开源语音合成库 coqui TTS 使用记录

1 介绍

功能：可以克隆声音；可以转换声音。支持多语言。

GitHub https://github.com/coqui-ai/TTS

在线试玩（效果不如本地demo） https://huggingface.co/spaces/coqui/xtts

2 本地搭建demo

搭建环境

conda create -n coqui python=3.10

conda activate coqui

pip install TTS （可以自动安装需要的依赖包，也可以根据requirements.txt逐个安装依赖包）

运行时其他缺的包直接pip即可（貌似就只有一个）

下载源代码和模型

GitHub https://github.com/coqui-ai/TTS 版本dbf1a08

模型地址 https://huggingface.co/coqui/XTTS-v2/tree/main

测试脚本

import torch
from TTS.api import TTS

## 查看模型列表
# for name in TTS().list_models().list_models():
#     print(name)

## Init TTS 初始化，传入模型和配置文件路径
device = "cuda" if torch.cuda.is_available() else "cpu"  # Get device
tts = TTS(model_path="/home/ze/coqui/mypath/models/model.pth", 
          config_path="/home/ze/coqui/mypath/models/config.json", 
          progress_bar=True).to(device)

## Text to speech to a file
# ## 英文
# tts.tts_to_file(text="A short story is a piece of prose fiction. It can typically be read in a single sitting and focuses on a self-contained incident or series of linked incidents, with the intent of evoking a single effect or mood.", 
#                 speaker_wav="mypath/audio/samples_en_sample.wav", 
#                 language="en", 
#                 file_path="output.wav")
# ## 中文
# tts.tts_to_file(text="龙能大能小，能升能隐；大则兴云吐雾，小则隐介藏形；升则飞腾于宇宙之间，隐则潜伏于波涛之内。方今春深，龙乘时变化，犹人得志而纵横四海。", 
#                 speaker_wav="mypath/audio/samples_zh-cn-sample.wav", 
#                 language="zh-cn", 
#                 file_path="output.wav")
## 指定中文音色，输出英文
tts.tts_to_file(text="A short story is a piece of prose fiction. It can typically be read in a single sitting and focuses on a self-contained incident or series of linked incidents, with the intent of evoking a single effect or mood.", 
                speaker_wav="mypath/audio/dragon.wav", 
                language="en", 
                file_path="output.wav")

遇到问题

报错 NotADirectoryError: [Errno 20] Not a directory: '/home/ze/coqui/mypath/models/model.pth/model.pth’

原因：代码接口存在bug，在/home/ze/coqui/TTS-dev/TTS/utils/synthesizer.py line192加载模型时没有按照接口定义。

解决：将home/ze/coqui/TTS-dev/TTS/utils/synthesizer.py line192调用语句self.tts_model.load_checkpoint()中参数tts_checkpoint改为模型所在路径，比如"/home/ze/coqui/mypath/models”