"].format(batch_size=self.batch_size, segment_size=self.segment_size))
+        self.logger.debug(translations["mdx_info_2"].format(overlap=self.overlap, hop_length=self.hop_length, enable_denoise=self.enable_denoise))
+        self.compensate = self.model_data["compensate"]
+        self.dim_f = self.model_data["mdx_dim_f_set"]
+        self.dim_t = 2 ** self.model_data["mdx_dim_t_set"]
+        self.n_fft = self.model_data["mdx_n_fft_scale_set"]
+        self.config_yaml = self.model_data.get("config_yaml", None)
+        self.logger.debug(f"{translations['mdx_info_3']}: compensate = {self.compensate}, dim_f = {self.dim_f}, dim_t = {self.dim_t}, n_fft = {self.n_fft}")
+        self.logger.debug(f"{translations['mdx_info_3']}: config_yaml = {self.config_yaml}")
+        self.load_model()
+        self.n_bins = 0
+        self.trim = 0
+        self.chunk_size = 0
+        self.gen_size = 0
+        self.stft = None
+        self.primary_source = None
+        self.secondary_source = None
+        self.audio_file_path = None
+        self.audio_file_base = None
+
+
+    def load_model(self):
+        self.logger.debug(translations["load_model_onnx"])
+
+        if self.segment_size == self.dim_t:
+            ort_session_options = ort.SessionOptions()
+
+            if self.log_level > 10: ort_session_options.log_severity_level = 3
+            else: ort_session_options.log_severity_level = 0
+
+            ort_inference_session = ort.InferenceSession(self.model_path, providers=self.onnx_execution_provider, sess_options=ort_session_options)
+            self.model_run = lambda spek: ort_inference_session.run(None, {"input": spek.cpu().numpy()})[0]
+            self.logger.debug(translations["load_model_onnx_success"])
+        else:
+            if platform.system() == 'Windows':
+                onnx_model = onnx.load(self.model_path)
+                self.model_run = onnx2torch.convert(onnx_model)
+            else: self.model_run = onnx2torch.convert(self.model_path)
+   
+            self.model_run.to(self.torch_device).eval()
+            self.logger.warning(translations["onnx_to_pytorch"])
+
+    def separate(self, audio_file_path):
+        self.audio_file_path = audio_file_path
+        self.audio_file_base = os.path.splitext(os.path.basename(audio_file_path))[0]
+
+        self.logger.debug(translations["mix"].format(audio_file_path=self.audio_file_path))
+        mix = self.prepare_mix(self.audio_file_path)
+
+        self.logger.debug(translations["normalization_demix"])
+        mix = spec_utils.normalize(wave=mix, max_peak=self.normalization_threshold)
+
+        source = self.demix(mix)
+        self.logger.debug(translations["mix_success"])
+
+        output_files = []
+        self.logger.debug(translations["process_output_file"])
+
+        if not isinstance(self.primary_source, np.ndarray):
+            self.logger.debug(translations["primary_source"])
+            self.primary_source = spec_utils.normalize(wave=source, max_peak=self.normalization_threshold).T
+        if not isinstance(self.secondary_source, np.ndarray):
+            self.logger.debug(translations["secondary_source"])
+            raw_mix = self.demix(mix, is_match_mix=True)
+
+            if self.invert_using_spec:
+                self.logger.debug(translations["invert_using_spec"])
+                self.secondary_source = spec_utils.invert_stem(raw_mix, source)
+            else:
+                self.logger.debug(translations["invert_using_spec_2"])
+                self.secondary_source = mix.T - source.T
+
+        if not self.output_single_stem or self.output_single_stem.lower() == self.secondary_stem_name.lower():
+            self.secondary_stem_output_path = os.path.join(f"{self.audio_file_base}_({self.secondary_stem_name})_{self.model_name}.{self.output_format.lower()}")
+            self.logger.info(translations["save_secondary_stem_output_path"].format(stem_name=self.secondary_stem_name, stem_output_path=self.secondary_stem_output_path))
+            self.final_process(self.secondary_stem_output_path, self.secondary_source, self.secondary_stem_name)
+            output_files.append(self.secondary_stem_output_path)
+
+        if not self.output_single_stem or self.output_single_stem.lower() == self.primary_stem_name.lower():
+            self.primary_stem_output_path = os.path.join(f"{self.audio_file_base}_({self.primary_stem_name})_{self.model_name}.{self.output_format.lower()}")
+
+            if not isinstance(self.primary_source, np.ndarray): self.primary_source = source.T
+
+            self.logger.info(translations["save_secondary_stem_output_path"].format(stem_name=self.primary_stem_name, stem_output_path=self.primary_stem_output_path))
+            self.final_process(self.primary_stem_output_path, self.primary_source, self.primary_stem_name)
+            output_files.append(self.primary_stem_output_path)
+
+        return output_files
+
+    def initialize_model_settings(self):
+        self.logger.debug(translations["starting_model"])
+
+        self.n_bins = self.n_fft // 2 + 1
+        self.trim = self.n_fft // 2
+
+        self.chunk_size = self.hop_length * (self.segment_size - 1)
+        self.gen_size = self.chunk_size - 2 * self.trim
+
+        self.stft = STFT(self.logger, self.n_fft, self.hop_length, self.dim_f, self.torch_device)
+
+        self.logger.debug(f"{translations['input_info']}: n_fft = {self.n_fft} hop_length = {self.hop_length} dim_f = {self.dim_f}")
+        self.logger.debug(f"{translations['model_settings']}: n_bins = {self.n_bins}, Trim = {self.trim}, chunk_size = {self.chunk_size}, gen_size = {self.gen_size}")
+
+    def initialize_mix(self, mix, is_ckpt=False):
+        self.logger.debug(translations["initialize_mix"].format(is_ckpt=is_ckpt, shape=mix.shape))
+
+        if mix.shape[0] != 2:
+            error_message = translations["!=2"].format(shape=mix.shape[0])
+            self.logger.error(error_message)
+            raise ValueError(error_message)
+
+        if is_ckpt:
+            self.logger.debug(translations["process_check"])
+            pad = self.gen_size + self.trim - (mix.shape[-1] % self.gen_size)
+            self.logger.debug(f"{translations['cache']}: {pad}")
+
+            mixture = np.concatenate((np.zeros((2, self.trim), dtype="float32"), mix, np.zeros((2, pad), dtype="float32")), 1)
+
+            num_chunks = mixture.shape[-1] // self.gen_size
+            self.logger.debug(translations["shape"].format(shape=mixture.shape, num_chunks=num_chunks))
+
+            mix_waves = [mixture[:, i * self.gen_size : i * self.gen_size + self.chunk_size] for i in range(num_chunks)]
+        else:
+            self.logger.debug(translations["process_no_check"])
+            mix_waves = []
+            n_sample = mix.shape[1]
+
+            pad = self.gen_size - n_sample % self.gen_size
+            self.logger.debug(translations["n_sample_or_pad"].format(n_sample=n_sample, pad=pad))
+
+            mix_p = np.concatenate((np.zeros((2, self.trim)), mix, np.zeros((2, pad)), np.zeros((2, self.trim))), 1)
+            self.logger.debug(f"{translations['shape_2']}: {mix_p.shape}")
+
+            i = 0
+            while i < n_sample + pad:
+                waves = np.array(mix_p[:, i : i + self.chunk_size])
+                mix_waves.append(waves)
+
+                self.logger.debug(translations["process_part"].format(mix_waves=len(mix_waves), i=i, ii=i + self.chunk_size))
+                i += self.gen_size
+
+        mix_waves_tensor = torch.tensor(mix_waves, dtype=torch.float32).to(self.torch_device)
+        self.logger.debug(translations["mix_waves_to_tensor"].format(shape=mix_waves_tensor.shape))
+
+        return mix_waves_tensor, pad
+
+    def demix(self, mix, is_match_mix=False):
+        self.logger.debug(f"{translations['demix_is_match_mix']}: {is_match_mix}...")
+        self.initialize_model_settings()
+
+        org_mix = mix
+        self.logger.debug(f"{translations['mix_shape']}: {org_mix.shape}")
+
+        tar_waves_ = []
+
+        if is_match_mix:
+            chunk_size = self.hop_length * (self.segment_size - 1)
+            overlap = 0.02
+            self.logger.debug(translations["chunk_size_or_overlap"].format(chunk_size=chunk_size, overlap=overlap))
+        else:
+            chunk_size = self.chunk_size
+            overlap = self.overlap
+            self.logger.debug(translations["chunk_size_or_overlap_standard"].format(chunk_size=chunk_size, overlap=overlap))
+
+
+        gen_size = chunk_size - 2 * self.trim
+        self.logger.debug(f"{translations['calc_size']}: {gen_size}")
+
+
+        pad = gen_size + self.trim - ((mix.shape[-1]) % gen_size)
+
+        mixture = np.concatenate((np.zeros((2, self.trim), dtype="float32"), mix, np.zeros((2, pad), dtype="float32")), 1)
+        self.logger.debug(f"{translations['mix_cache']}: {mixture.shape}")
+
+        step = int((1 - overlap) * chunk_size)
+        self.logger.debug(translations["step_or_overlap"].format(step=step, overlap=overlap))
+
+        result = np.zeros((1, 2, mixture.shape[-1]), dtype=np.float32)
+        divider = np.zeros((1, 2, mixture.shape[-1]), dtype=np.float32)
+
+        total = 0
+        total_chunks = (mixture.shape[-1] + step - 1) // step
+        self.logger.debug(f"{translations['all_process_part']}: {total_chunks}")
+
+        for i in tqdm(range(0, mixture.shape[-1], step)):
+            total += 1
+            start = i
+            end = min(i + chunk_size, mixture.shape[-1])
+            self.logger.debug(translations["process_part_2"].format(total=total, total_chunks=total_chunks, start=start, end=end))
+
+            chunk_size_actual = end - start
+            window = None
+
+            if overlap != 0:
+                window = np.hanning(chunk_size_actual)
+                window = np.tile(window[None, None, :], (1, 2, 1))
+                self.logger.debug(translations["window"])
+
+            mix_part_ = mixture[:, start:end]
+            
+            if end != i + chunk_size:
+                pad_size = (i + chunk_size) - end
+                mix_part_ = np.concatenate((mix_part_, np.zeros((2, pad_size), dtype="float32")), axis=-1)
+
+            mix_part = torch.tensor([mix_part_], dtype=torch.float32).to(self.torch_device)
+
+            mix_waves = mix_part.split(self.batch_size)
+            total_batches = len(mix_waves)
+            self.logger.debug(f"{translations['mix_or_batch']}: {total_batches}")
+
+            with torch.no_grad():
+                batches_processed = 0
+                for mix_wave in mix_waves:
+                    batches_processed += 1
+                    self.logger.debug(f"{translations['mix_wave']} {batches_processed}/{total_batches}")
+
+                    tar_waves = self.run_model(mix_wave, is_match_mix=is_match_mix)
+
+                    if window is not None:
+                        tar_waves[..., :chunk_size_actual] *= window
+                        divider[..., start:end] += window
+                    else: divider[..., start:end] += 1
+
+                    result[..., start:end] += tar_waves[..., : end - start]
+
+
+        self.logger.debug(translations["normalization_2"])
+        tar_waves = result / divider
+        tar_waves_.append(tar_waves)
+
+        tar_waves_ = np.vstack(tar_waves_)[:, :, self.trim : -self.trim]
+        tar_waves = np.concatenate(tar_waves_, axis=-1)[:, : mix.shape[-1]]
+
+        source = tar_waves[:, 0:None]
+        self.logger.debug(f"{translations['tar_waves']}: {tar_waves.shape}")
+
+        if not is_match_mix:
+            source *= self.compensate
+            self.logger.debug(translations["mix_match"])
+
+        self.logger.debug(translations["mix_success"])
+        return source
+
+    def run_model(self, mix, is_match_mix=False):
+        spek = self.stft(mix.to(self.torch_device))
+        self.logger.debug(translations["stft_2"].format(shape=spek.shape))
+
+        spek[:, :, :3, :] *= 0
+
+        if is_match_mix:
+            spec_pred = spek.cpu().numpy()
+            self.logger.debug(translations["is_match_mix"])
+        else:
+            if self.enable_denoise:
+                spec_pred_neg = self.model_run(-spek)  
+                spec_pred_pos = self.model_run(spek)
+                spec_pred = (spec_pred_neg * -0.5) + (spec_pred_pos * 0.5)
+                self.logger.debug(translations["enable_denoise"])
+            else:
+                spec_pred = self.model_run(spek)
+                self.logger.debug(translations["no_denoise"])
+
+        result = self.stft.inverse(torch.tensor(spec_pred).to(self.torch_device)).cpu().detach().numpy()
+        self.logger.debug(f"{translations['stft']}: {result.shape}")
+
+        return result
+
+class STFT:
+    def __init__(self, logger, n_fft, hop_length, dim_f, device):
+        self.logger = logger
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.dim_f = dim_f
+        self.device = device
+        self.hann_window = torch.hann_window(window_length=self.n_fft, periodic=True)
+
+    def __call__(self, input_tensor):
+        is_non_standard_device = not input_tensor.device.type in ["cuda", "cpu"]
+
+        if is_non_standard_device: input_tensor = input_tensor.cpu()
+
+        stft_window = self.hann_window.to(input_tensor.device)
+
+        batch_dimensions = input_tensor.shape[:-2]
+        channel_dim, time_dim = input_tensor.shape[-2:]
+
+        reshaped_tensor = input_tensor.reshape([-1, time_dim])
+        stft_output = torch.stft(reshaped_tensor, n_fft=self.n_fft, hop_length=self.hop_length, window=stft_window, center=True, return_complex=False)
+
+        permuted_stft_output = stft_output.permute([0, 3, 1, 2])
+
+        final_output = permuted_stft_output.reshape([*batch_dimensions, channel_dim, 2, -1, permuted_stft_output.shape[-1]]).reshape([*batch_dimensions, channel_dim * 2, -1, permuted_stft_output.shape[-1]])
+
+        if is_non_standard_device: final_output = final_output.to(self.device)
+
+        return final_output[..., : self.dim_f, :]
+
+    def pad_frequency_dimension(self, input_tensor, batch_dimensions, channel_dim, freq_dim, time_dim, num_freq_bins):
+        freq_padding = torch.zeros([*batch_dimensions, channel_dim, num_freq_bins - freq_dim, time_dim]).to(input_tensor.device)
+        padded_tensor = torch.cat([input_tensor, freq_padding], -2)
+
+        return padded_tensor
+
+    def calculate_inverse_dimensions(self, input_tensor):
+        batch_dimensions = input_tensor.shape[:-3]
+        channel_dim, freq_dim, time_dim = input_tensor.shape[-3:]
+
+        num_freq_bins = self.n_fft // 2 + 1
+
+        return batch_dimensions, channel_dim, freq_dim, time_dim, num_freq_bins
+
+    def prepare_for_istft(self, padded_tensor, batch_dimensions, channel_dim, num_freq_bins, time_dim):
+        reshaped_tensor = padded_tensor.reshape([*batch_dimensions, channel_dim // 2, 2, num_freq_bins, time_dim])
+        flattened_tensor = reshaped_tensor.reshape([-1, 2, num_freq_bins, time_dim])
+        permuted_tensor = flattened_tensor.permute([0, 2, 3, 1])
+        complex_tensor = permuted_tensor[..., 0] + permuted_tensor[..., 1] * 1.0j
+
+        return complex_tensor
+
+    def inverse(self, input_tensor):
+        is_non_standard_device = not input_tensor.device.type in ["cuda", "cpu"]
+
+        if is_non_standard_device: input_tensor = input_tensor.cpu()
+
+        stft_window = self.hann_window.to(input_tensor.device)
+
+        batch_dimensions, channel_dim, freq_dim, time_dim, num_freq_bins = self.calculate_inverse_dimensions(input_tensor)
+
+        padded_tensor = self.pad_frequency_dimension(input_tensor, batch_dimensions, channel_dim, freq_dim, time_dim, num_freq_bins)
+
+        complex_tensor = self.prepare_for_istft(padded_tensor, batch_dimensions, channel_dim, num_freq_bins, time_dim)
+
+        istft_result = torch.istft(complex_tensor, n_fft=self.n_fft, hop_length=self.hop_length, window=stft_window, center=True)
+
+        final_output = istft_result.reshape([*batch_dimensions, 2, -1])
+
+        if is_non_standard_device: final_output = final_output.to(self.device)
+
+        return final_output

main/library/predictors/FCPE.py

@@ -0,0 +1,600 @@
+import os
+import math
+import torch
+import librosa
+import numpy as np
+import torch.nn as nn
+import soundfile as sf
+import torch.utils.data
+import torch.nn.functional as F
+
+from torch import nn
+from typing import Union
+from functools import partial
+from einops import rearrange, repeat
+from torchaudio.transforms import Resample
+from local_attention import LocalAttention
+from librosa.filters import mel as librosa_mel_fn
+from torch.nn.utils.parametrizations import weight_norm
+
+os.environ["LRU_CACHE_CAPACITY"] = "3"
+
+def load_wav_to_torch(full_path, target_sr=None, return_empty_on_exception=False):
+    try:
+        data, sample_rate = sf.read(full_path, always_2d=True)
+    except Exception as e:
+        print(f"{full_path}: {e}")
+        if return_empty_on_exception: return [], sample_rate or target_sr or 48000
+        else: raise
+
+    data = data[:, 0] if len(data.shape) > 1 else data
+    assert len(data) > 2
+
+    max_mag = (-np.iinfo(data.dtype).min if np.issubdtype(data.dtype, np.integer) else max(np.amax(data), -np.amin(data)))
+    max_mag = ((2**31) + 1 if max_mag > (2**15) else ((2**15) + 1 if max_mag > 1.01 else 1.0))
+    data = torch.FloatTensor(data.astype(np.float32)) / max_mag
+
+    if (torch.isinf(data) | torch.isnan(data)).any() and return_empty_on_exception: return [], sample_rate or target_sr or 48000
+    if target_sr is not None and sample_rate != target_sr:
+        data = torch.from_numpy(librosa.core.resample(data.numpy(), orig_sr=sample_rate, target_sr=target_sr))
+        sample_rate = target_sr
+
+    return data, sample_rate
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+class STFT:
+    def __init__(self, sr=22050, n_mels=80, n_fft=1024, win_size=1024, hop_length=256, fmin=20, fmax=11025, clip_val=1e-5):
+        self.target_sr = sr
+        self.n_mels = n_mels
+        self.n_fft = n_fft
+        self.win_size = win_size
+        self.hop_length = hop_length
+        self.fmin = fmin
+        self.fmax = fmax
+        self.clip_val = clip_val
+        self.mel_basis = {}
+        self.hann_window = {}
+
+    def get_mel(self, y, keyshift=0, speed=1, center=False, train=False):
+        sample_rate = self.target_sr
+        n_mels = self.n_mels
+        n_fft = self.n_fft
+        win_size = self.win_size
+        hop_length = self.hop_length
+        fmin = self.fmin
+        fmax = self.fmax
+        clip_val = self.clip_val
+
+        factor = 2 ** (keyshift / 12)
+        n_fft_new = int(np.round(n_fft * factor))
+        win_size_new = int(np.round(win_size * factor))
+        hop_length_new = int(np.round(hop_length * speed))
+
+        mel_basis = self.mel_basis if not train else {}
+        hann_window = self.hann_window if not train else {}
+        mel_basis_key = str(fmax) + "_" + str(y.device)
+
+        if mel_basis_key not in mel_basis:
+            mel = librosa_mel_fn(sr=sample_rate, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax)
+            mel_basis[mel_basis_key] = torch.from_numpy(mel).float().to(y.device)
+
+        keyshift_key = str(keyshift) + "_" + str(y.device)
+
+        if keyshift_key not in hann_window: hann_window[keyshift_key] = torch.hann_window(win_size_new).to(y.device)
+
+        pad_left = (win_size_new - hop_length_new) // 2
+        pad_right = max((win_size_new - hop_length_new + 1) // 2, win_size_new - y.size(-1) - pad_left)
+        mode = "reflect" if pad_right < y.size(-1) else "constant"
+        y = torch.nn.functional.pad(y.unsqueeze(1), (pad_left, pad_right), mode=mode)
+        y = y.squeeze(1)
+
+        spec = torch.stft(y, n_fft_new, hop_length=hop_length_new, win_length=win_size_new, window=hann_window[keyshift_key], center=center, pad_mode="reflect", normalized=False, onesided=True, return_complex=True)
+        spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + (1e-9))
+
+        if keyshift != 0:
+            size = n_fft // 2 + 1
+            resize = spec.size(1)
+            spec = (F.pad(spec, (0, 0, 0, size - resize)) if resize < size else spec[:, :size, :])
+            spec = spec * win_size / win_size_new
+        spec = torch.matmul(mel_basis[mel_basis_key], spec)
+        spec = dynamic_range_compression_torch(spec, clip_val=clip_val)
+        return spec
+
+    def __call__(self, audiopath):
+        audio, sr = load_wav_to_torch(audiopath, target_sr=self.target_sr)
+        spect = self.get_mel(audio.unsqueeze(0)).squeeze(0)
+        return spect
+
+stft = STFT()
+
+def softmax_kernel(data, *, projection_matrix, is_query, normalize_data=True, eps=1e-4, device=None):
+    b, h, *_ = data.shape
+
+    data_normalizer = (data.shape[-1] ** -0.25) if normalize_data else 1.0
+
+    ratio = projection_matrix.shape[0] ** -0.5
+    projection = repeat(projection_matrix, "j d -> b h j d", b=b, h=h)
+    projection = projection.type_as(data)
+    data_dash = torch.einsum("...id,...jd->...ij", (data_normalizer * data), projection)
+
+    diag_data = data**2
+    diag_data = torch.sum(diag_data, dim=-1)
+    diag_data = (diag_data / 2.0) * (data_normalizer**2)
+    diag_data = diag_data.unsqueeze(dim=-1)
+
+    if is_query: data_dash = ratio * (torch.exp(data_dash - diag_data - torch.max(data_dash, dim=-1, keepdim=True).values) + eps)
+    else: data_dash = ratio * (torch.exp(data_dash - diag_data + eps))
+
+    return data_dash.type_as(data)
+
+def orthogonal_matrix_chunk(cols, qr_uniform_q=False, device=None):
+    unstructured_block = torch.randn((cols, cols), device=device)
+    q, r = torch.linalg.qr(unstructured_block.cpu(), mode="reduced")
+    q, r = map(lambda t: t.to(device), (q, r))
+
+    if qr_uniform_q:
+        d = torch.diag(r, 0)
+        q *= d.sign()
+
+    return q.t()
+
+def exists(val):
+    return val is not None
+def empty(tensor):
+    return tensor.numel() == 0
+def default(val, d):
+    return val if exists(val) else d
+def cast_tuple(val):
+    return (val,) if not isinstance(val, tuple) else val
+
+class PCmer(nn.Module):
+    def __init__(self, num_layers, num_heads, dim_model, dim_keys, dim_values, residual_dropout, attention_dropout):
+        super().__init__()
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.dim_model = dim_model
+        self.dim_values = dim_values
+        self.dim_keys = dim_keys
+        self.residual_dropout = residual_dropout
+        self.attention_dropout = attention_dropout
+
+        self._layers = nn.ModuleList([_EncoderLayer(self) for _ in range(num_layers)])
+
+    def forward(self, phone, mask=None):
+        for layer in self._layers:
+            phone = layer(phone, mask)
+
+        return phone
+
+class _EncoderLayer(nn.Module):
+    def __init__(self, parent: PCmer):
+        super().__init__()
+        self.conformer = ConformerConvModule(parent.dim_model)
+        self.norm = nn.LayerNorm(parent.dim_model)
+        self.dropout = nn.Dropout(parent.residual_dropout)
+        self.attn = SelfAttention(dim=parent.dim_model, heads=parent.num_heads, causal=False)
+
+    def forward(self, phone, mask=None):
+        phone = phone + (self.attn(self.norm(phone), mask=mask))
+        phone = phone + (self.conformer(phone))
+        return phone
+
+def calc_same_padding(kernel_size):
+    pad = kernel_size // 2
+    return (pad, pad - (kernel_size + 1) % 2)
+
+class Swish(nn.Module):
+    def forward(self, x):
+        return x * x.sigmoid()
+
+class Transpose(nn.Module):
+    def __init__(self, dims):
+        super().__init__()
+        assert len(dims) == 2, "dims == 2"
+        self.dims = dims
+
+    def forward(self, x):
+        return x.transpose(*self.dims)
+
+class GLU(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        out, gate = x.chunk(2, dim=self.dim)
+        return out * gate.sigmoid()
+
+class DepthWiseConv1d(nn.Module):
+    def __init__(self, chan_in, chan_out, kernel_size, padding):
+        super().__init__()
+        self.padding = padding
+        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size, groups=chan_in)
+
+    def forward(self, x):
+        x = F.pad(x, self.padding)
+        return self.conv(x)
+
+class ConformerConvModule(nn.Module):
+    def __init__(self, dim, causal=False, expansion_factor=2, kernel_size=31, dropout=0.0):
+        super().__init__()
+
+        inner_dim = dim * expansion_factor
+        padding = calc_same_padding(kernel_size) if not causal else (kernel_size - 1, 0)
+
+        self.net = nn.Sequential(nn.LayerNorm(dim), Transpose((1, 2)), nn.Conv1d(dim, inner_dim * 2, 1), GLU(dim=1), DepthWiseConv1d(inner_dim, inner_dim, kernel_size=kernel_size, padding=padding), Swish(), nn.Conv1d(inner_dim, dim, 1), Transpose((1, 2)), nn.Dropout(dropout))
+
+    def forward(self, x):
+        return self.net(x)
+
+def linear_attention(q, k, v):
+    if v is None:
+        out = torch.einsum("...ed,...nd->...ne", k, q)
+        return out
+    else:
+        k_cumsum = k.sum(dim=-2)
+        D_inv = 1.0 / (torch.einsum("...nd,...d->...n", q, k_cumsum.type_as(q)) + 1e-8)
+        context = torch.einsum("...nd,...ne->...de", k, v)
+        out = torch.einsum("...de,...nd,...n->...ne", context, q, D_inv)
+        return out
+
+def gaussian_orthogonal_random_matrix(nb_rows, nb_columns, scaling=0, qr_uniform_q=False, device=None):
+    nb_full_blocks = int(nb_rows / nb_columns)
+    block_list = []
+
+    for _ in range(nb_full_blocks):
+        q = orthogonal_matrix_chunk(nb_columns, qr_uniform_q=qr_uniform_q, device=device)
+        block_list.append(q)
+
+    remaining_rows = nb_rows - nb_full_blocks * nb_columns
+
+    if remaining_rows > 0:
+        q = orthogonal_matrix_chunk(nb_columns, qr_uniform_q=qr_uniform_q, device=device)
+        block_list.append(q[:remaining_rows])
+
+    final_matrix = torch.cat(block_list)
+
+    if scaling == 0: multiplier = torch.randn((nb_rows, nb_columns), device=device).norm(dim=1)
+    elif scaling == 1: multiplier = math.sqrt((float(nb_columns))) * torch.ones((nb_rows,), device=device)
+    else: raise ValueError(f"Chia không hợp lệ {scaling}")
+
+    return torch.diag(multiplier) @ final_matrix
+
+class FastAttention(nn.Module):
+    def __init__(self, dim_heads, nb_features=None, ortho_scaling=0, causal=False, generalized_attention=False, kernel_fn=nn.ReLU(), qr_uniform_q=False, no_projection=False):
+        super().__init__()
+        nb_features = default(nb_features, int(dim_heads * math.log(dim_heads)))
+
+        self.dim_heads = dim_heads
+        self.nb_features = nb_features
+        self.ortho_scaling = ortho_scaling
+
+        self.create_projection = partial(gaussian_orthogonal_random_matrix, nb_rows=self.nb_features, nb_columns=dim_heads, scaling=ortho_scaling, qr_uniform_q=qr_uniform_q)
+        projection_matrix = self.create_projection()
+        self.register_buffer("projection_matrix", projection_matrix)
+
+        self.generalized_attention = generalized_attention
+        self.kernel_fn = kernel_fn
+        self.no_projection = no_projection
+        self.causal = causal
+
+    @torch.no_grad()
+    def redraw_projection_matrix(self):
+        projections = self.create_projection()
+        self.projection_matrix.copy_(projections)
+        del projections
+
+    def forward(self, q, k, v):
+        device = q.device
+
+        if self.no_projection:
+            q = q.softmax(dim=-1)
+            k = torch.exp(k) if self.causal else k.softmax(dim=-2)
+        else:
+            create_kernel = partial(softmax_kernel, projection_matrix=self.projection_matrix, device=device)
+
+            q = create_kernel(q, is_query=True)
+            k = create_kernel(k, is_query=False)
+
+        attn_fn = linear_attention if not self.causal else self.causal_linear_fn
+
+        if v is None:
+            out = attn_fn(q, k, None)
+            return out
+        else:
+            out = attn_fn(q, k, v)
+            return out
+
+class SelfAttention(nn.Module):
+    def __init__(self, dim, causal=False, heads=8, dim_head=64, local_heads=0, local_window_size=256, nb_features=None, feature_redraw_interval=1000, generalized_attention=False, kernel_fn=nn.ReLU(), qr_uniform_q=False, dropout=0.0, no_projection=False):
+        super().__init__()
+        assert dim % heads == 0
+        dim_head = default(dim_head, dim // heads)
+        inner_dim = dim_head * heads
+        self.fast_attention = FastAttention(dim_head, nb_features, causal=causal, generalized_attention=generalized_attention, kernel_fn=kernel_fn, qr_uniform_q=qr_uniform_q, no_projection=no_projection)
+        self.heads = heads
+        self.global_heads = heads - local_heads
+        self.local_attn = (LocalAttention(window_size=local_window_size, causal=causal, autopad=True, dropout=dropout, look_forward=int(not causal), rel_pos_emb_config=(dim_head, local_heads)) if local_heads > 0 else None)
+        self.to_q = nn.Linear(dim, inner_dim)
+        self.to_k = nn.Linear(dim, inner_dim)
+        self.to_v = nn.Linear(dim, inner_dim)
+        self.to_out = nn.Linear(inner_dim, dim)
+        self.dropout = nn.Dropout(dropout)
+
+    @torch.no_grad()
+    def redraw_projection_matrix(self):
+        self.fast_attention.redraw_projection_matrix()
+
+    def forward(self, x, context=None, mask=None, context_mask=None, name=None, inference=False, **kwargs):
+        _, _, _, h, gh = *x.shape, self.heads, self.global_heads
+
+        cross_attend = exists(context)
+        context = default(context, x)
+        context_mask = default(context_mask, mask) if not cross_attend else context_mask
+        q, k, v = self.to_q(x), self.to_k(context), self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v))
+        (q, lq), (k, lk), (v, lv) = map(lambda t: (t[:, :gh], t[:, gh:]), (q, k, v))
+
+        attn_outs = []
+
+        if not empty(q):
+            if exists(context_mask):
+                global_mask = context_mask[:, None, :, None]
+                v.masked_fill_(~global_mask, 0.0)
+
+            if cross_attend: pass  
+            else: out = self.fast_attention(q, k, v)
+
+            attn_outs.append(out)
+
+        if not empty(lq):
+            assert (not cross_attend), "not cross_attend"
+            out = self.local_attn(lq, lk, lv, input_mask=mask)
+            attn_outs.append(out)
+
+        out = torch.cat(attn_outs, dim=1)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        out = self.to_out(out)
+        return self.dropout(out)
+
+def l2_regularization(model, l2_alpha):
+    l2_loss = []
+    for module in model.modules():
+        if type(module) is nn.Conv2d: l2_loss.append((module.weight**2).sum() / 2.0)
+
+    return l2_alpha * sum(l2_loss)
+
+class _FCPE(nn.Module):
+    def __init__(self, input_channel=128, out_dims=360, n_layers=12, n_chans=512, use_siren=False, use_full=False, loss_mse_scale=10, loss_l2_regularization=False, loss_l2_regularization_scale=1, loss_grad1_mse=False, loss_grad1_mse_scale=1, f0_max=1975.5, f0_min=32.70, confidence=False, threshold=0.05, use_input_conv=True):
+        super().__init__()
+        if use_siren: raise ValueError("Siren not support")
+        if use_full: raise ValueError("Model full not support")
+
+        self.loss_mse_scale = loss_mse_scale if (loss_mse_scale is not None) else 10
+        self.loss_l2_regularization = (loss_l2_regularization if (loss_l2_regularization is not None) else False)
+        self.loss_l2_regularization_scale = (loss_l2_regularization_scale if (loss_l2_regularization_scale is not None) else 1)
+        self.loss_grad1_mse = loss_grad1_mse if (loss_grad1_mse is not None) else False
+        self.loss_grad1_mse_scale = (loss_grad1_mse_scale if (loss_grad1_mse_scale is not None) else 1)
+        self.f0_max = f0_max if (f0_max is not None) else 1975.5
+        self.f0_min = f0_min if (f0_min is not None) else 32.70
+        self.confidence = confidence if (confidence is not None) else False
+        self.threshold = threshold if (threshold is not None) else 0.05
+        self.use_input_conv = use_input_conv if (use_input_conv is not None) else True
+        self.cent_table_b = torch.Tensor(np.linspace(self.f0_to_cent(torch.Tensor([f0_min]))[0], self.f0_to_cent(torch.Tensor([f0_max]))[0], out_dims))
+        self.register_buffer("cent_table", self.cent_table_b)
+
+        _leaky = nn.LeakyReLU()
+
+        self.stack = nn.Sequential(nn.Conv1d(input_channel, n_chans, 3, 1, 1), nn.GroupNorm(4, n_chans), _leaky, nn.Conv1d(n_chans, n_chans, 3, 1, 1))
+        self.decoder = PCmer(num_layers=n_layers, num_heads=8, dim_model=n_chans, dim_keys=n_chans, dim_values=n_chans, residual_dropout=0.1, attention_dropout=0.1)
+        self.norm = nn.LayerNorm(n_chans)
+        self.n_out = out_dims
+        self.dense_out = weight_norm(nn.Linear(n_chans, self.n_out))
+
+    def forward(self, mel, infer=True, gt_f0=None, return_hz_f0=False, cdecoder="local_argmax"):
+        if cdecoder == "argmax": self.cdecoder = self.cents_decoder
+        elif cdecoder == "local_argmax": self.cdecoder = self.cents_local_decoder
+
+        x = (self.stack(mel.transpose(1, 2)).transpose(1, 2) if self.use_input_conv else mel)
+        x = self.decoder(x)
+        x = self.norm(x)
+        x = self.dense_out(x)
+        x = torch.sigmoid(x)
+
+        if not infer:
+            gt_cent_f0 = self.f0_to_cent(gt_f0)
+            gt_cent_f0 = self.gaussian_blurred_cent(gt_cent_f0)
+            loss_all = self.loss_mse_scale * F.binary_cross_entropy(x, gt_cent_f0)
+
+            if self.loss_l2_regularization: loss_all = loss_all + l2_regularization(model=self, l2_alpha=self.loss_l2_regularization_scale)
+
+            x = loss_all
+
+        if infer:
+            x = self.cdecoder(x)
+            x = self.cent_to_f0(x)
+            x = (1 + x / 700).log() if not return_hz_f0 else x
+
+        return x
+
+    def cents_decoder(self, y, mask=True):
+        B, N, _ = y.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        rtn = torch.sum(ci * y, dim=-1, keepdim=True) / torch.sum(y, dim=-1, keepdim=True)
+        if mask:
+            confident = torch.max(y, dim=-1, keepdim=True)[0]
+            confident_mask = torch.ones_like(confident)
+            confident_mask[confident <= self.threshold] = float("-INF")
+            rtn = rtn * confident_mask
+
+        return (rtn, confident) if self.confidence else rtn
+
+    def cents_local_decoder(self, y, mask=True):
+        B, N, _ = y.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        confident, max_index = torch.max(y, dim=-1, keepdim=True)
+        local_argmax_index = torch.arange(0, 9).to(max_index.device) + (max_index - 4)
+        local_argmax_index = torch.clamp(local_argmax_index, 0, self.n_out - 1)
+        ci_l = torch.gather(ci, -1, local_argmax_index)
+        y_l = torch.gather(y, -1, local_argmax_index)
+        rtn = torch.sum(ci_l * y_l, dim=-1, keepdim=True) / torch.sum(y_l, dim=-1, keepdim=True)
+
+        if mask:
+            confident_mask = torch.ones_like(confident)
+            confident_mask[confident <= self.threshold] = float("-INF")
+            rtn = rtn * confident_mask
+
+        return (rtn, confident) if self.confidence else rtn
+
+    def cent_to_f0(self, cent):
+        return 10.0 * 2 ** (cent / 1200.0)
+
+    def f0_to_cent(self, f0):
+        return 1200.0 * torch.log2(f0 / 10.0)
+
+    def gaussian_blurred_cent(self, cents):
+        mask = (cents > 0.1) & (cents < (1200.0 * np.log2(self.f0_max / 10.0)))
+        B, N, _ = cents.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        return torch.exp(-torch.square(ci - cents) / 1250) * mask.float()
+
+class FCPEInfer:
+    def __init__(self, model_path, device=None, dtype=torch.float32):
+        if device is None: device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        self.device = device
+        ckpt = torch.load(model_path, map_location=torch.device(self.device))
+        self.args = DotDict(ckpt["config"])
+        self.dtype = dtype
+        model = _FCPE(input_channel=self.args.model.input_channel, out_dims=self.args.model.out_dims, n_layers=self.args.model.n_layers, n_chans=self.args.model.n_chans, use_siren=self.args.model.use_siren, use_full=self.args.model.use_full, loss_mse_scale=self.args.loss.loss_mse_scale, loss_l2_regularization=self.args.loss.loss_l2_regularization, loss_l2_regularization_scale=self.args.loss.loss_l2_regularization_scale, loss_grad1_mse=self.args.loss.loss_grad1_mse, loss_grad1_mse_scale=self.args.loss.loss_grad1_mse_scale, f0_max=self.args.model.f0_max, f0_min=self.args.model.f0_min, confidence=self.args.model.confidence)
+        model.to(self.device).to(self.dtype)
+        model.load_state_dict(ckpt["model"])
+        model.eval()
+        self.model = model
+        self.wav2mel = Wav2Mel(self.args, dtype=self.dtype, device=self.device)
+
+    @torch.no_grad()
+    def __call__(self, audio, sr, threshold=0.05):
+        self.model.threshold = threshold
+        audio = audio[None, :]
+        mel = self.wav2mel(audio=audio, sample_rate=sr).to(self.dtype)
+        f0 = self.model(mel=mel, infer=True, return_hz_f0=True)
+        return f0
+
+class Wav2Mel:
+    def __init__(self, args, device=None, dtype=torch.float32):
+        self.sample_rate = args.mel.sampling_rate
+        self.hop_size = args.mel.hop_size
+
+        if device is None: device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        self.device = device
+        self.dtype = dtype
+        self.stft = STFT(args.mel.sampling_rate, args.mel.num_mels, args.mel.n_fft, args.mel.win_size, args.mel.hop_size, args.mel.fmin, args.mel.fmax)
+        self.resample_kernel = {}
+
+    def extract_nvstft(self, audio, keyshift=0, train=False):
+        mel = self.stft.get_mel(audio, keyshift=keyshift, train=train).transpose(1, 2)
+        return mel
+
+    def extract_mel(self, audio, sample_rate, keyshift=0, train=False):
+        audio = audio.to(self.dtype).to(self.device)
+
+        if sample_rate == self.sample_rate: audio_res = audio
+        else:
+            key_str = str(sample_rate)
+
+            if key_str not in self.resample_kernel: self.resample_kernel[key_str] = Resample(sample_rate, self.sample_rate, lowpass_filter_width=128)
+
+            self.resample_kernel[key_str] = (self.resample_kernel[key_str].to(self.dtype).to(self.device))
+            audio_res = self.resample_kernel[key_str](audio)
+
+        mel = self.extract_nvstft(audio_res, keyshift=keyshift, train=train) 
+        n_frames = int(audio.shape[1] // self.hop_size) + 1
+        mel = (torch.cat((mel, mel[:, -1:, :]), 1) if n_frames > int(mel.shape[1]) else mel)
+        mel = mel[:, :n_frames, :] if n_frames < int(mel.shape[1]) else mel
+        return mel
+
+    def __call__(self, audio, sample_rate, keyshift=0, train=False):
+        return self.extract_mel(audio, sample_rate, keyshift=keyshift, train=train)
+
+class DotDict(dict):
+    def __getattr__(*args):
+        val = dict.get(*args)
+        return DotDict(val) if type(val) is dict else val
+
+    __setattr__ = dict.__setitem__
+    __delattr__ = dict.__delitem__
+
+class F0Predictor(object):
+    def compute_f0(self, wav, p_len): pass
+    def compute_f0_uv(self, wav, p_len): pass
+
+class FCPE(F0Predictor):
+    def __init__(self, model_path, hop_length=512, f0_min=50, f0_max=1100, dtype=torch.float32, device=None, sample_rate=44100, threshold=0.05):
+        self.fcpe = FCPEInfer(model_path, device=device, dtype=dtype)
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        self.threshold = threshold
+        self.sample_rate = sample_rate
+        self.dtype = dtype
+        self.name = "fcpe"
+
+    def repeat_expand(self, content: Union[torch.Tensor, np.ndarray], target_len, mode = "nearest"):
+        ndim = content.ndim
+        content = (content[None, None] if ndim == 1 else content[None] if ndim == 2 else content)
+        assert content.ndim == 3
+        is_np = isinstance(content, np.ndarray)
+        content = torch.from_numpy(content) if is_np else content
+        results = torch.nn.functional.interpolate(content, size=target_len, mode=mode)
+        results = results.numpy() if is_np else results
+        return results[0, 0] if ndim == 1 else results[0] if ndim == 2 else results
+
+    def post_process(self, x, sample_rate, f0, pad_to):
+        f0 = (torch.from_numpy(f0).float().to(x.device) if isinstance(f0, np.ndarray) else f0)
+        f0 = self.repeat_expand(f0, pad_to) if pad_to is not None else f0
+
+        vuv_vector = torch.zeros_like(f0)
+        vuv_vector[f0 > 0.0] = 1.0
+        vuv_vector[f0 <= 0.0] = 0.0
+
+        nzindex = torch.nonzero(f0).squeeze()
+        f0 = torch.index_select(f0, dim=0, index=nzindex).cpu().numpy()
+        time_org = self.hop_length / sample_rate * nzindex.cpu().numpy()
+        time_frame = np.arange(pad_to) * self.hop_length / sample_rate
+
+        vuv_vector = F.interpolate(vuv_vector[None, None, :], size=pad_to)[0][0]
+
+        if f0.shape[0] <= 0: return np.zeros(pad_to), vuv_vector.cpu().numpy()
+        if f0.shape[0] == 1: return np.ones(pad_to) * f0[0], vuv_vector.cpu().numpy()
+
+        f0 = np.interp(time_frame, time_org, f0, left=f0[0], right=f0[-1])
+        return f0, vuv_vector.cpu().numpy()
+
+    def compute_f0(self, wav, p_len=None):
+        x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
+        p_len = x.shape[0] // self.hop_length if p_len is None else p_len
+        f0 = self.fcpe(x, sr=self.sample_rate, threshold=self.threshold)[0, :, 0]
+
+        if torch.all(f0 == 0): return f0.cpu().numpy() if p_len is None else np.zeros(p_len), (f0.cpu().numpy() if p_len is None else np.zeros(p_len))
+
+        return self.post_process(x, self.sample_rate, f0, p_len)[0]
+    def compute_f0_uv(self, wav, p_len=None):
+        x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
+        p_len = x.shape[0] // self.hop_length if p_len is None else p_len
+        f0 = self.fcpe(x, sr=self.sample_rate, threshold=self.threshold)[0, :, 0]
+
+        if torch.all(f0 == 0): return f0.cpu().numpy() if p_len is None else np.zeros(p_len), (f0.cpu().numpy() if p_len is None else np.zeros(p_len))
+        
+        return self.post_process(x, self.sample_rate, f0, p_len)

main/library/predictors/RMVPE.py

@@ -0,0 +1,270 @@
+import torch
+import numpy as np
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import List
+from librosa.filters import mel
+
+N_MELS = 128
+N_CLASS = 360
+
+
+class ConvBlockRes(nn.Module):
+    def __init__(self, in_channels, out_channels, momentum=0.01):
+        super(ConvBlockRes, self).__init__()
+        self.conv = nn.Sequential(nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False), nn.BatchNorm2d(out_channels, momentum=momentum), nn.ReLU(), nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False), nn.BatchNorm2d(out_channels, momentum=momentum), nn.ReLU())
+
+        if in_channels != out_channels:
+            self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
+            self.is_shortcut = True
+        else: self.is_shortcut = False
+
+    def forward(self, x):
+        if self.is_shortcut: return self.conv(x) + self.shortcut(x)
+        else: return self.conv(x) + x
+
+class ResEncoderBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01):
+        super(ResEncoderBlock, self).__init__()
+        self.n_blocks = n_blocks
+        self.conv = nn.ModuleList()
+        self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
+
+        for _ in range(n_blocks - 1):
+            self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
+
+        self.kernel_size = kernel_size
+
+        if self.kernel_size is not None:
+            self.pool = nn.AvgPool2d(kernel_size=kernel_size)
+
+    def forward(self, x):
+        for i in range(self.n_blocks):
+            x = self.conv[i](x)
+
+        if self.kernel_size is not None: return x, self.pool(x)
+        else: return x
+
+class Encoder(nn.Module):
+    def __init__(self, in_channels, in_size, n_encoders, kernel_size, n_blocks, out_channels=16, momentum=0.01):
+        super(Encoder, self).__init__()
+        self.n_encoders = n_encoders
+        self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
+        self.layers = nn.ModuleList()
+        self.latent_channels = []
+
+        for i in range(self.n_encoders):
+            self.layers.append(ResEncoderBlock(in_channels, out_channels, kernel_size, n_blocks, momentum=momentum))
+            self.latent_channels.append([out_channels, in_size])
+            in_channels = out_channels
+            out_channels *= 2
+            in_size //= 2
+            
+        self.out_size = in_size
+        self.out_channel = out_channels
+
+    def forward(self, x: torch.Tensor):
+        concat_tensors: List[torch.Tensor] = []
+        x = self.bn(x)
+
+        for i in range(self.n_encoders):
+            t, x = self.layers[i](x)
+            concat_tensors.append(t)
+
+        return x, concat_tensors
+
+class Intermediate(nn.Module):
+    def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
+        super(Intermediate, self).__init__()
+        self.n_inters = n_inters
+        self.layers = nn.ModuleList()
+        self.layers.append(ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum))
+
+        for _ in range(self.n_inters - 1):
+            self.layers.append(ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum))
+
+    def forward(self, x):
+        for i in range(self.n_inters):
+            x = self.layers[i](x)
+        return x
+
+class ResDecoderBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
+        super(ResDecoderBlock, self).__init__()
+        out_padding = (0, 1) if stride == (1, 2) else (1, 1)
+        self.n_blocks = n_blocks
+        self.conv1 = nn.Sequential(nn.ConvTranspose2d(in_channels=in_channels, out_channels=out_channels, kernel_size=(3, 3), stride=stride, padding=(1, 1), output_padding=out_padding, bias=False), nn.BatchNorm2d(out_channels, momentum=momentum), nn.ReLU())
+        self.conv2 = nn.ModuleList()
+        self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
+        for _ in range(n_blocks - 1):
+            self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
+
+    def forward(self, x, concat_tensor):
+        x = self.conv1(x)
+        x = torch.cat((x, concat_tensor), dim=1)
+        for i in range(self.n_blocks):
+            x = self.conv2[i](x)
+
+        return x
+
+class Decoder(nn.Module):
+    def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
+        super(Decoder, self).__init__()
+        self.layers = nn.ModuleList()
+        self.n_decoders = n_decoders
+
+        for _ in range(self.n_decoders):
+            out_channels = in_channels // 2
+            self.layers.append(ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum))
+            in_channels = out_channels
+
+    def forward(self, x, concat_tensors):
+        for i in range(self.n_decoders):
+            x = self.layers[i](x, concat_tensors[-1 - i])
+
+        return x
+
+class DeepUnet(nn.Module):
+    def __init__(self, kernel_size, n_blocks, en_de_layers=5, inter_layers=4, in_channels=1, en_out_channels=16):
+        super(DeepUnet, self).__init__()
+        self.encoder = Encoder(in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels)
+        self.intermediate = Intermediate(self.encoder.out_channel // 2, self.encoder.out_channel, inter_layers, n_blocks)
+        self.decoder = Decoder(self.encoder.out_channel, en_de_layers, kernel_size, n_blocks)
+
+    def forward(self, x):
+        x, concat_tensors = self.encoder(x)
+        x = self.intermediate(x)
+        x = self.decoder(x, concat_tensors)
+        return x
+
+class E2E(nn.Module):
+    def __init__(self, n_blocks, n_gru, kernel_size, en_de_layers=5, inter_layers=4, in_channels=1, en_out_channels=16):
+        super(E2E, self).__init__()
+        self.unet = DeepUnet(kernel_size, n_blocks, en_de_layers, inter_layers, in_channels, en_out_channels)
+        self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
+
+        if n_gru: self.fc = nn.Sequential(BiGRU(3 * 128, 256, n_gru), nn.Linear(512, N_CLASS), nn.Dropout(0.25), nn.Sigmoid())
+        else: self.fc = nn.Sequential(nn.Linear(3 * N_MELS, N_CLASS), nn.Dropout(0.25), nn.Sigmoid())
+
+    def forward(self, mel):
+        mel = mel.transpose(-1, -2).unsqueeze(1)
+        x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
+        x = self.fc(x)
+        return x
+
+class MelSpectrogram(torch.nn.Module):
+    def __init__(self, is_half, n_mel_channels, sample_rate, win_length, hop_length, n_fft=None, mel_fmin=0, mel_fmax=None, clamp=1e-5):
+        super().__init__()
+        n_fft = win_length if n_fft is None else n_fft
+        self.hann_window = {}
+        mel_basis = mel(sr=sample_rate, n_fft=n_fft, n_mels=n_mel_channels, fmin=mel_fmin, fmax=mel_fmax, htk=True)
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+        self.n_fft = win_length if n_fft is None else n_fft
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.sample_rate = sample_rate
+        self.n_mel_channels = n_mel_channels
+        self.clamp = clamp
+        self.is_half = is_half
+
+    def forward(self, audio, keyshift=0, speed=1, center=True):
+        factor = 2 ** (keyshift / 12)
+        n_fft_new = int(np.round(self.n_fft * factor))
+        win_length_new = int(np.round(self.win_length * factor))
+        hop_length_new = int(np.round(self.hop_length * speed))
+        keyshift_key = str(keyshift) + "_" + str(audio.device)
+
+        if keyshift_key not in self.hann_window: self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(audio.device)
+
+        fft = torch.stft(audio, n_fft=n_fft_new, hop_length=hop_length_new, win_length=win_length_new, window=self.hann_window[keyshift_key], center=center, return_complex=True)
+        magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
+
+        if keyshift != 0:
+            size = self.n_fft // 2 + 1
+            resize = magnitude.size(1)
+
+            if resize < size: magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
+
+            magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
+
+        mel_output = torch.matmul(self.mel_basis, magnitude)
+
+        if self.is_half: mel_output = mel_output.half()
+
+        log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
+        return log_mel_spec
+
+class RMVPE:
+    def __init__(self, model_path, is_half, device=None):
+        self.resample_kernel = {}
+        model = E2E(4, 1, (2, 2))
+        ckpt = torch.load(model_path, map_location="cpu")
+        model.load_state_dict(ckpt)
+        model.eval()
+
+        if is_half: model = model.half()
+
+        self.model = model
+        self.resample_kernel = {}
+        self.is_half = is_half
+        self.device = device
+        self.mel_extractor = MelSpectrogram(is_half, N_MELS, 16000, 1024, 160, None, 30, 8000).to(device)
+        self.model = self.model.to(device)
+        cents_mapping = 20 * np.arange(N_CLASS) + 1997.3794084376191
+        self.cents_mapping = np.pad(cents_mapping, (4, 4))
+
+    def mel2hidden(self, mel):
+        with torch.no_grad():
+            n_frames = mel.shape[-1]
+            mel = F.pad(mel, (0, 32 * ((n_frames - 1) // 32 + 1) - n_frames), mode="reflect")
+            hidden = self.model(mel)
+            return hidden[:, :n_frames]
+
+    def decode(self, hidden, thred=0.03):
+        cents_pred = self.to_local_average_cents(hidden, thred=thred)
+        f0 = 10 * (2 ** (cents_pred / 1200))
+        f0[f0 == 10] = 0
+        return f0
+
+    def infer_from_audio(self, audio, thred=0.03):
+        audio = torch.from_numpy(audio).float().to(self.device).unsqueeze(0)
+        mel = self.mel_extractor(audio, center=True)
+        hidden = self.mel2hidden(mel)
+        hidden = hidden.squeeze(0).cpu().numpy()
+
+        if self.is_half: hidden = hidden.astype("float32")
+
+        f0 = self.decode(hidden, thred=thred)
+        return f0
+
+    def to_local_average_cents(self, salience, thred=0.05):
+        center = np.argmax(salience, axis=1)
+        salience = np.pad(salience, ((0, 0), (4, 4)))
+        center += 4
+        todo_salience = []
+        todo_cents_mapping = []
+        starts = center - 4
+        ends = center + 5
+
+        for idx in range(salience.shape[0]):
+            todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
+            todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
+
+        todo_salience = np.array(todo_salience)
+        todo_cents_mapping = np.array(todo_cents_mapping)
+        product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
+        weight_sum = np.sum(todo_salience, 1)
+        devided = product_sum / weight_sum
+        maxx = np.max(salience, axis=1)
+        devided[maxx <= thred] = 0
+        return devided
+
+class BiGRU(nn.Module):
+    def __init__(self, input_features, hidden_features, num_layers):
+        super(BiGRU, self).__init__()
+        self.gru = nn.GRU(input_features, hidden_features, num_layers=num_layers, batch_first=True, bidirectional=True)
+
+    def forward(self, x):
+        return self.gru(x)[0]