CosyVoice-300M 轻量级语音合成模型：2025 最新音色克隆教程，3 秒快速适配智能客服

conda create -n cosyvoice python=3.8
conda activate cosyvoice

pip install torch torchaudio librosa

wget https://modelscope.cn/models/damo/speech_cosyvoice-300m/snapshots/v1.0.0/cosyvoice-300m.pth

import torch
from cosyvoice import CosyVoice

model = CosyVoice('cosyvoice-300m.pth')

import torchaudio

prompt_speech, _ = torchaudio.load('reference.wav', )

text = "欢迎致电XX客服，请问有什么可以帮您？"
output = model.inference_zero_shot(text, prompt_speech)
torchaudio.save('output.wav', output, )

pip install fastapi uvicorn

from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI()

class TTSRequest(BaseModel):
    text: str
    prompt: str

@app.post("/tts")
async def tts(request: TTSRequest):
    output = model.inference_zero_shot(request.text, request.prompt)
    return {"audio": output.tolist()}

uvicorn server:app --host 0.0.0.0 --port 

import requests

url = "http://localhost:8000/tts"
data = {"text": "请问我的订单什么时候发货？", "prompt": "reference.wav"}
response = requests.post(url, json=data)