app.py

import os
import argparse
import gradio as gr
import yaml

from modules.utils.paths import (FASTER_WHISPER_MODELS_DIR, DIARIZATION_MODELS_DIR, OUTPUT_DIR, WHISPER_MODELS_DIR,
                                 INSANELY_FAST_WHISPER_MODELS_DIR, NLLB_MODELS_DIR, DEFAULT_PARAMETERS_CONFIG_PATH,
                                 UVR_MODELS_DIR)
from modules.utils.files_manager import load_yaml
from modules.whisper.whisper_factory import WhisperFactory
from modules.whisper.faster_whisper_inference import FasterWhisperInference
from modules.whisper.insanely_fast_whisper_inference import InsanelyFastWhisperInference
from modules.translation.nllb_inference import NLLBInference
from modules.ui.htmls import *
from modules.utils.cli_manager import str2bool
from modules.utils.youtube_manager import get_ytmetas
from modules.translation.deepl_api import DeepLAPI
from modules.whisper.whisper_parameter import *

### Device info ###
import torch
import torchaudio
import torch.cuda as cuda
import platform
from transformers import __version__ as transformers_version

device = "cuda" if torch.cuda.is_available() else "cpu"
num_gpus = cuda.device_count() if torch.cuda.is_available() else 0
cuda_version = torch.version.cuda if torch.cuda.is_available() else "N/A"
cudnn_version = torch.backends.cudnn.version() if torch.cuda.is_available() else "N/A"
os_info = platform.system() + " " + platform.release() + " " + platform.machine()

# Get the available VRAM for each GPU (if available)
vram_info = []
if torch.cuda.is_available():
    for i in range(cuda.device_count()):
        gpu_properties = cuda.get_device_properties(i)
        vram_info.append(f"**GPU {i}: {gpu_properties.total_memory / 1024**3:.2f} GB**")

pytorch_version = torch.__version__
torchaudio_version = torchaudio.__version__ if 'torchaudio' in dir() else "N/A"

device_info = f"""Running on: **{device}**

    Number of GPUs available: **{num_gpus}**

    CUDA version: **{cuda_version}**

    CuDNN version: **{cudnn_version}**

    PyTorch version: **{pytorch_version}**

    Torchaudio version: **{torchaudio_version}**

    Transformers version: **{transformers_version}**

    Operating system: **{os_info}**

    Available VRAM: 
    \t {', '.join(vram_info) if vram_info else '**N/A**'}
"""
### End Device info ###

class App:
    def __init__(self, args):
        self.args = args
        #self.app = gr.Blocks(css=CSS, theme=self.args.theme, delete_cache=(60, 3600))
        self.app = gr.Blocks(css=CSS,theme=gr.themes.Ocean(), title="Whisper - Automatic speech recognition", delete_cache=(60, 3600))
        self.whisper_inf = WhisperFactory.create_whisper_inference(
            whisper_type=self.args.whisper_type,
            whisper_model_dir=self.args.whisper_model_dir,
            faster_whisper_model_dir=self.args.faster_whisper_model_dir,
            insanely_fast_whisper_model_dir=self.args.insanely_fast_whisper_model_dir,
            uvr_model_dir=self.args.uvr_model_dir,
            output_dir=self.args.output_dir,
        )
        self.nllb_inf = NLLBInference(
            model_dir=self.args.nllb_model_dir,
            output_dir=os.path.join(self.args.output_dir, "translations")
        )
        self.deepl_api = DeepLAPI(
            output_dir=os.path.join(self.args.output_dir, "translations")
        )
        self.default_params = load_yaml(DEFAULT_PARAMETERS_CONFIG_PATH)
        print(f"Use \"{self.args.whisper_type}\" implementation")
        print(f"Device \"{self.whisper_inf.device}\" is detected")

    def create_whisper_parameters(self):

        whisper_params = self.default_params["whisper"]
        diarization_params = self.default_params["diarization"]
        vad_params = self.default_params["vad"]
        uvr_params = self.default_params["bgm_separation"]

        #Translation integration
        translation_params = self.default_params["translation"]
        nllb_params = translation_params["nllb"]

        with gr.Row():
            with gr.Column(scale=4):
                with gr.Row():
                    dd_model = gr.Dropdown(choices=self.whisper_inf.available_models, value=whisper_params["model_size"],label="Model", info="Larger models increase transcription quality, but reduce performance", interactive=True)
                    dd_lang = gr.Dropdown(choices=["Automatic Detection"] + self.whisper_inf.available_langs,value=whisper_params["lang"], label="Language", info="If the language is known upfront, always set it manually", interactive=True)
                    dd_file_format = gr.Dropdown(choices=["TXT","SRT"], value="TXT", label="Output format", info="Output preview format", interactive=True, visible=False)
                with gr.Row():
                    dd_translate_model = gr.Dropdown(choices=self.nllb_inf.available_models, value=nllb_params["model_size"],label="Model", info="Model used for translation", interactive=True)
                    dd_target_lang = gr.Dropdown(choices=["English","Dutch","French","German"], value=nllb_params["target_lang"],label="Language", info="Language used for output translation", interactive=True)
            with gr.Column(scale=1):
                with gr.Row():
                    cb_timestamp = gr.Checkbox(value=whisper_params["add_timestamp"], label="Add timestamp to output file",interactive=True)
                with gr.Row():
                    cb_translate = gr.Checkbox(value=whisper_params["is_translate"], label="Translate to English", info="Translate using OpenAI Whisper's built-in module",interactive=True)
                #with gr.Row():
                    cb_translate_output = gr.Checkbox(value=translation_params["translate_output"], label="Translate to selected language", info="Translate using Facebook's NLLB",interactive=True)
                
        with gr.Accordion("Speaker diarization", open=False, visible=True):
            cb_diarize = gr.Checkbox(value=diarization_params["is_diarize"],label="Use diarization",interactive=True)
            tb_hf_token = gr.Text(label="Token", value=diarization_params["hf_token"],info="An access token is required to use diarization & can be created [here](https://hf.co/settings/tokens). If not done yet for your account, you need to accept the terms & conditions of [diarization](https://huggingface.co/pyannote/speaker-diarization-3.1) & [segmentation](https://huggingface.co/pyannote/segmentation-3.0)")
            dd_diarization_device = gr.Dropdown(label="Device",
                                                choices=self.whisper_inf.diarizer.get_available_device(),
                                                value=self.whisper_inf.diarizer.get_device(),
                                                interactive=True, visible=False)

        with gr.Accordion("Preprocessing options", open=False, visible=True):
        
            gr.Markdown("<i><b>Note: ⚠ Experimental features (Use with caution)</b></i>")

            with gr.Accordion("Voice Detection Filter", open=False, visible=True):
                cb_vad_filter = gr.Checkbox(label="Enable Silero VAD Filter", value=vad_params["vad_filter"],
                                            info="Enable to transcribe only detected voice parts",
                                            interactive=True)
                sd_threshold = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label="Speech Threshold",
                                         value=vad_params["threshold"],
                                         info="Lower it to be more sensitive to small sounds")
                nb_min_speech_duration_ms = gr.Number(label="Minimum Speech Duration (ms)", precision=0,
                                                      value=vad_params["min_speech_duration_ms"],
                                                      info="Final speech chunks shorter than this time are thrown out")
                nb_max_speech_duration_s = gr.Number(label="Maximum Speech Duration (s)",
                                                     value=vad_params["max_speech_duration_s"],
                                                     info="Maximum duration of speech chunks in seconds")
                nb_min_silence_duration_ms = gr.Number(label="Minimum Silence Duration (ms)", precision=0,
                                                       value=vad_params["min_silence_duration_ms"],
                                                       info="In the end of each speech chunk wait for this time"
                                                            " before separating it")
                nb_speech_pad_ms = gr.Number(label="Speech Padding (ms)", precision=0, value=vad_params["speech_pad_ms"],
                                             info="Final speech chunks are padded by this time each side")
    
            with gr.Accordion("Background Music Remover Filter", open=False):
                cb_bgm_separation = gr.Checkbox(label="Enable Background Music Remover Filter", value=uvr_params["is_separate_bgm"],
                                                info="Enable to remove background music by submodel before transcribing",
                                                interactive=True)
                dd_uvr_device = gr.Dropdown(label="Device",
                                                value=self.whisper_inf.music_separator.device,
                                                choices=self.whisper_inf.music_separator.available_devices,
                                                interactive=True, visible=False)
                dd_uvr_model_size = gr.Dropdown(label="Model", value=uvr_params["model_size"],
                                                choices=self.whisper_inf.music_separator.available_models,
                                                interactive=True)
                nb_uvr_segment_size = gr.Number(label="Segment Size", value=uvr_params["segment_size"], precision=0,
                                                interactive=True, visible=False)
                cb_uvr_save_file = gr.Checkbox(label="Save separated files to output", value=uvr_params["save_file"],
                                               interactive=True, visible=False)
                cb_uvr_enable_offload = gr.Checkbox(label="Offload sub model after removing background music",value=uvr_params["enable_offload"],
                                                interactive=True, visible=False)
        
        with gr.Accordion("Advanced processing options", open=False, visible=False):
            nb_beam_size = gr.Number(label="Beam Size", value=whisper_params["beam_size"], precision=0, interactive=True,
                                     info="Beam size to use for decoding.")
            nb_log_prob_threshold = gr.Number(label="Log Probability Threshold", value=whisper_params["log_prob_threshold"], interactive=True,
                                              info="If the average log probability over sampled tokens is below this value, treat as failed.")
            nb_no_speech_threshold = gr.Number(label="No Speech Threshold", value=whisper_params["no_speech_threshold"], interactive=True,
                                               info="If the no speech probability is higher than this value AND the average log probability over sampled tokens is below 'Log Prob Threshold', consider the segment as silent.")
            dd_compute_type = gr.Dropdown(label="Compute Type", choices=self.whisper_inf.available_compute_types,
                                          value=self.whisper_inf.current_compute_type, interactive=True,
                                          allow_custom_value=True,
                                          info="Select the type of computation to perform.")
            nb_best_of = gr.Number(label="Best Of", value=whisper_params["best_of"], interactive=True,
                                   info="Number of candidates when sampling with non-zero temperature.")
            nb_patience = gr.Number(label="Patience", value=whisper_params["patience"], interactive=True,
                                    info="Beam search patience factor.")
            cb_condition_on_previous_text = gr.Checkbox(label="Condition On Previous Text", value=whisper_params["condition_on_previous_text"],
                                                        interactive=True,
                                                        info="Condition on previous text during decoding.")
            sld_prompt_reset_on_temperature = gr.Slider(label="Prompt Reset On Temperature", value=whisper_params["prompt_reset_on_temperature"],
                                                        minimum=0, maximum=1, step=0.01, interactive=True,
                                                        info="Resets prompt if temperature is above this value."
                                                             " Arg has effect only if 'Condition On Previous Text' is True.")
            tb_initial_prompt = gr.Textbox(label="Initial Prompt", value=None, interactive=True,
                                           info="Initial prompt to use for decoding.")
            sd_temperature = gr.Slider(label="Temperature", value=whisper_params["temperature"], minimum=0.0,
                                       step=0.01, maximum=1.0, interactive=True,
                                       info="Temperature for sampling. It can be a tuple of temperatures, which will be successively used upon failures according to either `Compression Ratio Threshold` or `Log Prob Threshold`.")
            nb_compression_ratio_threshold = gr.Number(label="Compression Ratio Threshold", value=whisper_params["compression_ratio_threshold"],
                                                       interactive=True,
                                                       info="If the gzip compression ratio is above this value, treat as failed.")
            nb_chunk_length = gr.Number(label="Chunk Length (s)", value=lambda: whisper_params["chunk_length"],
                                        precision=0,
                                        info="The length of audio segments. If it is not None, it will overwrite the default chunk_length of the FeatureExtractor.")
            with gr.Group(visible=isinstance(self.whisper_inf, FasterWhisperInference)):
                nb_length_penalty = gr.Number(label="Length Penalty", value=whisper_params["length_penalty"],
                                              info="Exponential length penalty constant.")
                nb_repetition_penalty = gr.Number(label="Repetition Penalty", value=whisper_params["repetition_penalty"],
                                                  info="Penalty applied to the score of previously generated tokens (set > 1 to penalize).")
                nb_no_repeat_ngram_size = gr.Number(label="No Repeat N-gram Size", value=whisper_params["no_repeat_ngram_size"],
                                                    precision=0,
                                                    info="Prevent repetitions of n-grams with this size (set 0 to disable).")
                tb_prefix = gr.Textbox(label="Prefix", value=lambda: whisper_params["prefix"],
                                       info="Optional text to provide as a prefix for the first window.")
                cb_suppress_blank = gr.Checkbox(label="Suppress Blank", value=whisper_params["suppress_blank"],
                                                info="Suppress blank outputs at the beginning of the sampling.")
                tb_suppress_tokens = gr.Textbox(label="Suppress Tokens", value=whisper_params["suppress_tokens"],
                                                info="List of token IDs to suppress. -1 will suppress a default set of symbols as defined in the model config.json file.")
                nb_max_initial_timestamp = gr.Number(label="Max Initial Timestamp", value=whisper_params["max_initial_timestamp"],
                                                     info="The initial timestamp cannot be later than this.")
                cb_word_timestamps = gr.Checkbox(label="Word Timestamps", value=whisper_params["word_timestamps"],
                                                 info="Extract word-level timestamps using the cross-attention pattern and dynamic time warping, and include the timestamps for each word in each segment.")
                tb_prepend_punctuations = gr.Textbox(label="Prepend Punctuations", value=whisper_params["prepend_punctuations"],
                                                     info="If 'Word Timestamps' is True, merge these punctuation symbols with the next word.")
                tb_append_punctuations = gr.Textbox(label="Append Punctuations", value=whisper_params["append_punctuations"],
                                                    info="If 'Word Timestamps' is True, merge these punctuation symbols with the previous word.")
                nb_max_new_tokens = gr.Number(label="Max New Tokens", value=lambda: whisper_params["max_new_tokens"],
                                              precision=0,
                                              info="Maximum number of new tokens to generate per-chunk. If not set, the maximum will be set by the default max_length.")
                nb_hallucination_silence_threshold = gr.Number(label="Hallucination Silence Threshold (sec)",
                                                               value=lambda: whisper_params["hallucination_silence_threshold"],
                                                               info="When 'Word Timestamps' is True, skip silent periods longer than this threshold (in seconds) when a possible hallucination is detected.")
                tb_hotwords = gr.Textbox(label="Hotwords", value=lambda: whisper_params["hotwords"],
                                         info="Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None.")
                nb_language_detection_threshold = gr.Number(label="Language Detection Threshold", value=lambda: whisper_params["language_detection_threshold"],
                                                            info="If the maximum probability of the language tokens is higher than this value, the language is detected.")
                nb_language_detection_segments = gr.Number(label="Language Detection Segments", value=lambda: whisper_params["language_detection_segments"],
                                                           precision=0,
                                                           info="Number of segments to consider for the language detection.")
            with gr.Group(visible=isinstance(self.whisper_inf, InsanelyFastWhisperInference)):
                nb_batch_size = gr.Number(label="Batch Size", value=whisper_params["batch_size"], precision=0)
    

        #dd_model.change(fn=self.on_change_models, inputs=[dd_model], outputs=[cb_translate])

        return (
            WhisperParameters(
                model_size=dd_model, lang=dd_lang, is_translate=cb_translate, beam_size=nb_beam_size,
                log_prob_threshold=nb_log_prob_threshold, no_speech_threshold=nb_no_speech_threshold,
                compute_type=dd_compute_type, best_of=nb_best_of, patience=nb_patience,
                condition_on_previous_text=cb_condition_on_previous_text, initial_prompt=tb_initial_prompt,
                temperature=sd_temperature, compression_ratio_threshold=nb_compression_ratio_threshold,
                vad_filter=cb_vad_filter, threshold=sd_threshold, min_speech_duration_ms=nb_min_speech_duration_ms,
                max_speech_duration_s=nb_max_speech_duration_s, min_silence_duration_ms=nb_min_silence_duration_ms,
                speech_pad_ms=nb_speech_pad_ms, chunk_length=nb_chunk_length, batch_size=nb_batch_size,
                is_diarize=cb_diarize, hf_token=tb_hf_token, diarization_device=dd_diarization_device,
                length_penalty=nb_length_penalty, repetition_penalty=nb_repetition_penalty,
                no_repeat_ngram_size=nb_no_repeat_ngram_size, prefix=tb_prefix, suppress_blank=cb_suppress_blank,
                suppress_tokens=tb_suppress_tokens, max_initial_timestamp=nb_max_initial_timestamp,
                word_timestamps=cb_word_timestamps, prepend_punctuations=tb_prepend_punctuations,
                append_punctuations=tb_append_punctuations, max_new_tokens=nb_max_new_tokens,
                hallucination_silence_threshold=nb_hallucination_silence_threshold, hotwords=tb_hotwords,
                language_detection_threshold=nb_language_detection_threshold,
                language_detection_segments=nb_language_detection_segments,
                prompt_reset_on_temperature=sld_prompt_reset_on_temperature, is_bgm_separate=cb_bgm_separation,
                uvr_device=dd_uvr_device, uvr_model_size=dd_uvr_model_size, uvr_segment_size=nb_uvr_segment_size,
                uvr_save_file=cb_uvr_save_file, uvr_enable_offload=cb_uvr_enable_offload
            ),
            dd_file_format,
            cb_timestamp,
            cb_translate_output,
            dd_translate_model,
            dd_target_lang
        )

    def launch(self):
        translation_params = self.default_params["translation"]
        deepl_params = translation_params["deepl"]
        nllb_params = translation_params["nllb"]
        uvr_params = self.default_params["bgm_separation"]

        with self.app:
            with gr.Row():
                with gr.Column():
                    gr.Markdown(MARKDOWN, elem_id="md_project")
            with gr.Tabs():
                with gr.TabItem("Audio upload/record"):  # tab1
                    with gr.Column():
                        #input_file = gr.Files(type="filepath", label="Upload File here")
                        #input_file = gr.File(type="filepath", label="Upload audio/video file here")
                        input_file = gr.Audio(type='filepath', elem_id="audio_input", show_download_button=True)
                        tb_input_folder = gr.Textbox(label="Input Folder Path (Optional)",
                                                     info="Optional: Specify the folder path where the input files are located, if you prefer to use local files instead of uploading them."
                                                          " Leave this field empty if you do not wish to use a local path.",
                                                     visible=self.args.colab,
                                                     value="")

                    whisper_params, dd_file_format, cb_timestamp, cb_translate_output, dd_translate_model, dd_target_lang = self.create_whisper_parameters()

                    with gr.Row():
                        btn_run = gr.Button("Transcribe", variant="primary")
                        btn_reset = gr.Button(value="Reset")
                        btn_reset.click(None,js="window.location.reload()")
                    with gr.Row():
                        with gr.Column(scale=4):
                            tb_indicator = gr.Textbox(label="Output preview (Always review output generated by AI models)", show_copy_button=True, show_label=True)
                        with gr.Column(scale=1):
                            tb_info = gr.Textbox(label="Output info", interactive=False, show_copy_button=True)
                            files_subtitles = gr.Files(label="Output data", interactive=False, file_count="multiple")
                            # btn_openfolder = gr.Button('📂', scale=1)                        

                    params = [input_file, tb_input_folder, dd_file_format, cb_timestamp, cb_translate_output, dd_translate_model, dd_target_lang]
                    btn_run.click(fn=self.whisper_inf.transcribe_file,
                                  inputs=params + whisper_params.as_list(),
                                  outputs=[tb_indicator, files_subtitles, tb_info])
                    # btn_openfolder.click(fn=lambda: self.open_folder("outputs"), inputs=None, outputs=None)

                with gr.TabItem("Device info"):  # tab2
                    with gr.Column():
                        gr.Markdown(device_info, label="Hardware info & installed packages")

        # Launch the app with optional gradio settings
        args = self.args

        self.app.queue(
            api_open=args.api_open
        ).launch(
            share=args.share,
            server_name=args.server_name,
            server_port=args.server_port,
            auth=(args.username, args.password) if args.username and args.password else None,
            root_path=args.root_path,
            inbrowser=args.inbrowser
        )

    @staticmethod
    def open_folder(folder_path: str):
        if os.path.exists(folder_path):
            os.system(f"start {folder_path}")
        else:
            os.makedirs(folder_path, exist_ok=True)
            print(f"The directory path {folder_path} has newly created.")

    @staticmethod
    def on_change_models(model_size: str):
        translatable_model = ["large", "large-v1", "large-v2", "large-v3"]
        if model_size not in translatable_model:
            return gr.Checkbox(visible=False, value=False, interactive=False)
           #return gr.Checkbox(visible=True, value=False, label="Translate to English (large models only)", interactive=False)
        else:
            return gr.Checkbox(visible=True, value=False, label="Translate to English", interactive=True)

# Create the parser for command-line arguments
parser = argparse.ArgumentParser()
parser.add_argument('--whisper_type', type=str, default="faster-whisper",
                    help='A type of the whisper implementation between: ["whisper", "faster-whisper", "insanely-fast-whisper"]')
parser.add_argument('--share', type=str2bool, default=False, nargs='?', const=True, help='Gradio share value')
parser.add_argument('--server_name', type=str, default=None, help='Gradio server host')
parser.add_argument('--server_port', type=int, default=None, help='Gradio server port')
parser.add_argument('--root_path', type=str, default=None, help='Gradio root path')
parser.add_argument('--username', type=str, default=None, help='Gradio authentication username')
parser.add_argument('--password', type=str, default=None, help='Gradio authentication password')
parser.add_argument('--theme', type=str, default=None, help='Gradio Blocks theme')
parser.add_argument('--colab', type=str2bool, default=False, nargs='?', const=True, help='Is colab user or not')
parser.add_argument('--api_open', type=str2bool, default=False, nargs='?', const=True, help='Enable api or not in Gradio')
parser.add_argument('--inbrowser', type=str2bool, default=True, nargs='?', const=True, help='Whether to automatically start Gradio app or not')
parser.add_argument('--whisper_model_dir', type=str, default=WHISPER_MODELS_DIR,
                    help='Directory path of the whisper model')
parser.add_argument('--faster_whisper_model_dir', type=str, default=FASTER_WHISPER_MODELS_DIR,
                    help='Directory path of the faster-whisper model')
parser.add_argument('--insanely_fast_whisper_model_dir', type=str,
                    default=INSANELY_FAST_WHISPER_MODELS_DIR,
                    help='Directory path of the insanely-fast-whisper model')
parser.add_argument('--diarization_model_dir', type=str, default=DIARIZATION_MODELS_DIR,
                    help='Directory path of the diarization model')
parser.add_argument('--nllb_model_dir', type=str, default=NLLB_MODELS_DIR,
                    help='Directory path of the Facebook NLLB model')
parser.add_argument('--uvr_model_dir', type=str, default=UVR_MODELS_DIR,
                    help='Directory path of the UVR model')
parser.add_argument('--output_dir', type=str, default=OUTPUT_DIR, help='Directory path of the outputs')
_args = parser.parse_args()

if __name__ == "__main__":
    app = App(args=_args)
    app.launch()