ChatTTS（Chat Text To Speech），專為對話場景設計的文本生成語音(TTS)模型，適用于大型語言模型(LLM)助手的對話任務，以及諸如對話式音頻和視頻介紹等應用。支持中文和英文，還可以穿插笑聲、說話間的停頓、以及語氣詞等。

1 下載模型

通過git-lfs工具包下載：

sudo apt install git-lfs

下載文件：

git lfs install
git clone https://www.modelscope.cn/pzc163/chatTTS.git ChatTTS-Model

如果因網絡不佳下載中斷，可以通過以下命令在中斷后繼續下載：

git lfs pull

2 安裝 ChatTTS 依賴包

下載ChatTTS官網GitHub源碼：

git clone https://gitcode.com/2noise/ChatTTS.git ChatTTS

安裝Python依賴包:

cd ChatTTS
pip install -r requirements.txt

torch需要使用2.2.2

3 ChatTTS 中文文本轉音頻文件

ChatTTS 官網的樣例代碼 API 可能會跑不起來。

ChatTTS-01.py

import ChatTTS
import torch
import torchaudio# 下載的ChatTTS模型文件目錄，請按照實際情況替換
MODEL_PATH = '/path/to/ChatTTS-Model'# 初始化并加載模型，特別注意加載模型參數，官網樣例代碼已經過時
chat = ChatTTS.Chat()
chat.load_models(vocos_config_path=f'{MODEL_PATH}/config/vocos.yaml',vocos_ckpt_path=f'{MODEL_PATH}/asset/Vocos.pt',gpt_config_path=f'{MODEL_PATH}/config/gpt.yaml',gpt_ckpt_path=f'{MODEL_PATH}/asset/GPT.pt',decoder_config_path=f'{MODEL_PATH}/config/decoder.yaml',decoder_ckpt_path=f'{MODEL_PATH}/asset/Decoder.pt',tokenizer_path=f'{MODEL_PATH}/asset/tokenizer.pt',
)# 需要轉化為音頻的文本內容
text = '中文文本'# 文本轉為音頻
wavs = chat.infer(text, use_decoder=True)# 保存音頻文件到本地文件（采樣率為24000Hz）
torchaudio.save("./output/output-01.wav", torch.from_numpy(wavs[0]), 24000)

運行Python代碼：python ChatTTS-01.py

也可以在文本轉換成語音之后，直接播放語音內容：

from IPython.display import Audio# 播放生成的音頻（autoplay=True 代表自動播放）
Audio(wavs[0], rate=24000, autoplay=True)

4 搭建 WebUI 界面

4.1 安裝 Python 依賴包

pip install omegaconf~=2.3.0 transformers~=4.41.1
pip install tqdm einops vector_quantize_pytorch vocos
pip install modelscope gradio

運行 Python 程序，即可看到可視化界面，可以隨意輸入文本來生成音頻文件了。

ChatTTS-WebUI.py

import randomimport ChatTTS
import gradio as gr
import numpy as np
import torch
from ChatTTS.infer.api import refine_text, infer_codeprint('啟動ChatTTS WebUI......')# WebUI設置
WEB_HOST = '127.0.0.1'
WEB_PORT = 8089MODEL_PATH = '/path/to/ChatTTS-Model'chat = ChatTTS.Chat()
chat.load_models(vocos_config_path=f'{MODEL_PATH}/config/vocos.yaml',vocos_ckpt_path=f'{MODEL_PATH}/asset/Vocos.pt',gpt_config_path=f'{MODEL_PATH}/config/gpt.yaml',gpt_ckpt_path=f'{MODEL_PATH}/asset/GPT.pt',decoder_config_path=f'{MODEL_PATH}/config/decoder.yaml',decoder_ckpt_path=f'{MODEL_PATH}/asset/Decoder.pt',tokenizer_path=f'{MODEL_PATH}/asset/tokenizer.pt',
)def generate_seed():new_seed = random.randint(1, 100000000)return {"__type__": "update","value": new_seed}def generate_audio(text, temperature, top_P, top_K, audio_seed_input, text_seed_input, refine_text_flag):torch.manual_seed(audio_seed_input)rand_spk = torch.randn(768)params_infer_code = {'spk_emb': rand_spk,'temperature': temperature,'top_P': top_P,'top_K': top_K,}params_refine_text = {'prompt': '[oral_2][laugh_0][break_6]'}torch.manual_seed(text_seed_input)text_tokens = refine_text(chat.pretrain_models, text, **params_refine_text)['ids']text_tokens = [i[i < chat.pretrain_models['tokenizer'].convert_tokens_to_ids('[break_0]')] for i in text_tokens]text = chat.pretrain_models['tokenizer'].batch_decode(text_tokens)# result = infer_code(chat.pretrain_models, text, **params_infer_code, return_hidden=True)print(f'ChatTTS微調文本：{text}')wav = chat.infer(text,params_refine_text=params_refine_text,params_infer_code=params_infer_code,use_decoder=True,skip_refine_text=True,)audio_data = np.array(wav[0]).flatten()sample_rate = 24000text_data = text[0] if isinstance(text, list) else textreturn [(sample_rate, audio_data), text_data]def main():with gr.Blocks() as demo:default_text = "文字"text_input = gr.Textbox(label="輸入文本", lines=4, placeholder="Please Input Text...", value=default_text)with gr.Row():refine_text_checkbox = gr.Checkbox(label="文本微調開關", value=True)temperature_slider = gr.Slider(minimum=0.00001, maximum=1.0, step=0.00001, value=0.8, label="語音溫度參數")top_p_slider = gr.Slider(minimum=0.1, maximum=0.9, step=0.05, value=0.7, label="語音top_P采樣參數")top_k_slider = gr.Slider(minimum=1, maximum=20, step=1, value=20, label="語音top_K采樣參數")with gr.Row():audio_seed_input = gr.Number(value=42, label="語音隨機數")generate_audio_seed = gr.Button("\U0001F3B2")text_seed_input = gr.Number(value=42, label="文本隨機數")generate_text_seed = gr.Button("\U0001F3B2")generate_button = gr.Button("文本生成語音")text_output = gr.Textbox(label="微調文本", interactive=False)audio_output = gr.Audio(label="語音")generate_audio_seed.click(generate_seed,inputs=[],outputs=audio_seed_input)generate_text_seed.click(generate_seed,inputs=[],outputs=text_seed_input)generate_button.click(generate_audio,inputs=[text_input, temperature_slider, top_p_slider, top_k_slider, audio_seed_input, text_seed_input, refine_text_checkbox],outputs=[audio_output, text_output, ])# 啟動WebUIdemo.launch(server_name='127.0.0.1', server_port=8089, share=False, show_api=False, )if __name__ == '__main__':main()