模型介绍
中文整理模型说明
- 模型来自 modelscope 网站 OpenBMB/VoxCPM-0.5B 。
- 参考文本和音频示例来自 https://huggingface.co/openbmb/VoxCPM-0.5B 。
- 经过实测,该模型生成速度和质量相当可观,可以应用于语音合成和实时聊天等场景。
合规风险
使用模型前请阅读官方 README zh.md 说明,以下两点需要特别注意:
- **禁止** 将 voxcpm-0.5b 用于任何 **非法或不道德** 的目的。
- 推荐在任何用此模型生成并 **公开分享** 的内容描述中明确标注为 **AI 生成**。
下载模型
SDK 下载
Git 下载
本地应用
默认为 gpu 推理,推荐安装 cuda 。
环境
推荐使用虚拟环境进行包管理,基础环境要求: python >= 3.11。
程序
这里使用 gradio 构建一个简单的本地界面示例。
这里使用 fastapi 构建一个接口推理示例,响应音频文件在 data/audio file 字段中,采用 base64 编码,采样率 16000hz 。
python
import asyncio
import base64
import logging
import os
import uuid
from concurrent.futures import ThreadPoolExecutor
from contextlib import asynccontextmanager
from pathlib import Path
from typing import Annotated
import coloredlogs
import numpy as np
import soundfile as sf
import uvicorn
from fastapi import FastAPI, File, Form, Header, UploadFile
from voxcpm import VoxCPM
获取日志对象
logger = logging.getLogger( name )
logger.setLevel(logging.DEBUG)
创建日志文件
log file path = "app.log"
file formatter = logging.Formatter(
fmt="%(levelname)s %(asctime)s - [t-%(threadName)s] (%(filename)s:%(lineno)d) %(funcName)s : %(message)s",
datefmt="%Y-%m-%d %H:%M:%S"
)
file handler = logging.FileHandler(log file path, encoding="utf-8")
file handler.setFormatter(file formatter)
logger.addHandler(file handler)
创建彩色日志输出
coloredlogs.install(
level="DEBUG",
datefmt="%Y-%m-%d %H:%M:%S",
milliseconds=True,
fmt="%(levelname)s %(asctime)s - [t-%(threadName)s] (%(filename)s:%(lineno)d) %(funcName)s : %(message)s",
field styles={
"levelname": {"color": "green"},
"asctime": {"color": "magenta"},
"name": {"color": "blue", "bold": True},
"thread": {"color": "yellow"},
"funcName": {"color": "green"},
},
)
class TTSModel:
"""
TTS 模型:VoxCPM
"""
def init (self, model path: Path, num threads: int = 4):
self. model path: Path = model path
self. num threads: int = num threads
self. model: VoxCPM None = None
self. inference executor: ThreadPoolExecutor = ThreadPoolExecutor(max workers=num threads,
thread name prefix="tts")
def load model(self):
"""加载模型(单线程执行)"""
logger.info(f"Initializing TTS model: VoxCPM from {self. model path}")
if self. model is not None:
return
if not self. model path.exists():
raise ValueError(f"Model path does not exist: {self. model path}")
try:
self. model = VoxCPM.from pretrained(str(self. model path))
logger.info(f"Successfully loaded VoxCPM model from {self. model path}")
except Exception as e:
raise RuntimeError(f"Failed to load VoxCPM model: {e}")
def generate speech(
self,
prompt speech: Path None,
prompt speech enhancement: bool,
prompt text: str,
cfg value: float,
inference timestep: int,
target text: str,
text normalization: bool,
):
"""同步执行语音合成"""
if self. model is None:
raise RuntimeError("TTS model has not been loaded or has been released")
logger.debug(
f"Generating speech: target text='{target text[:50]}...', "
f"cfg={cfg value}, steps={inference timestep}, normalize={text normalization}"
)
try:
# 调用模型生成
wav = self. model.generate(
text=target text,
prompt wav path=str(prompt speech) if prompt speech else None,
prompt text=prompt text if prompt speech else None,
cfg value=cfg value,
inference timesteps=inference timestep,
normalize=text normalization,
denoise=prompt speech enhancement,
retry badcase=True,
retry badcase max times=3,
)
# 确保输出是 numpy 数组,采样率通常为16000Hz
if not isinstance(wav, np.ndarray):
raise ValueError("Generated audio is not a numpy array")
# 保存为临时 WAV 文件
output file name = f"{uuid.uuid4()}.wav"
output file path = tmp audio path.joinpath(output file name)
sf.write(output file path, wav, samplerate=16000, subtype="PCM 16")
logger.info(f"Generated audio saved to: {output file path}")
# 读取临时 WAV 文件,转换为 base64 字符串
output file base64 = base64.b64encode(output file path.read bytes()).decode("utf-8")
# 返回音频信息
audio info = sf.info(output file path)
return {
"code": 200,
"msg": "success",
"data": {
"id": str(uuid.uuid4()),
"audio file": output file base64,
"sample rate": audio info.samplerate,
"duration": audio info.duration,
"channels": audio info.channels,
"format": "wav",
"metadata": {
"target text": target text,
"prompt text": prompt text,
"cfg value": cfg value,
"inference timesteps": inference timestep,
"text normalization": text normalization,
"prompt speech enhancement": prompt speech enhancement,
},
},
}
except Exception as e:
raise RuntimeError(f"Failed to generate speech: {e}")
async def generate speech async(
self,
prompt speech: Path None,
prompt speech enhancement: bool,
prompt text: str,
cfg value: float,
inference timestep: int,
target text: str,
text normalization: bool,
):
"""异步执行语音合成"""
loop = asyncio.get event loop()
return await loop.run in executor(
self. inference executor,
self.generate speech,
prompt speech,
prompt speech enhancement,
prompt text,
cfg value,
inference timestep,
target text,
text normalization,
)
def release(self):
"""释放模型资源"""
if self. model:
logger.info("Releasing VoxCPM model...")
self. model = None
self. inference executor.shutdown(wait=True)
全局配置
API KEY = os.getenv("API KEY", f"Bearer {uuid.uuid4()}")
logger.info(f"API KEY: {API KEY}")
tmp audio path = Path(os.getenv("TMP AUDIO PATH", "./tmp"))
tts model: VoxCPM None = None
def init app():
"""初始化应用"""
# 模型路径
model path = Path(os.getenv("MODEL PATH", "xiaowangge/voxcpm-0.5b"))
if not model path.exists():
raise ValueError(f"MODEL PATH({model path}) does not exist")
# 初始化临时音频存放目录
if not tmp audio path.exists():
tmp audio path.mkdir(parents=True, exist ok=True)
logger.info(f"Created tmp audio path: {tmp audio path}")
# 推理线程数
max executors = int(os.getenv("MAX EXECUTORS", "4"))
if max executors
优先采用原作者提供的中文模型卡片;仅有英文说明时由闲社AI翻译整理。重要参数、许可证和商用范围请以原文为准。
你将获得什么
完整仓库方式包含该固定版本中的权重、配置、分词器及说明文件;指定版本方式只下载所选文件或整组分片。多模态模型的投影文件、配置等依赖,请按作者说明配齐。
完整仓库可能包含多个权重格式,因此需要的磁盘空间可能大于单一格式的模型大小。下载文件不等于完成模型部署。
使用前须知
运行模型需要的设备、工具与依赖,以原作者说明为准。
是否允许商用、修改和分发,请查看模型许可证。
闲社提供资料与镜像下载方法,不代表已运行模型或完成安全审计。下载的代码应先检查,再执行。
资料或下载有问题?
登录后可提交反馈。
