Ubuntu下部署CosyVoice语音克隆项目

1、下载项目（科学上网环境下）

git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git

cd CosyVoice

git submodule update --init --recursive

2、创建并激活虚拟环境

uv venv --python 3.10 && source .venv/bin/activate

3、安装依赖库（无需科学上网环境）

UV_HTTP_TIMEOUT=600 uv pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --extra-index-url https://pypi.nvidia.com --index-strategy unsafe-best-match

安装魔搭库

uv pip install modelscope

卸载deepspeed

uv pip uninstall deepspeed

更新ruamel.yaml

uv pip install "ruamel.yaml==0.17.21" --force-reinstall

4、安装必要软件

sudo apt-get install sox libsox-dev ffmpeg -y

5、编写模型下载脚本，并执行下载（从魔搭平台下载，无需翻墙）

# download_cosyvoice_models.py

import os

# 👇 关键：提前禁用 CUDA

os.environ["CUDA_VISIBLE_DEVICES"] = ""

os.environ["USE_CUDA"] = "0"

os.environ["TRANSFORMERS_OFFLINE"] = "0"  # 允许联网下载

from modelscope import snapshot_download

models_to_download = [

    ('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', 'pretrained_models/Fun-CosyVoice3-0.5B'),

    ('iic/CosyVoice2-0.5B', 'pretrained_models/CosyVoice2-0.5B'),

    ('iic/CosyVoice-300M', 'pretrained_models/CosyVoice-300M'),

    ('iic/CosyVoice-300M-SFT', 'pretrained_models/CosyVoice-300M-SFT'),

    ('iic/CosyVoice-300M-Instruct', 'pretrained_models/CosyVoice-300M-Instruct'),

    ('iic/CosyVoice-ttsfrd', 'pretrained_models/CosyVoice-ttsfrd'),

]

if __name__ == "__main__":

    for model_id, local_dir in models_to_download:

        print(f"📥 Downloading {model_id} → {local_dir}")

        snapshot_download(model_id, local_dir=local_dir)

        print(f"✅ Done: {model_id}\n")

    print("🎉 All models downloaded!")

6、启动CosyVoice克隆服务

python webui.py

7、浏览器访问

http://127.0.0.1:8000/