本地搭建ChatTTS WebUi

声明

声明：本教程基于modelscope.cn的演示站进行本地搭建，环境为Windows

作者GitHub地址：https://github.com/2noise/ChatTTS

Webui体验地址：https://modelscope.cn/studios/AI-ModelScope/ChatTTS-demo/summary

第一步 克隆代码

先在终端输入以下内容，克隆modelscope的文件到本地

git clone https://www.modelscope.cn/studios/AI-ModelScope/ChatTTS-demo.git

克隆好之后进入文件目录

到了目录之后直接执行安装txt中的内容太慢了，换成国内源很快就能下好

pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

第二步 安装库

下好之后也不能立马执行，需要在手动安装一些库

依此执行以下命令

pip install modelscope -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install gradio -i https://pypi.tuna.tsinghua.edu.cn/simple

上面安装没有问题，执行以下代码

python app.py

启动等待程序下载完成（白条可能会卡住不动，因为他是一次显示很多的，看一下网络宽带有占用就行了，不要暂停程序!）

错误解决

我就遇到一个错误，说什么Windows不支持，然后我根据错误修改了一些代码我就能运行了，由于我第一次搭建是拿物理电脑搭建的，解决运行之后想着写一篇文章，然后用虚拟机继续搭建，遇到的错误又不一样了，烦死了，于是我直接把我物理机的core.py中的内容复制到虚拟机中的core.py就能运行了，直接复制以下代码覆盖克隆下来的core.py

core.py在ChatTTS目录下面

import os
import logging
from omegaconf import OmegaConf

import platform
import torch
from vocos import Vocos
from .model.dvae import DVAE
from .model.gpt import GPT_warpper
from .utils.gpu_utils import select_device
from .utils.infer_utils import count_invalid_characters, detect_language
from .utils.io_utils import get_latest_modified_file
from .infer.api import refine_text, infer_code

from huggingface_hub import snapshot_download

logging.basicConfig(level = logging.INFO)

class Chat:
    def __init__(self, ):
        self.pretrain_models = {}
        self.normalizer = {}
        self.logger = logging.getLogger(__name__)
        
    def check_model(self, level = logging.INFO, use_decoder = False):
        not_finish = False
        check_list = ['vocos', 'gpt', 'tokenizer']
        
        if use_decoder:
            check_list.append('decoder')
        else:
            check_list.append('dvae')
            
        for module in check_list:
            if module not in self.pretrain_models:
                self.logger.log(logging.WARNING, f'{module} not initialized.')
                not_finish = True
                
        if not not_finish:
            self.logger.log(level, f'All initialized.')
            
        return not not_finish
        
    def load_models(self, source='huggingface', force_redownload=False, local_path='<LOCAL_PATH>', **kwargs):
        if source == 'huggingface':
            hf_home = os.getenv('HF_HOME', os.path.expanduser("~/.cache/huggingface"))
            try:
                download_path = get_latest_modified_file(os.path.join(hf_home, 'hub/models--2Noise--ChatTTS/snapshots'))
            except:
                download_path = None
            if download_path is None or force_redownload: 
                self.logger.log(logging.INFO, f'Download from HF: https://huggingface.co/2Noise/ChatTTS')
                download_path = snapshot_download(repo_id="2Noise/ChatTTS", allow_patterns=["*.pt", "*.yaml"])
            else:
                self.logger.log(logging.INFO, f'Load from cache: {download_path}')
        elif source == 'local':
            self.logger.log(logging.INFO, f'Load from local: {local_path}')
            download_path = local_path

        self._load(**{k: os.path.join(download_path, v) for k, v in OmegaConf.load(os.path.join(download_path, 'config', 'path.yaml')).items()}, **kwargs)
        
    def _load(
        self, 
        vocos_config_path: str = None, 
        vocos_ckpt_path: str = None,
        dvae_config_path: str = None,
        dvae_ckpt_path: str = None,
        gpt_config_path: str = None,
        gpt_ckpt_path: str = None,
        decoder_config_path: str = None,
        decoder_ckpt_path: str = None,
        tokenizer_path: str = None,
        device: str = None,
        compile: bool = True,
    ):
        if not device:
            device = select_device(4096)
            self.logger.log(logging.INFO, f'use {device}')
            
        if vocos_config_path:
            vocos = Vocos.from_hparams(vocos_config_path).to(device).eval()
            assert vocos_ckpt_path, 'vocos_ckpt_path should not be None'
            vocos.load_state_dict(torch.load(vocos_ckpt_path))
            self.pretrain_models['vocos'] = vocos
            self.logger.log(logging.INFO, 'vocos loaded.')
        
        if dvae_config_path:
            cfg = OmegaConf.load(dvae_config_path)
            dvae = DVAE(**cfg).to(device).eval()
            assert dvae_ckpt_path, 'dvae_ckpt_path should not be None'
            dvae.load_state_dict(torch.load(dvae_ckpt_path, map_location='cpu'))
            self.pretrain_models['dvae'] = dvae
            self.logger.log(logging.INFO, 'dvae loaded.')
            
        if gpt_config_path:
            cfg = OmegaConf.load(gpt_config_path)
            gpt = GPT_warpper(**cfg).to(device).eval()
            assert gpt_ckpt_path, 'gpt_ckpt_path should not be None'
            gpt.load_state_dict(torch.load(gpt_ckpt_path, map_location='cpu'))
            if platform.system() != 'Windows':
                gpt.gpt.forward = torch.compile(gpt.gpt.forward, backend='inductor', dynamic=True)
            self.pretrain_models['gpt'] = gpt
            spk_stat_path = os.path.join(os.path.dirname(gpt_ckpt_path), 'spk_stat.pt')
            assert os.path.exists(spk_stat_path), f'Missing spk_stat.pt: {spk_stat_path}'
            self.pretrain_models['spk_stat'] = torch.load(spk_stat_path).to(device)
            self.logger.log(logging.INFO, 'gpt loaded.')
            
        if decoder_config_path:
            cfg = OmegaConf.load(decoder_config_path)
            decoder = DVAE(**cfg).to(device).eval()
            assert decoder_ckpt_path, 'decoder_ckpt_path should not be None'
            decoder.load_state_dict(torch.load(decoder_ckpt_path, map_location='cpu'))
            self.pretrain_models['decoder'] = decoder
            self.logger.log(logging.INFO, 'decoder loaded.')
        
        if tokenizer_path:
            tokenizer = torch.load(tokenizer_path, map_location='cpu')
            tokenizer.padding_side = 'left'
            self.pretrain_models['tokenizer'] = tokenizer
            self.logger.log(logging.INFO, 'tokenizer loaded.')
            
        self.check_model()
    
    def infer(
        self, 
        text, 
        skip_refine_text=False, 
        refine_text_only=False, 
        params_refine_text={}, 
        params_infer_code={'prompt':'[speed_5]'}, 
        use_decoder=True,
        do_text_normalization=False,
        lang=None,
    ):
        
        assert self.check_model(use_decoder=use_decoder)
        
        if not isinstance(text, list): 
            text = [text]
        
        if do_text_normalization:
            for i, t in enumerate(text):
                _lang = detect_language(t) if lang is None else lang
                self.init_normalizer(_lang)
                text[i] = self.normalizer[_lang].normalize(t, verbose=False, punct_post_process=True)
            
        for i in text:
            invalid_characters = count_invalid_characters(i)
            if len(invalid_characters):
                self.logger.log(logging.WARNING, f'Invalid characters found! : {invalid_characters}')
                
        if not skip_refine_text:
            text_tokens = refine_text(self.pretrain_models, text, **params_refine_text)['ids']
            text_tokens = [i[i < self.pretrain_models['tokenizer'].convert_tokens_to_ids('[break_0]')] for i in text_tokens]
            text = self.pretrain_models['tokenizer'].batch_decode(text_tokens)
            if refine_text_only:
                return text
            
        text = [params_infer_code.get('prompt', '') + i for i in text]
        params_infer_code.pop('prompt', '')
        result = infer_code(self.pretrain_models, text, **params_infer_code, return_hidden=use_decoder)
        
        if use_decoder:
            mel_spec = [self.pretrain_models['decoder'](i[None].permute(0,2,1)) for i in result['hiddens']]
        else:
            mel_spec = [self.pretrain_models['dvae'](i[None].permute(0,2,1)) for i in result['ids']]
            
        wav = [self.pretrain_models['vocos'].decode(i).cpu().numpy() for i in mel_spec]
        
        return wav
    
    def sample_random_speaker(self, ):
        
        dim = self.pretrain_models['gpt'].gpt.layers[0].mlp.gate_proj.in_features
        std, mean = self.pretrain_models['spk_stat'].chunk(2)
        return torch.randn(dim, device=std.device) * std + mean
    
    def init_normalizer(self, lang):
        
        if lang not in self.normalizer:
            from nemo_text_processing.text_normalization.normalize import Normalizer
            self.normalizer[lang] = Normalizer(input_case='cased', lang=lang)

运行结果

执行app.py之后会得到一个地址，访问即可

访问就可以生成了