ChatTTS 开源文本转语音模型本地部署 API 使用和搭建 WebUI 界面

ChatTTS（Chat Text To Speech），专为对话场景设计的文本生成语音(TTS)模型，适用于大型语言模型(LLM)助手的对话任务，以及诸如对话式音频和视频介绍等应用。支持中文和英文，还可以穿插笑声、说话间的停顿、以及语气词等。

1 下载模型

通过git-lfs工具包下载：

sudo apt install git-lfs

下载文件：

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

如果因网络不佳下载中断，可以通过以下命令在中断后继续下载：

git lfs pull

2 安装 ChatTTS 依赖包

下载ChatTTS官网GitHub源码：

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

安装Python依赖包:

cd ChatTTS
pip install -r requirements.txt

torch需要使用2.2.2

3 ChatTTS 中文文本转音频文件

ChatTTS 官网的样例代码 API 可能会跑不起来。

ChatTTS-01.py

import ChatTTS
import torch
import torchaudio

# 下载的ChatTTS模型文件目录，请按照实际情况替换
MODEL_PATH = '/path/to/ChatTTS-Model'

# 初始化并加载模型，特别注意加载模型参数，官网样例代码已经过时
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',
)

# 需要转化为音频的文本内容
text = '中文文本'

# 文本转为音频
wavs = chat.infer(text, use_decoder=True)

# 保存音频文件到本地文件（采样率为24000Hz）
torchaudio.save("./output/output-01.wav", torch.from_numpy(wavs[0]), 24000)

运行Python代码：python ChatTTS-01.py

也可以在文本转换成语音之后，直接播放语音内容：

from IPython.display import Audio

# 播放生成的音频（autoplay=True 代表自动播放）
Audio(wavs[0], rate=24000, autoplay=True)

4 搭建 WebUI 界面

4.1 安装 Python 依赖包

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

运行 Python 程序，即可看到可视化界面，可以随意输入文本来生成音频文件了。

ChatTTS-WebUI.py

import random

import ChatTTS
import gradio as gr
import numpy as np
import torch
from ChatTTS.infer.api import refine_text, infer_code

print('启动ChatTTS WebUI......')

# WebUI设置
WEB_HOST = '127.0.0.1'
WEB_PORT = 8089

MODEL_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微调文本：{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 = 24000
    text_data = text[0] if isinstance(text, list) else text

    return [(sample_rate, audio_data), text_data]


def main():
    with gr.Blocks() as demo:
        default_text = "文字"
        text_input = gr.Textbox(label="输入文本", lines=4, placeholder="Please Input Text...", value=default_text)

        with gr.Row():
            refine_text_checkbox = gr.Checkbox(label="文本微调开关", value=True)
            temperature_slider = gr.Slider(minimum=0.00001, maximum=1.0, step=0.00001, value=0.8, label="语音温度参数")
            top_p_slider = gr.Slider(minimum=0.1, maximum=0.9, step=0.05, value=0.7, label="语音top_P采样参数")
            top_k_slider = gr.Slider(minimum=1, maximum=20, step=1, value=20, label="语音top_K采样参数")

        with gr.Row():
            audio_seed_input = gr.Number(value=42, label="语音随机数")
            generate_audio_seed = gr.Button("\U0001F3B2")
            text_seed_input = gr.Number(value=42, label="文本随机数")
            generate_text_seed = gr.Button("\U0001F3B2")

        generate_button = gr.Button("文本生成语音")

        text_output = gr.Textbox(label="微调文本", interactive=False)
        audio_output = gr.Audio(label="语音")

        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, ])

    # 启动WebUI
    demo.launch(server_name='127.0.0.1', server_port=8089, share=False, show_api=False, )


if __name__ == '__main__':
    main()