init

README.md

@@ -1,13 +1,12 @@
 ---
-title: BreezyVoice Playground
-emoji: 📚
-colorFrom: red
-colorTo: yellow
+title: BreezyVoice
+emoji: 🏆
+colorFrom: green
+colorTo: green
 sdk: gradio
-sdk_version: 5.16.0
+sdk_version: 5.12.0
 app_file: app.py
 pinned: false
-short_description: Playground of BreezyVoice
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,257 @@
+# Copyright (c) 2025 MediaTek Reserch Inc (authors: Chan-Jan Hsu)
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Liu Yue)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+import sys
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append('{}/third_party/Matcha-TTS'.format(ROOT_DIR))
+
+import argparse
+import gradio as gr
+import numpy as np
+import torch
+torch.set_num_threads(1)
+import torchaudio
+import random
+import librosa
+from transformers import pipeline
+import subprocess
+from scipy.signal import resample
+
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+
+from cosyvoice.cli.cosyvoice import CosyVoice
+from cosyvoice.utils.file_utils import load_wav, speed_change
+
+#logging.basicConfig(level=logging.DEBUG,
+#                    format='%(asctime)s %(levelname)s %(message)s')
+
+def generate_seed():
+    seed = random.randint(1, 100000000)
+    return {
+        "__type__": "update",
+        "value": seed
+    }
+
+def set_all_random_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+max_val = 0.8
+def postprocess(speech, top_db=60, hop_length=220, win_length=440):
+    speech, _ = librosa.effects.trim(
+        speech, top_db=top_db,
+        frame_length=win_length,
+        hop_length=hop_length
+    )
+    if speech.abs().max() > max_val:
+        speech = speech / speech.abs().max() * max_val
+    speech = torch.concat([speech, torch.zeros(1, int(target_sr * 0.2))], dim=1)
+    return speech
+
+def generate_audio(tts_text, prompt_text, prompt_wav_upload, prompt_wav_record, seed, select_which):
+    if select_which == "上傳檔案" and prompt_wav_upload is not None:
+        prompt_wav = prompt_wav_upload
+    elif select_which == "麥克風" and prompt_wav_record is not None:
+        prompt_wav = prompt_wav_record
+    else:
+        prompt_wav = None
+    # if instruct mode, please make sure that model is iic/CosyVoice-300M-Instruct and not cross_lingual mode
+
+    prompt_speech_16k = postprocess(load_wav(prompt_wav, prompt_sr))
+    set_all_random_seed(seed)
+    output = cosyvoice.inference_zero_shot(tts_text, prompt_text, prompt_speech_16k)
+    speed_factor = 1
+    if speed_factor != 1.0:
+        #try:
+            #audio_data, sample_rate = speed_change(output["tts_speech"], target_sr, str(speed_factor))
+            #audio_data = audio_data.numpy().flatten()
+        new_length = int(len(output['tts_speech']) / speed_factor)
+        audio_data = resample(output['tts_speech'], new_length)
+        # except Exception as e:
+        #     print(f"Failed to change speed of audio: \n{e}")
+    else:
+        audio_data = output['tts_speech'].numpy().flatten()
+
+    return (target_sr, audio_data)
+
+
+def generate_text(prompt_wav_upload, prompt_wav_record, select_which):
+    # Determine which input to use based on the selection in select_which
+    if select_which == "上傳檔案" and prompt_wav_upload is not None:
+        prompt_wav = prompt_wav_upload
+        LAST_UPLOADED = "upload"
+    elif select_which == "麥克風" and prompt_wav_record is not None:
+        prompt_wav = prompt_wav_record
+        LAST_UPLOADED = "record"
+    else:
+        prompt_wav = None
+        LAST_UPLOADED = None
+    print(select_which)
+    # Process with ASR pipeline
+    if prompt_wav:
+        results = asr_pipeline(prompt_wav)
+        return results['text']
+    return "No valid input detected."
+
+# LAST_UPLOADED = ""
+# def switch_selected(select_which):
+#     # Check the file type (assuming WAV file)
+#     if select_which == "上傳檔案" and prompt_wav_upload is not None:
+#         prompt_wav = prompt_wav_upload
+#         LAST_UPLOADED = "upload"
+#     elif select_which == "麥克風" and prompt_wav_record is not None:
+#         prompt_wav = prompt_wav_record
+#     return "麥克風"
+
+def demo_get_audio(tts_text):
+    sample_wav = 'sample.wav'
+    speech, sample_rate = torchaudio.load(sample_wav)
+    
+    return sample_rate, speech
+def main():
+    with gr.Blocks(title="BreezyVoice 語音合成系統", theme="default") as demo:
+        # Title and About section at the top
+        gr.Markdown("# BreezyVoice 語音合成系統")
+        gr.Markdown(
+            """## 僅需5秒語音樣本，就可輸出擬真人聲。
+
+            <img src="https://raw.githubusercontent.com/Splend1d/BreezyVoice/main/images/flowchart.png" alt="Flowchart" width="600"/>
+
+            #### 此沙盒使用 Huggingface CPU，請預期大於200 ��的推理時間，您可以考慮以下方法加速：
+            1. 複製這個 Space（僅當執行需要排隊時）
+            2. 複製至本地GPU執行（請參考[指南](https://huggingface.co/docs/hub/en/spaces-overview)）或使用[kaggle](https://www.kaggle.com/code/a24998667/breezyvoice-playground)
+            3. 複製至本地CPU執行（請參考[指南](https://huggingface.co/docs/hub/en/spaces-overview)）
+
+            為了加快推理速度，g2pw注音標註並未被啟動。
+
+            免責聲明：此沙盒在一次性容器地端執行，關閉後檔案將遭到刪除。此沙盒不屬於聯發創新基地，聯發創新基地無法獲得任何使用者輸入。"""
+        )
+
+        # All content arranged in a single column
+        with gr.Column():
+            # Configuration Section
+            
+
+            
+            # Grouping prompt audio inputs and auto speech recognition in one block using Markdown
+            gr.Markdown("### 步驟 1. 音訊樣本輸入 & 音訊樣本文本輸入")
+            gr.Markdown("選擇prompt音訊檔案或錄製prompt音訊，並手動校對自動產生的音訊樣本文本。")
+            prompt_wav_upload = gr.Audio(
+                sources='upload',
+                type='filepath',
+                label='選擇prompt音訊檔案（確保取樣率不低於16khz）'
+            )
+            prompt_wav_record = gr.Audio(
+                sources='microphone',
+                type='filepath',
+                label='錄製prompt音訊檔案'
+            )
+            
+            with gr.Blocks():
+                select_which = gr.Radio(["上傳檔案", "麥克風"], label="音訊來源", interactive=True )
+            with gr.Blocks():
+                prompt_text = gr.Textbox(
+                    label="音訊樣本文本輸入(此欄位應與音檔內容完全相同)",
+                    lines=2,
+                    placeholder="音訊樣本文本"
+                )
+
+            # Automatic speech recognition when either prompt audio input changes
+            def a(X):
+                return "上傳檔案"
+            prompt_wav_upload.change(
+                fn=a,#lambda file: "上傳檔案",
+                inputs=[prompt_wav_upload],
+                outputs=select_which
+            )
+
+            
+            
+            
+
+            prompt_wav_record.change(
+                fn=lambda recording: "麥克風",
+                inputs=[prompt_wav_record],
+                outputs=select_which
+            )
+
+            select_which.change(
+                fn=generate_text,
+                inputs=[prompt_wav_upload, prompt_wav_record, select_which],
+                outputs=prompt_text
+            )
+            # select_which.change(
+            #     fn=switch_selected,
+            #     inputs=[select_which],
+            #     outputs= None
+            # )
+            # Input Section: Synthesis Text
+
+            gr.Markdown("### 步驟 2.合成文本輸入")
+            tts_text = gr.Textbox(
+                label="輸入想要合成的文本",
+                lines=2,
+                placeholder="請輸入想要合成的文本...",
+                value="你好，歡迎光臨"
+            )
+
+
+            # Output Section
+            gr.Markdown("### 步驟 3. 合成音訊")
+            # Generation button for audio synthesis (triggered manually)
+
+            with gr.Accordion("進階設定", open=False):
+                seed = gr.Number(value=0, label="隨機推理種子")
+                #seed_button = gr.Button("隨機")
+                seed_button = gr.Button(value="\U0001F3B2生成隨機推理種子\U0001F3B2")
+                speed_factor = 1
+                # speed_factor = gr.Slider(
+                #     minimum=0.25,
+                #     maximum=4,
+                #     step=0.05,
+                #     label="語速",
+                #     value=1.0,
+                #     interactive=True
+                # )
+
+            generate_button = gr.Button("生成音訊")
+            audio_output = gr.Audio(label="合成音訊")
+
+            # Set up callbacks for seed generation and audio synthesis
+            seed_button.click(fn=generate_seed, inputs=[], outputs=seed)
+            generate_button.click(
+                fn=generate_audio,
+                inputs=[tts_text, prompt_text, prompt_wav_upload, prompt_wav_record, seed, select_which],
+                outputs=audio_output
+            )
+
+        demo.queue(max_size=4, default_concurrency_limit=2)
+        demo.launch()
+        
+if __name__ == '__main__':
+    cosyvoice = CosyVoice('Splend1dchan/BreezyVoice')
+    asr_pipeline = pipeline(
+        "automatic-speech-recognition",
+        model="openai/whisper-tiny",
+        tokenizer="openai/whisper-tiny",
+        device=0  # Use GPU (if available); set to -1 for CPU
+    )
+    sft_spk = cosyvoice.list_avaliable_spks()
+    prompt_sr, target_sr = 16000, 22050
+    default_data = np.zeros(target_sr)
+    main()

cosyvoice/bin/inference.py

@@ -0,0 +1,114 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import argparse
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+import os
+
+import torch
+from torch.utils.data import DataLoader
+import torchaudio
+from hyperpyyaml import load_hyperpyyaml
+from tqdm import tqdm
+from cosyvoice.cli.model import CosyVoiceModel
+
+from cosyvoice.dataset.dataset import Dataset
+
+def get_args():
+    parser = argparse.ArgumentParser(description='inference with your model')
+    parser.add_argument('--config', required=True, help='config file')
+    parser.add_argument('--prompt_data', required=True, help='prompt data file')
+    parser.add_argument('--prompt_utt2data', required=True, help='prompt data file')
+    parser.add_argument('--tts_text', required=True, help='tts input file')
+    parser.add_argument('--llm_model', required=True, help='llm model file')
+    parser.add_argument('--flow_model', required=True, help='flow model file')
+    parser.add_argument('--hifigan_model', required=True, help='hifigan model file')
+    parser.add_argument('--gpu',
+                        type=int,
+                        default=-1,
+                        help='gpu id for this rank, -1 for cpu')
+    parser.add_argument('--mode',
+                        default='sft',
+                        choices=['sft', 'zero_shot'],
+                        help='inference mode')
+    parser.add_argument('--result_dir', required=True, help='asr result file')
+    args = parser.parse_args()
+    print(args)
+    return args
+
+
+def main():
+    args = get_args()
+    logging.basicConfig(level=logging.DEBUG,
+                        format='%(asctime)s %(levelname)s %(message)s')
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
+
+    # Init cosyvoice models from configs
+    use_cuda = args.gpu >= 0 and torch.cuda.is_available()
+    device = torch.device('cuda' if use_cuda else 'cpu')
+    with open(args.config, 'r') as f:
+        configs = load_hyperpyyaml(f)
+
+    model = CosyVoiceModel(configs['llm'], configs['flow'], configs['hift'])
+    model.load(args.llm_model, args.flow_model, args.hifigan_model)
+
+    test_dataset = Dataset(args.prompt_data, data_pipeline=configs['data_pipeline'], mode='inference', shuffle=False, partition=False, tts_file=args.tts_text, prompt_utt2data=args.prompt_utt2data)
+    test_data_loader = DataLoader(test_dataset, batch_size=None, num_workers=0)
+
+    del configs
+    os.makedirs(args.result_dir, exist_ok=True)
+    fn = os.path.join(args.result_dir, 'wav.scp')
+    f = open(fn, 'w')
+    with torch.no_grad():
+        for batch_idx, batch in tqdm(enumerate(test_data_loader)):
+            utts = batch["utts"]
+            assert len(utts) == 1, "inference mode only support batchsize 1"
+            text = batch["text"]
+            text_token = batch["text_token"].to(device)
+            text_token_len = batch["text_token_len"].to(device)
+            tts_text = batch["tts_text"]
+            tts_index = batch["tts_index"]
+            tts_text_token = batch["tts_text_token"].to(device)
+            tts_text_token_len = batch["tts_text_token_len"].to(device)
+            speech_token = batch["speech_token"].to(device)
+            speech_token_len = batch["speech_token_len"].to(device)
+            speech_feat = batch["speech_feat"].to(device)
+            speech_feat_len = batch["speech_feat_len"].to(device)
+            utt_embedding = batch["utt_embedding"].to(device)
+            spk_embedding = batch["spk_embedding"].to(device)
+            if args.mode == 'sft':
+                model_input = {'text': tts_text_token, 'text_len': tts_text_token_len,
+                               'llm_embedding': spk_embedding, 'flow_embedding': spk_embedding}
+            else:
+                model_input = {'text': tts_text_token, 'text_len': tts_text_token_len,
+                               'prompt_text': text_token, 'prompt_text_len': text_token_len,
+                               'llm_prompt_speech_token': speech_token, 'llm_prompt_speech_token_len': speech_token_len,
+                               'flow_prompt_speech_token': speech_token, 'flow_prompt_speech_token_len': speech_token_len,
+                               'prompt_speech_feat': speech_feat, 'prompt_speech_feat_len': speech_feat_len,
+                               'llm_embedding': utt_embedding, 'flow_embedding': utt_embedding}
+            model_output = model.inference(**model_input)
+            tts_key = '{}_{}'.format(utts[0], tts_index[0])
+            tts_fn = os.path.join(args.result_dir, '{}.wav'.format(tts_key))
+            torchaudio.save(tts_fn, model_output['tts_speech'], sample_rate=22050)
+            f.write('{} {}\n'.format(tts_key, tts_fn))
+            f.flush()
+    f.close()
+    logging.info('Result wav.scp saved in {}'.format(fn))
+
+
+if __name__ == '__main__':
+    main()

cosyvoice/bin/train.py

@@ -0,0 +1,136 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+import argparse
+import datetime
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+from copy import deepcopy
+import torch
+import torch.distributed as dist
+import deepspeed
+
+from hyperpyyaml import load_hyperpyyaml
+
+from torch.distributed.elastic.multiprocessing.errors import record
+
+from cosyvoice.utils.executor import Executor
+from cosyvoice.utils.train_utils import (
+    init_distributed,
+    init_dataset_and_dataloader,
+    init_optimizer_and_scheduler,
+    init_summarywriter, save_model,
+    wrap_cuda_model, check_modify_and_save_config)
+
+
+def get_args():
+    parser = argparse.ArgumentParser(description='training your network')
+    parser.add_argument('--train_engine',
+                        default='torch_ddp',
+                        choices=['torch_ddp', 'deepspeed'],
+                        help='Engine for paralleled training')
+    parser.add_argument('--model', required=True, help='model which will be trained')
+    parser.add_argument('--config', required=True, help='config file')
+    parser.add_argument('--train_data', required=True, help='train data file')
+    parser.add_argument('--cv_data', required=True, help='cv data file')
+    parser.add_argument('--checkpoint', help='checkpoint model')
+    parser.add_argument('--model_dir', required=True, help='save model dir')
+    parser.add_argument('--tensorboard_dir',
+                        default='tensorboard',
+                        help='tensorboard log dir')
+    parser.add_argument('--ddp.dist_backend',
+                        dest='dist_backend',
+                        default='nccl',
+                        choices=['nccl', 'gloo'],
+                        help='distributed backend')
+    parser.add_argument('--num_workers',
+                        default=0,
+                        type=int,
+                        help='num of subprocess workers for reading')
+    parser.add_argument('--prefetch',
+                        default=100,
+                        type=int,
+                        help='prefetch number')
+    parser.add_argument('--pin_memory',
+                        action='store_true',
+                        default=False,
+                        help='Use pinned memory buffers used for reading')
+    parser.add_argument('--deepspeed.save_states',
+                        dest='save_states',
+                        default='model_only',
+                        choices=['model_only', 'model+optimizer'],
+                        help='save model/optimizer states')
+    parser.add_argument('--timeout',
+                        default=30,
+                        type=int,
+                        help='timeout (in seconds) of cosyvoice_join.')
+    parser = deepspeed.add_config_arguments(parser)
+    args = parser.parse_args()
+    return args
+
+
+@record
+def main():
+    args = get_args()
+    logging.basicConfig(level=logging.DEBUG,
+                        format='%(asctime)s %(levelname)s %(message)s')
+
+    override_dict = {k: None for k in ['llm', 'flow', 'hift'] if k != args.model}
+    with open(args.config, 'r') as f:
+        configs = load_hyperpyyaml(f, overrides=override_dict)
+    configs['train_conf'].update(vars(args))
+
+    # Init env for ddp
+    init_distributed(args)
+
+    # Get dataset & dataloader
+    train_dataset, cv_dataset, train_data_loader, cv_data_loader = \
+        init_dataset_and_dataloader(args, configs)
+
+    # Do some sanity checks and save config to arsg.model_dir
+    configs = check_modify_and_save_config(args, configs)
+
+    # Tensorboard summary
+    writer = init_summarywriter(args)
+
+    # load checkpoint
+    model = configs[args.model]
+    if args.checkpoint is not None:
+        model.load_state_dict(torch.load(args.checkpoint, map_location='cpu'))
+
+    # Dispatch model from cpu to gpu
+    model = wrap_cuda_model(args, model)
+
+    # Get optimizer & scheduler
+    model, optimizer, scheduler = init_optimizer_and_scheduler(args, configs, model)
+
+    # Save init checkpoints
+    info_dict = deepcopy(configs['train_conf'])
+    save_model(model, 'init', info_dict)
+
+    # Get executor
+    executor = Executor()
+
+    # Start training loop
+    for epoch in range(info_dict['max_epoch']):
+        executor.epoch = epoch
+        train_dataset.set_epoch(epoch)
+        dist.barrier()
+        group_join = dist.new_group(backend="gloo", timeout=datetime.timedelta(seconds=args.timeout))
+        executor.train_one_epoc(model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, group_join)
+        dist.destroy_process_group(group_join)
+
+if __name__ == '__main__':
+    main()

cosyvoice/cli/cosyvoice.py

@@ -0,0 +1,83 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+import torch
+from hyperpyyaml import load_hyperpyyaml
+from huggingface_hub import snapshot_download
+from cosyvoice.cli.frontend import CosyVoiceFrontEnd
+from cosyvoice.cli.model import CosyVoiceModel
+
+class CosyVoice:
+
+    def __init__(self, model_dir):
+        instruct = True if '-Instruct' in model_dir else False
+        self.model_dir = model_dir
+        if not os.path.exists(model_dir):
+            model_dir = snapshot_download(model_dir)
+        with open('{}/cosyvoice.yaml'.format(model_dir), 'r') as f:
+            configs = load_hyperpyyaml(f)
+        self.frontend = CosyVoiceFrontEnd(configs['get_tokenizer'],
+                                          configs['feat_extractor'],
+                                          '{}/campplus.onnx'.format(model_dir),
+                                          '{}/speech_tokenizer_v1.onnx'.format(model_dir),
+                                          '{}/spk2info.pt'.format(model_dir),
+                                          instruct,
+                                          configs['allowed_special'])
+        self.model = CosyVoiceModel(configs['llm'], configs['flow'], configs['hift'])
+        self.model.load('{}/llm.pt'.format(model_dir),
+                        '{}/flow.pt'.format(model_dir),
+                        '{}/hift.pt'.format(model_dir))
+        del configs
+
+    def list_avaliable_spks(self):
+        spks = list(self.frontend.spk2info.keys())
+        return spks
+
+    def inference_sft(self, tts_text, spk_id):
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_sft(i, spk_id)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}
+
+    def inference_zero_shot(self, tts_text, prompt_text, prompt_speech_16k):
+        prompt_text = self.frontend.text_normalize(prompt_text, split=False)
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_zero_shot(i, prompt_text, prompt_speech_16k)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}
+
+    def inference_cross_lingual(self, tts_text, prompt_speech_16k):
+        if self.frontend.instruct is True:
+            raise ValueError('{} do not support cross_lingual inference'.format(self.model_dir))
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_cross_lingual(i, prompt_speech_16k)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}
+
+    def inference_instruct(self, tts_text, spk_id, instruct_text):
+        if self.frontend.instruct is False:
+            raise ValueError('{} do not support instruct inference'.format(self.model_dir))
+        instruct_text = self.frontend.text_normalize(instruct_text, split=False)
+        tts_speeches = []
+        for i in self.frontend.text_normalize(tts_text, split=True):
+            model_input = self.frontend.frontend_instruct(i, spk_id, instruct_text)
+            model_output = self.model.inference(**model_input)
+            tts_speeches.append(model_output['tts_speech'])
+        return {'tts_speech': torch.concat(tts_speeches, dim=1)}

cosyvoice/cli/frontend.py

@@ -0,0 +1,183 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from functools import partial
+import onnxruntime
+import torch
+import numpy as np
+import whisper
+from typing import Callable
+import torchaudio.compliance.kaldi as kaldi
+import torchaudio
+import os
+import re
+import inflect
+import subprocess
+try:
+    import ttsfrd
+    use_ttsfrd = True
+except ImportError:
+    print("failed to import ttsfrd, use WeTextProcessing instead")
+    from tn.chinese.normalizer import Normalizer as ZhNormalizer
+    from tn.english.normalizer import Normalizer as EnNormalizer
+    use_ttsfrd = False
+from cosyvoice.utils.frontend_utils import contains_chinese, replace_blank, replace_corner_mark, remove_bracket, spell_out_number, split_paragraph
+
+
+class CosyVoiceFrontEnd:
+
+    def __init__(self,
+                 get_tokenizer: Callable,
+                 feat_extractor: Callable,
+                 campplus_model: str,
+                 speech_tokenizer_model: str,
+                 spk2info: str = '',
+                 instruct: bool = False,
+                 allowed_special: str = 'all'):
+        self.tokenizer = get_tokenizer()
+        self.feat_extractor = feat_extractor
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        option = onnxruntime.SessionOptions()
+        option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+        option.intra_op_num_threads = 1
+        self.campplus_session = onnxruntime.InferenceSession(campplus_model, sess_options=option, providers=["CPUExecutionProvider"])
+        self.speech_tokenizer_session = onnxruntime.InferenceSession(speech_tokenizer_model, sess_options=option, providers=["CUDAExecutionProvider"if torch.cuda.is_available() else "CPUExecutionProvider"])
+        if os.path.exists(spk2info):
+            self.spk2info = torch.load(spk2info, map_location=self.device)
+        self.instruct = instruct
+        self.allowed_special = allowed_special
+        self.inflect_parser = inflect.engine()
+        self.use_ttsfrd = use_ttsfrd
+        if self.use_ttsfrd:
+            self.frd = ttsfrd.TtsFrontendEngine()
+            ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+            #print("LOCATION",ttsfrd.__file__)
+            #print('TTSFRD FILES',os.listdir(ttsfrd.__file__))
+            if not os.path.exists('resource.zip'):
+                # Download the file if it does not exist
+                subprocess.run("wget https://huggingface.co/FunAudioLLM/CosyVoice-ttsfrd/resolve/main/resource.zip".split())
+
+            # Unzip the file if it exists
+            if not os.path.exists('resource'):
+                subprocess.run("unzip resource.zip".split())
+            else:
+                pass
+            #print(os.listdir())
+            #print(subprocess.run("pwd"))
+            print("root",ROOT_DIR)
+            assert self.frd.initialize('{}/../../resource'.format(ROOT_DIR)) is True, 'failed to initialize ttsfrd resource'
+            self.frd.set_lang_type('pinyin')
+            self.frd.enable_pinyin_mix(True)
+            self.frd.set_breakmodel_index(1)
+        else:
+            self.zh_tn_model = ZhNormalizer(remove_erhua=False, full_to_half=False)
+            self.en_tn_model = EnNormalizer()
+
+    def _extract_text_token(self, text):
+        text_token = self.tokenizer.encode(text, allowed_special=self.allowed_special)
+        text_token = torch.tensor([text_token], dtype=torch.int32).to(self.device)
+        text_token_len = torch.tensor([text_token.shape[1]], dtype=torch.int32).to(self.device)
+        return text_token, text_token_len
+
+    def _extract_speech_token(self, speech):
+        feat = whisper.log_mel_spectrogram(speech, n_mels=128)
+        speech_token = self.speech_tokenizer_session.run(None, {self.speech_tokenizer_session.get_inputs()[0].name: feat.detach().cpu().numpy(),
+                                                                self.speech_tokenizer_session.get_inputs()[1].name: np.array([feat.shape[2]], dtype=np.int32)})[0].flatten().tolist()
+        speech_token = torch.tensor([speech_token], dtype=torch.int32).to(self.device)
+        speech_token_len = torch.tensor([speech_token.shape[1]], dtype=torch.int32).to(self.device)
+        return speech_token, speech_token_len
+
+    def _extract_spk_embedding(self, speech):
+        feat = kaldi.fbank(speech,
+                           num_mel_bins=80,
+                           dither=0,
+                           sample_frequency=16000)
+        feat = feat - feat.mean(dim=0, keepdim=True)
+        embedding = self.campplus_session.run(None, {self.campplus_session.get_inputs()[0].name: feat.unsqueeze(dim=0).cpu().numpy()})[0].flatten().tolist()
+        embedding = torch.tensor([embedding]).to(self.device)
+        return embedding
+
+    def _extract_speech_feat(self, speech):
+        speech_feat = self.feat_extractor(speech).squeeze(dim=0).transpose(0, 1).to(self.device)
+        speech_feat = speech_feat.unsqueeze(dim=0)
+        speech_feat_len = torch.tensor([speech_feat.shape[1]], dtype=torch.int32).to(self.device)
+        return speech_feat, speech_feat_len
+
+    def text_normalize(self, text, split=True):
+        text = text.strip()
+        if contains_chinese(text):
+            if self.use_ttsfrd:
+                text = self.frd.get_frd_extra_info(text, 'input')
+            else:
+                text = self.zh_tn_model.normalize(text)
+            text = text.replace("\n", "")
+            text = replace_blank(text)
+            text = replace_corner_mark(text)
+            text = text.replace(".", "、")
+            text = text.replace(" - ", "，")
+            text = remove_bracket(text)
+            text = re.sub(r'[，,]+$', '。', text)
+            texts = [i for i in split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "zh", token_max_n=80,
+                                                token_min_n=60, merge_len=20,
+                                                comma_split=False)]
+        else:
+            if self.use_ttsfrd:
+                text = self.frd.get_frd_extra_info(text, 'input')
+            else:
+                text = self.en_tn_model.normalize(text)
+            text = spell_out_number(text, self.inflect_parser)
+            texts = [i for i in split_paragraph(text, partial(self.tokenizer.encode, allowed_special=self.allowed_special), "en", token_max_n=80,
+                                                token_min_n=60, merge_len=20,
+                                                comma_split=False)]
+        if split is False:
+            return text
+        return texts
+
+    def frontend_sft(self, tts_text, spk_id):
+        tts_text_token, tts_text_token_len = self._extract_text_token(tts_text)
+        embedding = self.spk2info[spk_id]['embedding']
+        model_input = {'text': tts_text_token, 'text_len': tts_text_token_len, 'llm_embedding': embedding, 'flow_embedding': embedding}
+        return model_input
+
+    def frontend_zero_shot(self, tts_text, prompt_text, prompt_speech_16k):
+        tts_text_token, tts_text_token_len = self._extract_text_token(tts_text)
+        prompt_text_token, prompt_text_token_len = self._extract_text_token(prompt_text)
+        prompt_speech_22050 = torchaudio.transforms.Resample(orig_freq=16000, new_freq=22050)(prompt_speech_16k)
+        speech_feat, speech_feat_len = self._extract_speech_feat(prompt_speech_22050)
+        speech_token, speech_token_len = self._extract_speech_token(prompt_speech_16k)
+        embedding = self._extract_spk_embedding(prompt_speech_16k)
+        model_input = {'text': tts_text_token, 'text_len': tts_text_token_len,
+                       'prompt_text': prompt_text_token, 'prompt_text_len': prompt_text_token_len,
+                       'llm_prompt_speech_token': speech_token, 'llm_prompt_speech_token_len': speech_token_len,
+                       'flow_prompt_speech_token': speech_token, 'flow_prompt_speech_token_len': speech_token_len,
+                       'prompt_speech_feat': speech_feat, 'prompt_speech_feat_len': speech_feat_len,
+                       'llm_embedding': embedding, 'flow_embedding': embedding}
+        return model_input
+
+    def frontend_cross_lingual(self, tts_text, prompt_speech_16k):
+        model_input = self.frontend_zero_shot(tts_text, '', prompt_speech_16k)
+        # in cross lingual mode, we remove prompt in llm
+        del model_input['prompt_text']
+        del model_input['prompt_text_len']
+        del model_input['llm_prompt_speech_token']
+        del model_input['llm_prompt_speech_token_len']
+        return model_input
+
+    def frontend_instruct(self, tts_text, spk_id, instruct_text):
+        model_input = self.frontend_sft(tts_text, spk_id)
+        # in instruct mode, we remove spk_embedding in llm due to information leakage
+        del model_input['llm_embedding']
+        instruct_text_token, instruct_text_token_len = self._extract_text_token(instruct_text + '<endofprompt>')
+        model_input['prompt_text'] = instruct_text_token
+        model_input['prompt_text_len'] = instruct_text_token_len
+        return model_input

cosyvoice/cli/model.py

@@ -0,0 +1,60 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+class CosyVoiceModel:
+
+    def __init__(self,
+                 llm: torch.nn.Module,
+                 flow: torch.nn.Module,
+                 hift: torch.nn.Module):
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.llm = llm
+        self.flow = flow
+        self.hift = hift
+
+    def load(self, llm_model, flow_model, hift_model):
+        self.llm.load_state_dict(torch.load(llm_model, map_location=self.device))
+        self.llm.to(self.device).eval()
+        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device))
+        self.flow.to(self.device).eval()
+        self.hift.load_state_dict(torch.load(hift_model, map_location=self.device))
+        self.hift.to(self.device).eval()
+
+    def inference(self, text, text_len, flow_embedding, llm_embedding=torch.zeros(0, 192),
+                  prompt_text=torch.zeros(1, 0, dtype=torch.int32), prompt_text_len=torch.zeros(1, dtype=torch.int32),
+                  llm_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), llm_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
+                  flow_prompt_speech_token=torch.zeros(1, 0, dtype=torch.int32), flow_prompt_speech_token_len=torch.zeros(1, dtype=torch.int32),
+                  prompt_speech_feat=torch.zeros(1, 0, 80), prompt_speech_feat_len=torch.zeros(1, dtype=torch.int32)):
+        tts_speech_token = self.llm.inference(text=text.to(self.device),
+                                              text_len=text_len.to(self.device),
+                                              prompt_text=prompt_text.to(self.device),
+                                              prompt_text_len=prompt_text_len.to(self.device),
+                                              prompt_speech_token=llm_prompt_speech_token.to(self.device),
+                                              prompt_speech_token_len=llm_prompt_speech_token_len.to(self.device),
+                                              embedding=llm_embedding.to(self.device),
+                                              beam_size=1,
+                                              sampling=25,
+                                              max_token_text_ratio=30,
+                                              min_token_text_ratio=3)
+        tts_mel = self.flow.inference(token=tts_speech_token,
+                                      token_len=torch.tensor([tts_speech_token.size(1)], dtype=torch.int32).to(self.device),
+                                      prompt_token=flow_prompt_speech_token.to(self.device),
+                                      prompt_token_len=flow_prompt_speech_token_len.to(self.device),
+                                      prompt_feat=prompt_speech_feat.to(self.device),
+                                      prompt_feat_len=prompt_speech_feat_len.to(self.device),
+                                      embedding=flow_embedding.to(self.device))
+        tts_speech = self.hift.inference(mel=tts_mel).cpu()
+        torch.cuda.empty_cache()
+        return {'tts_speech': tts_speech}

cosyvoice/dataset/dataset.py

@@ -0,0 +1,160 @@
+# Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+import json
+import math
+from functools import partial
+
+import torch
+import torch.distributed as dist
+from torch.utils.data import IterableDataset
+from cosyvoice.utils.file_utils import read_lists, read_json_lists
+
+
+class Processor(IterableDataset):
+
+    def __init__(self, source, f, *args, **kw):
+        assert callable(f)
+        self.source = source
+        self.f = f
+        self.args = args
+        self.kw = kw
+
+    def set_epoch(self, epoch):
+        self.source.set_epoch(epoch)
+
+    def __iter__(self):
+        """ Return an iterator over the source dataset processed by the
+            given processor.
+        """
+        assert self.source is not None
+        assert callable(self.f)
+        return self.f(iter(self.source), *self.args, **self.kw)
+
+    def apply(self, f):
+        assert callable(f)
+        return Processor(self, f, *self.args, **self.kw)
+
+
+class DistributedSampler:
+
+    def __init__(self, shuffle=True, partition=True):
+        self.epoch = -1
+        self.update()
+        self.shuffle = shuffle
+        self.partition = partition
+
+    def update(self):
+        assert dist.is_available()
+        if dist.is_initialized():
+            self.rank = dist.get_rank()
+            self.world_size = dist.get_world_size()
+        else:
+            self.rank = 0
+            self.world_size = 1
+        worker_info = torch.utils.data.get_worker_info()
+        if worker_info is None:
+            self.worker_id = 0
+            self.num_workers = 1
+        else:
+            self.worker_id = worker_info.id
+            self.num_workers = worker_info.num_workers
+        return dict(rank=self.rank,
+                    world_size=self.world_size,
+                    worker_id=self.worker_id,
+                    num_workers=self.num_workers)
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+    def sample(self, data):
+        """ Sample data according to rank/world_size/num_workers
+
+            Args:
+                data(List): input data list
+
+            Returns:
+                List: data list after sample
+        """
+        data = list(range(len(data)))
+        # force datalist even
+        if self.partition:
+            if self.shuffle:
+                random.Random(self.epoch).shuffle(data)
+            if len(data) < self.world_size:
+                data = data * math.ceil(self.world_size / len(data))
+                data = data[:self.world_size]
+            data = data[self.rank::self.world_size]
+        if len(data) < self.num_workers:
+            data = data * math.ceil(self.num_workers / len(data))
+            data = data[:self.num_workers]
+        data = data[self.worker_id::self.num_workers]
+        return data
+
+
+class DataList(IterableDataset):
+
+    def __init__(self, lists, shuffle=True, partition=True):
+        self.lists = lists
+        self.sampler = DistributedSampler(shuffle, partition)
+
+    def set_epoch(self, epoch):
+        self.sampler.set_epoch(epoch)
+
+    def __iter__(self):
+        sampler_info = self.sampler.update()
+        indexes = self.sampler.sample(self.lists)
+        for index in indexes:
+            data = dict(src=self.lists[index])
+            data.update(sampler_info)
+            yield data
+
+
+def Dataset(data_list_file,
+            data_pipeline,
+            mode='train',
+            shuffle=True,
+            partition=True,
+            tts_file='',
+            prompt_utt2data=''):
+    """ Construct dataset from arguments
+
+        We have two shuffle stage in the Dataset. The first is global
+        shuffle at shards tar/raw file level. The second is global shuffle
+        at training samples level.
+
+        Args:
+            data_type(str): raw/shard
+            tokenizer (BaseTokenizer): tokenizer to tokenize
+            partition(bool): whether to do data partition in terms of rank
+    """
+    assert mode in ['train', 'inference']
+    lists = read_lists(data_list_file)
+    if mode == 'inference':
+        with open(tts_file) as f:
+            tts_data = json.load(f)
+        utt2lists = read_json_lists(prompt_utt2data)
+        # filter unnecessary file in inference mode
+        lists = list(set([utt2lists[utt] for utt in tts_data.keys() if utt2lists[utt] in lists]))
+    dataset = DataList(lists,
+                       shuffle=shuffle,
+                       partition=partition)
+    if mode == 'inference':
+        # map partial arg tts_data in inference mode
+        data_pipeline[0] = partial(data_pipeline[0], tts_data=tts_data)
+    for func in data_pipeline:
+        dataset = Processor(dataset, func, mode=mode)
+    return dataset

cosyvoice/dataset/processor.py

@@ -0,0 +1,369 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+import random
+
+import pyarrow.parquet as pq
+from io import BytesIO
+import torch
+import torchaudio
+from torch.nn.utils.rnn import pad_sequence
+import torch.nn.functional as F
+
+torchaudio.set_audio_backend('soundfile')
+
+AUDIO_FORMAT_SETS = set(['flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma'])
+
+
+def parquet_opener(data, mode='train', tts_data={}):
+    """ Give url or local file, return file descriptor
+        Inplace operation.
+
+        Args:
+            data(Iterable[str]): url or local file list
+
+        Returns:
+            Iterable[{src, stream}]
+    """
+    for sample in data:
+        assert 'src' in sample
+        url = sample['src']
+        try:
+            df = pq.read_table(url).to_pandas()
+            for i in range(len(df)):
+                if mode == 'inference' and df.loc[i, 'utt'] not in tts_data:
+                    continue
+                sample.update(dict(df.loc[i]))
+                if mode == 'train':
+                    # NOTE do not return sample directly, must initialize a new dict
+                    yield {**sample}
+                else:
+                    for index, text in enumerate(tts_data[df.loc[i, 'utt']]):
+                        yield {**sample, 'tts_index': index, 'tts_text': text}
+        except Exception as ex:
+            logging.warning('Failed to open {}, ex info {}'.format(url, ex))
+
+def filter(data,
+           max_length=10240,
+           min_length=10,
+           token_max_length=200,
+           token_min_length=1,
+           min_output_input_ratio=0.0005,
+           max_output_input_ratio=1,
+           mode='train'):
+    """ Filter sample according to feature and label length
+        Inplace operation.
+
+        Args::
+            data: Iterable[{key, wav, label, sample_rate}]
+            max_length: drop utterance which is greater than max_length(10ms)
+            min_length: drop utterance which is less than min_length(10ms)
+            token_max_length: drop utterance which is greater than
+                token_max_length, especially when use char unit for
+                english modeling
+            token_min_length: drop utterance which is
+                less than token_max_length
+            min_output_input_ratio: minimal ration of
+                token_length / feats_length(10ms)
+            max_output_input_ratio: maximum ration of
+                token_length / feats_length(10ms)
+
+        Returns:
+            Iterable[{key, wav, label, sample_rate}]
+    """
+    for sample in data:
+        sample['speech'], sample['sample_rate'] = torchaudio.load(BytesIO(sample['audio_data']))
+        del sample['audio_data']
+        # sample['wav'] is torch.Tensor, we have 100 frames every second
+        num_frames = sample['speech'].size(1) / sample['sample_rate'] * 100
+        if num_frames < min_length:
+            continue
+        if num_frames > max_length:
+            continue
+        if len(sample['text_token']) < token_min_length:
+            continue
+        if len(sample['text_token']) > token_max_length:
+            continue
+        if len(sample['speech_token']) == 0:
+            continue
+        if num_frames != 0:
+            if len(sample['text_token']) / num_frames < min_output_input_ratio:
+                continue
+            if len(sample['text_token']) / num_frames > max_output_input_ratio:
+                continue
+        yield sample
+
+
+def resample(data, resample_rate=22050, min_sample_rate=16000, mode='train'):
+    """ Resample data.
+        Inplace operation.
+
+        Args:
+            data: Iterable[{key, wav, label, sample_rate}]
+            resample_rate: target resample rate
+
+        Returns:
+            Iterable[{key, wav, label, sample_rate}]
+    """
+    for sample in data:
+        assert 'sample_rate' in sample
+        assert 'speech' in sample
+        sample_rate = sample['sample_rate']
+        waveform = sample['speech']
+        if sample_rate != resample_rate:
+            if sample_rate < min_sample_rate:
+                continue
+            sample['sample_rate'] = resample_rate
+            sample['speech'] = torchaudio.transforms.Resample(
+                orig_freq=sample_rate, new_freq=resample_rate)(waveform)
+        max_val = sample['speech'].abs().max()
+        if max_val > 1:
+            sample['speech'] /= max_val
+        yield sample
+
+
+def compute_fbank(data,
+                  feat_extractor,
+                  mode='train'):
+    """ Extract fbank
+
+        Args:
+            data: Iterable[{key, wav, label, sample_rate}]
+
+        Returns:
+            Iterable[{key, feat, label}]
+    """
+    for sample in data:
+        assert 'sample_rate' in sample
+        assert 'speech' in sample
+        assert 'utt' in sample
+        assert 'text_token' in sample
+        waveform = sample['speech']
+        mat = feat_extractor(waveform).squeeze(dim=0).transpose(0, 1)
+        sample['speech_feat'] = mat
+        del sample['speech']
+        yield sample
+
+
+def parse_embedding(data, normalize, mode='train'):
+    """ Parse utt_embedding/spk_embedding
+
+        Args:
+            data: Iterable[{key, wav, label, sample_rate}]
+
+        Returns:
+            Iterable[{key, feat, label}]
+    """
+    for sample in data:
+        sample['utt_embedding'] = torch.tensor(sample['utt_embedding'], dtype=torch.float32)
+        sample['spk_embedding'] = torch.tensor(sample['spk_embedding'], dtype=torch.float32)
+        if normalize:
+            sample['utt_embedding'] = F.normalize(sample['utt_embedding'], dim=0)
+            sample['spk_embedding'] = F.normalize(sample['spk_embedding'], dim=0)
+        yield sample
+
+
+def tokenize(data, get_tokenizer, allowed_special, mode='train'):
+    """ Decode text to chars or BPE
+        Inplace operation
+
+        Args:
+            data: Iterable[{key, wav, txt, sample_rate}]
+
+        Returns:
+            Iterable[{key, wav, txt, tokens, label, sample_rate}]
+    """
+    tokenizer = get_tokenizer()
+    for sample in data:
+        assert 'text' in sample
+        sample['text_token'] = tokenizer.encode(sample['text'], allowed_special=allowed_special)
+        if mode == 'inference':
+            sample['tts_text_token'] = tokenizer.encode(sample['tts_text'], allowed_special=allowed_special)
+        yield sample
+
+
+def shuffle(data, shuffle_size=10000, mode='train'):
+    """ Local shuffle the data
+
+        Args:
+            data: Iterable[{key, feat, label}]
+            shuffle_size: buffer size for shuffle
+
+        Returns:
+            Iterable[{key, feat, label}]
+    """
+    buf = []
+    for sample in data:
+        buf.append(sample)
+        if len(buf) >= shuffle_size:
+            random.shuffle(buf)
+            for x in buf:
+                yield x
+            buf = []
+    # The sample left over
+    random.shuffle(buf)
+    for x in buf:
+        yield x
+
+
+def sort(data, sort_size=500, mode='train'):
+    """ Sort the data by feature length.
+        Sort is used after shuffle and before batch, so we can group
+        utts with similar lengths into a batch, and `sort_size` should
+        be less than `shuffle_size`
+
+        Args:
+            data: Iterable[{key, feat, label}]
+            sort_size: buffer size for sort
+
+        Returns:
+            Iterable[{key, feat, label}]
+    """
+
+    buf = []
+    for sample in data:
+        buf.append(sample)
+        if len(buf) >= sort_size:
+            buf.sort(key=lambda x: x['speech_feat'].size(0))
+            for x in buf:
+                yield x
+            buf = []
+    # The sample left over
+    buf.sort(key=lambda x: x['speech_feat'].size(0))
+    for x in buf:
+        yield x
+
+
+def static_batch(data, batch_size=16):
+    """ Static batch the data by `batch_size`
+
+        Args:
+            data: Iterable[{key, feat, label}]
+            batch_size: batch size
+
+        Returns:
+            Iterable[List[{key, feat, label}]]
+    """
+    buf = []
+    for sample in data:
+        buf.append(sample)
+        if len(buf) >= batch_size:
+            yield buf
+            buf = []
+    if len(buf) > 0:
+        yield buf
+
+
+def dynamic_batch(data, max_frames_in_batch=12000, mode='train'):
+    """ Dynamic batch the data until the total frames in batch
+        reach `max_frames_in_batch`
+
+        Args:
+            data: Iterable[{key, feat, label}]
+            max_frames_in_batch: max_frames in one batch
+
+        Returns:
+            Iterable[List[{key, feat, label}]]
+    """
+    buf = []
+    longest_frames = 0
+    for sample in data:
+        assert 'speech_feat' in sample
+        assert isinstance(sample['speech_feat'], torch.Tensor)
+        new_sample_frames = sample['speech_feat'].size(0)
+        longest_frames = max(longest_frames, new_sample_frames)
+        frames_after_padding = longest_frames * (len(buf) + 1)
+        if frames_after_padding > max_frames_in_batch:
+            yield buf
+            buf = [sample]
+            longest_frames = new_sample_frames
+        else:
+            buf.append(sample)
+    if len(buf) > 0:
+        yield buf
+
+
+def batch(data, batch_type='static', batch_size=16, max_frames_in_batch=12000, mode='train'):
+    """ Wrapper for static/dynamic batch
+    """
+    if mode == 'inference':
+        return static_batch(data, 1)
+    else:
+        if batch_type == 'static':
+            return static_batch(data, batch_size)
+        elif batch_type == 'dynamic':
+            return dynamic_batch(data, max_frames_in_batch)
+        else:
+            logging.fatal('Unsupported batch type {}'.format(batch_type))
+
+
+def padding(data, use_spk_embedding, mode='train'):
+    """ Padding the data into training data
+
+        Args:
+            data: Iterable[List[{key, feat, label}]]
+
+        Returns:
+            Iterable[Tuple(keys, feats, labels, feats lengths, label lengths)]
+    """
+    for sample in data:
+        assert isinstance(sample, list)
+        speech_feat_len = torch.tensor([x['speech_feat'].size(1) for x in sample],
+                                       dtype=torch.int32)
+        order = torch.argsort(speech_feat_len, descending=True)
+
+        utts = [sample[i]['utt'] for i in order]
+        speech_token = [torch.tensor(sample[i]['speech_token']) for i in order]
+        speech_token_len = torch.tensor([i.size(0) for i in speech_token], dtype=torch.int32)
+        speech_token = pad_sequence(speech_token,
+                                    batch_first=True,
+                                    padding_value=0)
+        speech_feat = [sample[i]['speech_feat'] for i in order]
+        speech_feat_len = torch.tensor([i.size(0) for i in speech_feat], dtype=torch.int32)
+        speech_feat = pad_sequence(speech_feat,
+                                   batch_first=True,
+                                   padding_value=0)
+        text = [sample[i]['text'] for i in order]
+        text_token = [torch.tensor(sample[i]['text_token']) for i in order]
+        text_token_len = torch.tensor([i.size(0) for i in text_token], dtype=torch.int32)
+        text_token = pad_sequence(text_token, batch_first=True, padding_value=0)
+        utt_embedding = torch.stack([sample[i]['utt_embedding'] for i in order], dim=0)
+        spk_embedding = torch.stack([sample[i]['spk_embedding'] for i in order], dim=0)
+        batch = {
+            "utts": utts,
+            "speech_token": speech_token,
+            "speech_token_len": speech_token_len,
+            "speech_feat": speech_feat,
+            "speech_feat_len": speech_feat_len,
+            "text": text,
+            "text_token": text_token,
+            "text_token_len": text_token_len,
+            "utt_embedding": utt_embedding,
+            "spk_embedding": spk_embedding,
+        }
+        if mode == 'inference':
+            tts_text = [sample[i]['tts_text'] for i in order]
+            tts_index = [sample[i]['tts_index'] for i in order]
+            tts_text_token = [torch.tensor(sample[i]['tts_text_token']) for i in order]
+            tts_text_token_len = torch.tensor([i.size(0) for i in tts_text_token], dtype=torch.int32)
+            tts_text_token = pad_sequence(tts_text_token, batch_first=True, padding_value=-1)
+            batch.update({'tts_text': tts_text,
+                          'tts_index': tts_index,
+                          'tts_text_token': tts_text_token,
+                          'tts_text_token_len': tts_text_token_len})
+        if use_spk_embedding is True:
+            batch["embedding"] = batch["spk_embedding"]
+        else:
+            batch["embedding"] = batch["utt_embedding"]
+        yield batch

cosyvoice/flow/decoder.py

@@ -0,0 +1,222 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn as nn
+from einops import pack, rearrange, repeat
+from matcha.models.components.decoder import SinusoidalPosEmb, Block1D, ResnetBlock1D, Downsample1D, TimestepEmbedding, Upsample1D
+from matcha.models.components.transformer import BasicTransformerBlock
+
+
+class ConditionalDecoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        channels=(256, 256),
+        dropout=0.05,
+        attention_head_dim=64,
+        n_blocks=1,
+        num_mid_blocks=2,
+        num_heads=4,
+        act_fn="snake",
+    ):
+        """
+        This decoder requires an input with the same shape of the target. So, if your text content
+        is shorter or longer than the outputs, please re-sampling it before feeding to the decoder.
+        """
+        super().__init__()
+        channels = tuple(channels)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.time_embeddings = SinusoidalPosEmb(in_channels)
+        time_embed_dim = channels[0] * 4
+        self.time_mlp = TimestepEmbedding(
+            in_channels=in_channels,
+            time_embed_dim=time_embed_dim,
+            act_fn="silu",
+        )
+        self.down_blocks = nn.ModuleList([])
+        self.mid_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        output_channel = in_channels
+        for i in range(len(channels)):  # pylint: disable=consider-using-enumerate
+            input_channel = output_channel
+            output_channel = channels[i]
+            is_last = i == len(channels) - 1
+            resnet = ResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            downsample = (
+                Downsample1D(output_channel) if not is_last else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+            self.down_blocks.append(nn.ModuleList([resnet, transformer_blocks, downsample]))
+
+        for i in range(num_mid_blocks):
+            input_channel = channels[-1]
+            out_channels = channels[-1]
+            resnet = ResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
+
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+
+            self.mid_blocks.append(nn.ModuleList([resnet, transformer_blocks]))
+
+        channels = channels[::-1] + (channels[0],)
+        for i in range(len(channels) - 1):
+            input_channel = channels[i] * 2
+            output_channel = channels[i + 1]
+            is_last = i == len(channels) - 2
+            resnet = ResnetBlock1D(
+                dim=input_channel,
+                dim_out=output_channel,
+                time_emb_dim=time_embed_dim,
+            )
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            upsample = (
+                Upsample1D(output_channel, use_conv_transpose=True)
+                if not is_last
+                else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+            self.up_blocks.append(nn.ModuleList([resnet, transformer_blocks, upsample]))
+        self.final_block = Block1D(channels[-1], channels[-1])
+        self.final_proj = nn.Conv1d(channels[-1], self.out_channels, 1)
+        self.initialize_weights()
+
+
+    def initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.GroupNorm):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, x, mask, mu, t, spks=None, cond=None):
+        """Forward pass of the UNet1DConditional model.
+
+        Args:
+            x (torch.Tensor): shape (batch_size, in_channels, time)
+            mask (_type_): shape (batch_size, 1, time)
+            t (_type_): shape (batch_size)
+            spks (_type_, optional): shape: (batch_size, condition_channels). Defaults to None.
+            cond (_type_, optional): placeholder for future use. Defaults to None.
+
+        Raises:
+            ValueError: _description_
+            ValueError: _description_
+
+        Returns:
+            _type_: _description_
+        """
+
+        t = self.time_embeddings(t)
+        t = self.time_mlp(t)
+
+        x = pack([x, mu], "b * t")[0]
+
+        if spks is not None:
+            spks = repeat(spks, "b c -> b c t", t=x.shape[-1])
+            x = pack([x, spks], "b * t")[0]
+        if cond is not None:
+            x = pack([x, cond], "b * t")[0]
+
+        hiddens = []
+        masks = [mask]
+        for resnet, transformer_blocks, downsample in self.down_blocks:
+            mask_down = masks[-1]
+            x = resnet(x, mask_down, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            attn_mask = torch.matmul(mask_down.transpose(1, 2).contiguous(), mask_down)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            hiddens.append(x)  # Save hidden states for skip connections
+            x = downsample(x * mask_down)
+            masks.append(mask_down[:, :, ::2])
+        masks = masks[:-1]
+        mask_mid = masks[-1]
+
+        for resnet, transformer_blocks in self.mid_blocks:
+            x = resnet(x, mask_mid, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            attn_mask = torch.matmul(mask_mid.transpose(1, 2).contiguous(), mask_mid)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+
+        for resnet, transformer_blocks, upsample in self.up_blocks:
+            mask_up = masks.pop()
+            skip = hiddens.pop()
+            x = pack([x[:, :, :skip.shape[-1]], skip], "b * t")[0]
+            x = resnet(x, mask_up, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            attn_mask = torch.matmul(mask_up.transpose(1, 2).contiguous(), mask_up)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            x = upsample(x * mask_up)
+        x = self.final_block(x, mask_up)
+        output = self.final_proj(x * mask_up)
+        return output * mask

cosyvoice/flow/flow.py

@@ -0,0 +1,141 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+import random
+from typing import Dict, Optional
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from omegaconf import DictConfig
+from cosyvoice.utils.mask import make_pad_mask
+
+
+class MaskedDiffWithXvec(torch.nn.Module):
+    def __init__(self,
+                 input_size: int = 512,
+                 output_size: int = 80,
+                 spk_embed_dim: int = 192,
+                 output_type: str = "mel",
+                 vocab_size: int = 4096,
+                 input_frame_rate: int = 50,
+                 only_mask_loss: bool = True,
+                 encoder: torch.nn.Module = None,
+                 length_regulator: torch.nn.Module = None,
+                 decoder: torch.nn.Module = None,
+                 decoder_conf: Dict = {'in_channels': 240, 'out_channel': 80, 'spk_emb_dim': 80, 'n_spks': 1, 'cfm_params': DictConfig({'sigma_min': 1e-06, 'solver': 'euler', 't_scheduler': 'cosine', 'training_cfg_rate': 0.2, 'inference_cfg_rate': 0.7, 'reg_loss_type': 'l1'}), 'decoder_params': {'channels': [256, 256], 'dropout': 0.0, 'attention_head_dim': 64, 'n_blocks': 4, 'num_mid_blocks': 12, 'num_heads': 8, 'act_fn': 'gelu'}},
+                 mel_feat_conf: Dict = {'n_fft': 1024, 'num_mels': 80, 'sampling_rate': 22050, 'hop_size': 256, 'win_size': 1024, 'fmin': 0, 'fmax': 8000}):
+        super().__init__()
+        self.input_size = input_size
+        self.output_size = output_size
+        self.decoder_conf = decoder_conf
+        self.mel_feat_conf = mel_feat_conf
+        self.vocab_size = vocab_size
+        self.output_type = output_type
+        self.input_frame_rate = input_frame_rate
+        logging.info(f"input frame rate={self.input_frame_rate}")
+        self.input_embedding = nn.Embedding(vocab_size, input_size)
+        self.spk_embed_affine_layer = torch.nn.Linear(spk_embed_dim, output_size)
+        self.encoder = encoder
+        self.encoder_proj = torch.nn.Linear(self.encoder.output_size(), output_size)
+        self.decoder = decoder
+        self.length_regulator = length_regulator
+        self.only_mask_loss = only_mask_loss
+
+    def forward(
+            self,
+            batch: dict,
+            device: torch.device,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        token = batch['speech_token'].to(device)
+        token_len = batch['speech_token_len'].to(device)
+        feat = batch['speech_feat'].to(device)
+        feat_len = batch['speech_feat_len'].to(device)
+        embedding = batch['embedding'].to(device)
+
+        # xvec projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+
+        # concat text and prompt_text
+        mask = (~make_pad_mask(token_len)).float().unsqueeze(-1).to(device)
+        token = self.input_embedding(torch.clamp(token, min=0)) * mask
+
+        # text encode
+        h, h_lengths = self.encoder(token, token_len)
+        h = self.encoder_proj(h)
+        h, h_lengths = self.length_regulator(h, feat_len)
+
+        # get conditions
+        conds = torch.zeros(feat.shape, device=token.device)
+        for i, j in enumerate(feat_len):
+            if random.random() < 0.5:
+                continue
+            index = random.randint(0, int(0.3 * j))
+            conds[i, :index] = feat[i, :index]
+        conds = conds.transpose(1, 2)
+
+        mask = (~make_pad_mask(feat_len)).to(h)
+        feat = F.interpolate(feat.unsqueeze(dim=1), size=h.shape[1:], mode="nearest").squeeze(dim=1)
+        loss, _ = self.decoder.compute_loss(
+            feat.transpose(1, 2).contiguous(),
+            mask.unsqueeze(1),
+            h.transpose(1, 2).contiguous(),
+            embedding,
+            cond=conds
+        )
+        return {'loss': loss}
+
+    @torch.inference_mode()
+    def inference(self,
+                  token,
+                  token_len,
+                  prompt_token,
+                  prompt_token_len,
+                  prompt_feat,
+                  prompt_feat_len,
+                  embedding):
+        assert token.shape[0] == 1
+        # xvec projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+
+        # concat text and prompt_text
+        token, token_len = torch.concat([prompt_token, token], dim=1), prompt_token_len + token_len
+        mask = (~make_pad_mask(token_len)).float().unsqueeze(-1).to(embedding)
+        token = self.input_embedding(torch.clamp(token, min=0)) * mask
+
+        # text encode
+        h, h_lengths = self.encoder(token, token_len)
+        h = self.encoder_proj(h)
+        feat_len = (token_len / 50 * 22050 / 256).int()
+        h, h_lengths = self.length_regulator(h, feat_len)
+
+        # get conditions
+        conds = torch.zeros([1, feat_len.max().item(), self.output_size], device=token.device)
+        if prompt_feat.shape[1] != 0:
+            for i, j in enumerate(prompt_feat_len):
+                conds[i, :j] = prompt_feat[i]
+        conds = conds.transpose(1, 2)
+
+        mask = (~make_pad_mask(feat_len)).to(h)
+        feat = self.decoder(
+            mu=h.transpose(1, 2).contiguous(),
+            mask=mask.unsqueeze(1),
+            spks=embedding,
+            cond=conds,
+            n_timesteps=10
+        )
+        if prompt_feat.shape[1] != 0:
+            feat = feat[:, :, prompt_feat.shape[1]:]
+        return feat

cosyvoice/flow/flow_matching.py

@@ -0,0 +1,138 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn.functional as F
+from matcha.models.components.flow_matching import BASECFM
+
+class ConditionalCFM(BASECFM):
+    def __init__(self, in_channels, cfm_params, n_spks=1, spk_emb_dim=64, estimator: torch.nn.Module = None):
+        super().__init__(
+            n_feats=in_channels,
+            cfm_params=cfm_params,
+            n_spks=n_spks,
+            spk_emb_dim=spk_emb_dim,
+        )
+        self.t_scheduler = cfm_params.t_scheduler
+        self.training_cfg_rate = cfm_params.training_cfg_rate
+        self.inference_cfg_rate = cfm_params.inference_cfg_rate
+        in_channels = in_channels + (spk_emb_dim if n_spks > 0 else 0)
+        # Just change the architecture of the estimator here
+        self.estimator = estimator
+
+    @torch.inference_mode()
+    def forward(self, mu, mask, n_timesteps, temperature=1.0, spks=None, cond=None):
+        """Forward diffusion
+
+        Args:
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            n_timesteps (int): number of diffusion steps
+            temperature (float, optional): temperature for scaling noise. Defaults to 1.0.
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+
+        Returns:
+            sample: generated mel-spectrogram
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        z = torch.randn_like(mu) * temperature
+        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device)
+        if self.t_scheduler == 'cosine':
+            t_span = 1 - torch.cos(t_span * 0.5 * torch.pi)
+        return self.solve_euler(z, t_span=t_span, mu=mu, mask=mask, spks=spks, cond=cond)
+
+    def solve_euler(self, x, t_span, mu, mask, spks, cond):
+        """
+        Fixed euler solver for ODEs.
+        Args:
+            x (torch.Tensor): random noise
+            t_span (torch.Tensor): n_timesteps interpolated
+                shape: (n_timesteps + 1,)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+        """
+        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]
+
+        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
+        # Or in future might add like a return_all_steps flag
+        sol = []
+
+        for step in range(1, len(t_span)):
+            dphi_dt = self.estimator(x, mask, mu, t, spks, cond)
+            # Classifier-Free Guidance inference introduced in VoiceBox
+            if self.inference_cfg_rate > 0:
+                cfg_dphi_dt = self.estimator(
+                    x, mask,
+                    torch.zeros_like(mu), t,
+                    torch.zeros_like(spks) if spks is not None else None,
+                    torch.zeros_like(cond)
+                )
+                dphi_dt = ((1.0 + self.inference_cfg_rate) * dphi_dt -
+                           self.inference_cfg_rate * cfg_dphi_dt)
+            x = x + dt * dphi_dt
+            t = t + dt
+            sol.append(x)
+            if step < len(t_span) - 1:
+                dt = t_span[step + 1] - t
+
+        return sol[-1]
+
+    def compute_loss(self, x1, mask, mu, spks=None, cond=None):
+        """Computes diffusion loss
+
+        Args:
+            x1 (torch.Tensor): Target
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): target mask
+                shape: (batch_size, 1, mel_timesteps)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            spks (torch.Tensor, optional): speaker embedding. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+
+        Returns:
+            loss: conditional flow matching loss
+            y: conditional flow
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        b, _, t = mu.shape
+
+        # random timestep
+        t = torch.rand([b, 1, 1], device=mu.device, dtype=mu.dtype)
+        if self.t_scheduler == 'cosine':
+            t = 1 - torch.cos(t * 0.5 * torch.pi)
+        # sample noise p(x_0)
+        z = torch.randn_like(x1)
+
+        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
+        u = x1 - (1 - self.sigma_min) * z
+
+        # during training, we randomly drop condition to trade off mode coverage and sample fidelity
+        if self.training_cfg_rate > 0:
+            cfg_mask = torch.rand(b, device=x1.device) > self.training_cfg_rate
+            mu = mu * cfg_mask.view(-1, 1, 1)
+            spks = spks * cfg_mask.view(-1, 1)
+            cond = cond * cfg_mask.view(-1, 1, 1)
+
+        pred = self.estimator(y, mask, mu, t.squeeze(), spks, cond)
+        loss = F.mse_loss(pred * mask, u * mask, reduction="sum") / (torch.sum(mask) * u.shape[1])
+        return loss, y

cosyvoice/flow/length_regulator.py

@@ -0,0 +1,49 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Tuple
+import torch.nn as nn
+from torch.nn import functional as F
+from cosyvoice.utils.mask import make_pad_mask
+
+
+class InterpolateRegulator(nn.Module):
+    def __init__(
+            self,
+            channels: int,
+            sampling_ratios: Tuple,
+            out_channels: int = None,
+            groups: int = 1,
+    ):
+        super().__init__()
+        self.sampling_ratios = sampling_ratios
+        out_channels = out_channels or channels
+        model = nn.ModuleList([])
+        if len(sampling_ratios) > 0:
+            for _ in sampling_ratios:
+                module = nn.Conv1d(channels, channels, 3, 1, 1)
+                norm = nn.GroupNorm(groups, channels)
+                act = nn.Mish()
+                model.extend([module, norm, act])
+        model.append(
+            nn.Conv1d(channels, out_channels, 1, 1)
+        )
+        self.model = nn.Sequential(*model)
+
+    def forward(self, x, ylens=None):
+        # x in (B, T, D)
+        mask = (~make_pad_mask(ylens)).to(x).unsqueeze(-1)
+        x = F.interpolate(x.transpose(1, 2).contiguous(), size=ylens.max(), mode='nearest')
+        out = self.model(x).transpose(1, 2).contiguous()
+        olens = ylens
+        return out * mask, olens

cosyvoice/hifigan/f0_predictor.py

@@ -0,0 +1,55 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Kai Hu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn as nn
+from torch.nn.utils import weight_norm
+
+
+class ConvRNNF0Predictor(nn.Module):
+    def __init__(self,
+                 num_class: int = 1,
+                 in_channels: int = 80,
+                 cond_channels: int = 512
+                 ):
+        super().__init__()
+
+        self.num_class = num_class
+        self.condnet = nn.Sequential(
+            weight_norm(
+                nn.Conv1d(in_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+        )
+        self.classifier = nn.Linear(in_features=cond_channels, out_features=self.num_class)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.condnet(x)
+        x = x.transpose(1, 2)
+        return torch.abs(self.classifier(x).squeeze(-1))

cosyvoice/hifigan/generator.py

@@ -0,0 +1,391 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Kai Hu)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""HIFI-GAN"""
+
+import typing as tp
+import numpy as np
+from scipy.signal import get_window
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import Conv1d
+from torch.nn import ConvTranspose1d
+from torch.nn.utils import remove_weight_norm
+from torch.nn.utils import weight_norm
+from torch.distributions.uniform import Uniform
+
+from cosyvoice.transformer.activation import Snake
+from cosyvoice.utils.common import get_padding
+from cosyvoice.utils.common import init_weights
+
+
+"""hifigan based generator implementation.
+
+This code is modified from https://github.com/jik876/hifi-gan
+ ,https://github.com/kan-bayashi/ParallelWaveGAN and
+ https://github.com/NVIDIA/BigVGAN
+
+"""
+class ResBlock(torch.nn.Module):
+    """Residual block module in HiFiGAN/BigVGAN."""
+    def __init__(
+        self,
+        channels: int = 512,
+        kernel_size: int = 3,
+        dilations: tp.List[int] = [1, 3, 5],
+    ):
+        super(ResBlock, self).__init__()
+        self.convs1 = nn.ModuleList()
+        self.convs2 = nn.ModuleList()
+
+        for dilation in dilations:
+            self.convs1.append(
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation,
+                        padding=get_padding(kernel_size, dilation)
+                    )
+                )
+            )
+            self.convs2.append(
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1)
+                    )
+                )
+            )
+        self.convs1.apply(init_weights)
+        self.convs2.apply(init_weights)
+        self.activations1 = nn.ModuleList([
+            Snake(channels, alpha_logscale=False)
+            for _ in range(len(self.convs1))
+        ])
+        self.activations2 = nn.ModuleList([
+            Snake(channels, alpha_logscale=False)
+            for _ in range(len(self.convs2))
+        ])
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        for idx in range(len(self.convs1)):
+            xt = self.activations1[idx](x)
+            xt = self.convs1[idx](xt)
+            xt = self.activations2[idx](xt)
+            xt = self.convs2[idx](xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for idx in range(len(self.convs1)):
+            remove_weight_norm(self.convs1[idx])
+            remove_weight_norm(self.convs2[idx])
+
+class SineGen(torch.nn.Module):
+    """ Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(self, samp_rate, harmonic_num=0,
+                 sine_amp=0.1, noise_std=0.003,
+                 voiced_threshold=0):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = (f0 > self.voiced_threshold).type(torch.float32)
+        return uv
+
+    @torch.no_grad()
+    def forward(self, f0):
+        """
+        :param f0: [B, 1, sample_len], Hz
+        :return: [B, 1, sample_len]
+        """
+
+        F_mat = torch.zeros((f0.size(0), self.harmonic_num + 1, f0.size(-1))).to(f0.device)
+        for i in range(self.harmonic_num + 1):
+            F_mat[:, i: i + 1, :] = f0 * (i + 1) / self.sampling_rate
+
+        theta_mat = 2 * np.pi * (torch.cumsum(F_mat, dim=-1) % 1)
+        u_dist = Uniform(low=-np.pi, high=np.pi)
+        phase_vec = u_dist.sample(sample_shape=(f0.size(0), self.harmonic_num + 1, 1)).to(F_mat.device)
+        phase_vec[:, 0, :] = 0
+
+        # generate sine waveforms
+        sine_waves = self.sine_amp * torch.sin(theta_mat + phase_vec)
+
+        # generate uv signal
+        uv = self._f02uv(f0)
+
+        # noise: for unvoiced should be similar to sine_amp
+        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
+        # .       for voiced regions is self.noise_std
+        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+        noise = noise_amp * torch.randn_like(sine_waves)
+
+        # first: set the unvoiced part to 0 by uv
+        # then: additive noise
+        sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """ SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(sampling_rate, harmonic_num,
+                                 sine_amp, add_noise_std, voiced_threshod)
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x):
+        """
+        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+        F0_sampled (batchsize, length, 1)
+        Sine_source (batchsize, length, 1)
+        noise_source (batchsize, length 1)
+        """
+        # source for harmonic branch
+        with torch.no_grad():
+            sine_wavs, uv, _ = self.l_sin_gen(x.transpose(1, 2))
+            sine_wavs = sine_wavs.transpose(1, 2)
+            uv = uv.transpose(1, 2)
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+
+        # source for noise branch, in the same shape as uv
+        noise = torch.randn_like(uv) * self.sine_amp / 3
+        return sine_merge, noise, uv
+
+
+class HiFTGenerator(nn.Module):
+    """
+    HiFTNet Generator: Neural Source Filter + ISTFTNet
+    https://arxiv.org/abs/2309.09493
+    """
+    def __init__(
+            self,
+            in_channels: int = 80,
+            base_channels: int = 512,
+            nb_harmonics: int = 8,
+            sampling_rate: int = 22050,
+            nsf_alpha: float = 0.1,
+            nsf_sigma: float = 0.003,
+            nsf_voiced_threshold: float = 10,
+            upsample_rates: tp.List[int] = [8, 8],
+            upsample_kernel_sizes: tp.List[int] = [16, 16],
+            istft_params: tp.Dict[str, int] = {"n_fft": 16, "hop_len": 4},
+            resblock_kernel_sizes: tp.List[int] = [3, 7, 11],
+            resblock_dilation_sizes: tp.List[tp.List[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            source_resblock_kernel_sizes: tp.List[int] = [7, 11],
+            source_resblock_dilation_sizes: tp.List[tp.List[int]] = [[1, 3, 5], [1, 3, 5]],
+            lrelu_slope: float = 0.1,
+            audio_limit: float = 0.99,
+            f0_predictor: torch.nn.Module = None,
+    ):
+        super(HiFTGenerator, self).__init__()
+
+        self.out_channels = 1
+        self.nb_harmonics = nb_harmonics
+        self.sampling_rate = sampling_rate
+        self.istft_params = istft_params
+        self.lrelu_slope = lrelu_slope
+        self.audio_limit = audio_limit
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sampling_rate,
+            upsample_scale=np.prod(upsample_rates) * istft_params["hop_len"],
+            harmonic_num=nb_harmonics,
+            sine_amp=nsf_alpha,
+            add_noise_std=nsf_sigma,
+            voiced_threshod=nsf_voiced_threshold)
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * istft_params["hop_len"])
+
+        self.conv_pre = weight_norm(
+            Conv1d(in_channels, base_channels, 7, 1, padding=3)
+        )
+
+        # Up
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        base_channels // (2**i),
+                        base_channels // (2**(i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        # Down
+        self.source_downs = nn.ModuleList()
+        self.source_resblocks = nn.ModuleList()
+        downsample_rates = [1] + upsample_rates[::-1][:-1]
+        downsample_cum_rates = np.cumprod(downsample_rates)
+        for i, (u, k, d) in enumerate(zip(downsample_cum_rates[::-1], source_resblock_kernel_sizes,
+                                          source_resblock_dilation_sizes)):
+            if u == 1:
+                self.source_downs.append(
+                    Conv1d(istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), 1, 1)
+                )
+            else:
+                self.source_downs.append(
+                    Conv1d(istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), u * 2, u, padding=(u // 2))
+                )
+
+            self.source_resblocks.append(
+                ResBlock(base_channels // (2 ** (i + 1)), k, d)
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = base_channels // (2**(i + 1))
+            for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+                self.resblocks.append(ResBlock(ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, istft_params["n_fft"] + 2, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+        self.reflection_pad = nn.ReflectionPad1d((1, 0))
+        self.stft_window = torch.from_numpy(get_window("hann", istft_params["n_fft"], fftbins=True).astype(np.float32))
+        self.f0_predictor = f0_predictor
+
+    def _f02source(self, f0: torch.Tensor) -> torch.Tensor:
+        f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
+
+        har_source, _, _ = self.m_source(f0)
+        return har_source.transpose(1, 2)
+
+    def _stft(self, x):
+        spec = torch.stft(
+            x,
+            self.istft_params["n_fft"], self.istft_params["hop_len"], self.istft_params["n_fft"], window=self.stft_window.to(x.device),
+            return_complex=True)
+        spec = torch.view_as_real(spec)  # [B, F, TT, 2]
+        return spec[..., 0], spec[..., 1]
+
+    def _istft(self, magnitude, phase):
+        magnitude = torch.clip(magnitude, max=1e2)
+        real = magnitude * torch.cos(phase)
+        img = magnitude * torch.sin(phase)
+        inverse_transform = torch.istft(torch.complex(real, img), self.istft_params["n_fft"], self.istft_params["hop_len"], self.istft_params["n_fft"], window=self.stft_window.to(magnitude.device))
+        return inverse_transform
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        f0 = self.f0_predictor(x)
+        s = self._f02source(f0)
+
+        s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
+        s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)
+
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, self.lrelu_slope)
+            x = self.ups[i](x)
+
+            if i == self.num_upsamples - 1:
+                x = self.reflection_pad(x)
+
+            # fusion
+            si = self.source_downs[i](s_stft)
+            si = self.source_resblocks[i](si)
+            x = x + si
+
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        magnitude = torch.exp(x[:, :self.istft_params["n_fft"] // 2 + 1, :])
+        phase = torch.sin(x[:, self.istft_params["n_fft"] // 2 + 1:, :])  # actually, sin is redundancy
+
+        x = self._istft(magnitude, phase)
+        x = torch.clamp(x, -self.audio_limit, self.audio_limit)
+        return x
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+        self.source_module.remove_weight_norm()
+        for l in self.source_downs:
+            remove_weight_norm(l)
+        for l in self.source_resblocks:
+            l.remove_weight_norm()
+
+    @torch.inference_mode()
+    def inference(self, mel: torch.Tensor) -> torch.Tensor:
+        return self.forward(x=mel)

cosyvoice/llm/llm.py

@@ -0,0 +1,206 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Union
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn.utils.rnn import pad_sequence, unpad_sequence
+from cosyvoice.utils.common import IGNORE_ID
+from cosyvoice.transformer.label_smoothing_loss import LabelSmoothingLoss
+from cosyvoice.utils.common import th_accuracy
+
+
+class TransformerLM(torch.nn.Module):
+    def __init__(
+            self,
+            text_encoder_input_size: int,
+            llm_input_size: int,
+            llm_output_size: int,
+            text_token_size: int,
+            speech_token_size: int,
+            text_encoder: torch.nn.Module,
+            llm: torch.nn.Module,
+            length_normalized_loss: bool = True,
+            lsm_weight: float = 0.0,
+            spk_embed_dim: int = 192,
+    ):
+        super().__init__()
+        self.llm_input_size = llm_input_size
+        self.speech_token_size = speech_token_size
+        # 1. build text token inputs related modules
+        self.text_embedding = torch.nn.Embedding(text_token_size, text_encoder_input_size)
+        self.text_encoder = text_encoder
+        self.text_encoder_affine_layer = nn.Linear(
+            self.text_encoder.output_size(),
+            llm_input_size
+        )
+
+        # 2. build speech token language model related modules
+        self.sos_eos = 0
+        self.task_id = 1
+        self.llm_embedding = torch.nn.Embedding(2, llm_input_size)
+        self.llm = llm
+        self.llm_decoder = nn.Linear(llm_output_size, speech_token_size + 1)
+        self.criterion_ce = LabelSmoothingLoss(
+            size=speech_token_size + 1,
+            padding_idx=IGNORE_ID,
+            smoothing=lsm_weight,
+            normalize_length=length_normalized_loss,
+        )
+
+        # 3. [Optional] build speech token related modules
+        self.speech_embedding = torch.nn.Embedding(speech_token_size, llm_input_size)
+        self.spk_embed_affine_layer = torch.nn.Linear(spk_embed_dim, llm_input_size)
+
+    def encode(
+            self,
+            text: torch.Tensor,
+            text_lengths: torch.Tensor,
+    ):
+        encoder_out, encoder_mask = self.text_encoder(text, text_lengths, decoding_chunk_size=1, num_decoding_left_chunks=-1)
+        encoder_out_lens = encoder_mask.squeeze(1).sum(1)
+        encoder_out = self.text_encoder_affine_layer(encoder_out)
+        return encoder_out, encoder_out_lens
+
+    def pad_unpad_sequence(self, sos_eos_emb, embedding, text_token, text_token_len, task_id_emb, speech_token, speech_token_len):
+        text_token = unpad_sequence(text_token, text_token_len.cpu(), batch_first=True)
+        speech_token = unpad_sequence(speech_token, speech_token_len.cpu(), batch_first=True)
+        lm_input = [torch.concat([sos_eos_emb.squeeze(dim=0), embedding[i], text_token[i], task_id_emb.squeeze(dim=0), speech_token[i]], dim=0) for i in range(len(text_token))]
+        lm_input_len = torch.tensor([i.size(0) for i in lm_input], dtype=torch.int32)
+        lm_input = pad_sequence(lm_input, batch_first=True, padding_value=IGNORE_ID)
+        return lm_input, lm_input_len
+
+    def forward(
+            self,
+            batch: dict,
+            device: torch.device,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        """
+        Args:
+            text: (B, L, D)
+            text_lengths: (B,)
+            audio: (B, T, N) or (B, T)
+            audio_lengths: (B,)
+        """
+        text_token = batch['text_token'].to(device)
+        text_token_len = batch['text_token_len'].to(device)
+        speech_token = batch['speech_token'].to(device)
+        speech_token_len = batch['speech_token_len'].to(device)
+        embedding = batch['embedding'].to(device)
+
+        # 1. prepare llm_target
+        lm_target = [torch.tensor([IGNORE_ID] * (2 + text_token_len[i]) + speech_token[i, :speech_token_len[i]].tolist() + [self.speech_token_size]) for i in range(text_token.size(0))]
+        lm_target = pad_sequence(lm_target, batch_first=True, padding_value=IGNORE_ID).to(device)
+
+        # 1. encode text_token
+        text_token = self.text_embedding(text_token)
+        text_token, text_token_len = self.encode(text_token, text_token_len)
+
+        # 2. embedding projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+        embedding = embedding.unsqueeze(1)
+
+        # 3. eos and task_id
+        sos_eos_emb = self.llm_embedding.weight[self.sos_eos].reshape(1, 1, -1)
+        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)
+
+        # 4. encode speech_token
+        speech_token = self.speech_embedding(speech_token)
+
+        # 5. unpad and pad
+        lm_input, lm_input_len = self.pad_unpad_sequence(sos_eos_emb, embedding, text_token, text_token_len, task_id_emb, speech_token, speech_token_len)
+
+        # 6. run lm forward
+        lm_output, lm_output_mask = self.llm(lm_input, lm_input_len.to(device))
+        logits = self.llm_decoder(lm_output)
+        loss = self.criterion_ce(logits, lm_target)
+        acc = th_accuracy(logits.view(-1, self.speech_token_size + 1), lm_target, ignore_label=IGNORE_ID)
+        return {'loss': loss, 'acc': acc}
+
+    def sampling_ids(
+            self,
+            weighted_scores: torch.Tensor,
+            sampling: Union[bool, int, float] = True,
+            beam_size: int = 1,
+            ignore_eos: bool = True,
+    ):
+        while True:
+            prob, indices = weighted_scores.softmax(dim=-1).topk(sampling)
+            top_ids = prob.multinomial(beam_size, replacement=True)
+            top_ids = indices[top_ids]
+            if (not ignore_eos) or (self.speech_token_size not in top_ids):
+                break
+        return top_ids
+
+    @torch.inference_mode()
+    def inference(
+            self,
+            text: torch.Tensor,
+            text_len: torch.Tensor,
+            prompt_text: torch.Tensor,
+            prompt_text_len: torch.Tensor,
+            prompt_speech_token: torch.Tensor,
+            prompt_speech_token_len: torch.Tensor,
+            embedding: torch.Tensor,
+            beam_size: int = 1,
+            sampling: int = 25,
+            max_token_text_ratio: float = 20,
+            min_token_text_ratio: float = 2,
+    ) -> torch.Tensor:
+        device = text.device
+        text = torch.concat([prompt_text, text], dim=1)
+        text_len += prompt_text_len
+        text = self.text_embedding(text)
+
+        # 1. encode text
+        text, text_len = self.encode(text, text_len)
+
+        # 2. encode embedding
+        if embedding.shape[0] != 0:
+            embedding = F.normalize(embedding, dim=1)
+            embedding = self.spk_embed_affine_layer(embedding)
+            embedding = embedding.unsqueeze(dim=1)
+        else:
+            embedding = torch.zeros(1, 0, self.llm_input_size).to(device)
+
+        # 3. concat llm_input
+        sos_eos_emb = self.llm_embedding.weight[self.sos_eos].reshape(1, 1, -1)
+        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)
+        if prompt_speech_token_len != 0:
+            prompt_speech_token_emb = self.speech_embedding(prompt_speech_token)
+        else:
+            prompt_speech_token_emb = torch.zeros(1, 0, self.llm_input_size).to(device)
+        lm_input = torch.concat([sos_eos_emb, embedding, text, task_id_emb, prompt_speech_token_emb], dim=1)
+
+        # 4. cal min/max_length
+        min_len = int((text_len - prompt_text_len) * min_token_text_ratio)
+        max_len = int((text_len - prompt_text_len) * max_token_text_ratio)
+
+        # 5. step by step decode
+        out_tokens = []
+        offset = 0
+        att_cache, cnn_cache = torch.zeros((0, 0, 0, 0), device=lm_input.device), torch.zeros((0, 0, 0, 0), device=lm_input.device)
+        for i in range(max_len):
+            y_pred, att_cache, cnn_cache = self.llm.forward_chunk(lm_input, offset=0, required_cache_size=-1, att_cache=att_cache, cnn_cache=cnn_cache,
+                                                                  att_mask=torch.tril(torch.ones((1, lm_input.shape[1], lm_input.shape[1]), device=lm_input.device)).to(torch.bool))
+            logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1)
+            top_ids = self.sampling_ids(logp.squeeze(dim=0), sampling, beam_size, ignore_eos=True if i < min_len else False).item()
+            if top_ids == self.speech_token_size:
+                break
+            out_tokens.append(top_ids)
+            offset += lm_input.size(1)
+            lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1)
+
+        return torch.tensor([out_tokens], dtype=torch.int64, device=device)