whisper-api
介绍
使用openai的开源项目winsper语音识别开源模型封装成openai chatgpt兼容接口
软件架构
使用uvicorn、fastapi、openai-whisper等开源库实现高性能接口
更多介绍 [https://blog.csdn.net/weixin_40986713/article/details/138712293](https://blog.csdn.net/weixin_40986713/article/details/138712293)
使用说明
- 下载代码
- 安装 ffmpeg https://ffmpeg.org/download.html
- 安装依赖 项目根目录下执行命令
pip install -r requirements.txt
- 运行代码 项目根目录下执行命令
python main.py
这里的 http://0.0.0.0:3003
就是连接地址。
启动类代码
import atexit
import json
import os
import tempfile
import time
import uvicorn
from fastapi import FastAPI, UploadFile, File, Security, HTTPException
from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
from whisper_script import WhisperHandler
app = FastAPI()
security = HTTPBearer()
env_bearer_token = 'sk-tarzan'
model_size = os.getenv("MODEL_SIZE", "base")
language = os.getenv("LANGUAGE", "Chinese")
def cleanup_temp_file(path):
if os.path.exists(path):
os.remove(path)
with open('options.json', 'r') as options:
# 使用json.load()函数读取并解析文件内容
load_options = json.load(options)
# 语音识别
@app.post("/v1/audio/transcriptions")
async def transcribe(file: UploadFile = File(...), credentials: HTTPAuthorizationCredentials = Security(security)):
if env_bearer_token is not None and credentials.credentials != env_bearer_token:
raise HTTPException(status_code=401, detail="Invalid token")
file_bytes = await file.read()
return {"text": audio_to_text(file_bytes, 'transcribe')}
# 语音翻译
@app.post("/v1/audio/translations")
async def translate(file: UploadFile = File(...), credentials: HTTPAuthorizationCredentials = Security(security)):
if env_bearer_token is not None and credentials.credentials != env_bearer_token:
raise HTTPException(status_code=401, detail="Invalid token")
file_bytes = await file.read()
return {"text": audio_to_text(file_bytes, 'translate')}
def audio_to_text(file_bytes, task):
start_time = time.time()
max_file_size = 500 * 1024 * 1024
if len(file_bytes) > max_file_size:
raise HTTPException(status_code=400, detail="File is too large")
temp_path = None
try:
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_audio:
temp_audio.write(file_bytes)
temp_path = temp_audio.name
model_size = load_options.get("model_size")
language = load_options.get("language")
prompts = {
"verbose": load_options.get("verbose"),
"temperature": load_options.get("temperature"),
"compression_ratio_threshold": load_options.get("compression_ratio_threshold"),
"logprob_threshold": load_options.get("logprob_threshold"),
"no_speech_threshold": load_options.get("no_speech_threshold"),
"condition_on_previous_text": load_options.get("condition_on_previous_text"),
"initial_prompt": load_options.get("initial_prompt"),
"word_timestamps": load_options.get("word_timestamps"),
"prepend_punctuations": load_options.get("prepend_punctuations"),
"append_punctuations": load_options.get("append_punctuations")
}
print('temp_path', temp_path)
handler = WhisperHandler(temp_path, model_size=model_size, language=language, task=task, prompt=prompts)
result = handler.transcribe()
except Exception as e:
raise HTTPException(status_code=500, detail=str(e))
finally:
atexit.register(cleanup_temp_file, temp_path)
end_time = time.time()
print(f"audio to text took {end_time - start_time:.2f} seconds")
return result['text']
if __name__ == "__main__":
token = os.getenv("ACCESS_TOKEN")
if token is not None:
env_bearer_token = token
try:
uvicorn.run("main:app", reload=True, host="0.0.0.0", port=3003)
except Exception as e:
print(f"API启动失败!\n报错:\n{e}")
docker
- docker打包命令
docker build -t whisper .
2.docker命令启动
gpu显卡模式
docker run -itd --name whisper-api -p 3003:3003 --gpus all --restart=always whisper
- 默认 ACCESS_TOKEN=sk-tarzan
cpu模式
docker run -itd --name whisper-api -p 3003:3003 --restart=always whisper
- 默认 ACCESS_TOKEN=sk-tarzan
鉴权模式
docker run -itd --name whisper-api -p 3003:3003-e ACCESS_TOKEN=yourtoken --gpus all --restart=always whisper
docker run -itd --name whisper-api -p 3003:3003-e ACCESS_TOKEN=yourtoken --restart=always whisper
- yourtoken 修改你设置的鉴权token,接口调用header 里传
Authorization:Bearer sk-tarzan
docker日志查看
docker logs -f [容器id或容器名称]
配置文件
options.json
{
"model_size": "base",
"language": "Chinese"
}
- 可结合one-api,接入FastGPT等rag开源项目使用,使用教程如下:
《Fastgpt接入Whisper本地模型实现语音输入》