Logo

菜单

基于Funasr+modelscope(Paraformer)分角色语音识别

吴书松
吴书松
发布于 2025-09-25 / 16 阅读
0
0

基于Funasr+modelscope(Paraformer)分角色语音识别

#大模型#语音

1、启动docker

docker run -p 10095:10095 -p 5000:5000 -it --privileged=true -v $PWD/funasr-runtime-resources/models:/workspace/models -v $PWD/funasr-runtime-resources/app:/workspace/app --name funasr-voice registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-cpu-0.4.6

含义

1、-p 10095:10095：funasr内部模型的ws端口

2、-p 5000:5000：内部自定义应用端口

3、-it：启动成功之后，进入容器内部

4、-v $PWD/funasr-runtime-resources/models:/workspace/models：funasr模型映射路径

5、-v $PWD/funasr-runtime-resources/app:/workspace/app：app自定义应用映射路径

6、--name funasr-voice：容器名称

7、registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-cpu-0.4.6：容器镜像，已上传云盘

1.1 安装环境

进入容器内部之后，安装环境

1、安装python的flask。启动服务

pip install flask

2、部署语音大模型

这里使用：modelscope 的 iic/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn

但是由于这个模型不支持onnx，所以直接通过Funasr的run_server.sh是没办法启动的，下面采用python方式启动

2.1、下载模型

进入容器内部之后，进入路径：/workspace/app

执行下面下载命令，模型已经上传云盘

modelscope download --model iic/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn --local_dir ./models/iic/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn
modelscope download --model iic/speech_fsmn_vad_zh-cn-16k-common-pytorch --local_dir ./models/iic/speech_fsmn_vad_zh-cn-16k-common-pytorch
modelscope download --model iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch --local_dir ./models/iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch
modelscope download --model iic/speech_campplus_sv_zh-cn_16k-common --local_dir ./models/iic/speech_campplus_sv_zh-cn_16k-common

2.2、创建python服务

2.2.1、在线运行

每次启动python都会在线下载模型，启动之后，使用不会再次下载

在线运行，就不需要执行上面模型下载步骤

touch voice_server.py
from flask import Flask, request, jsonify
from funasr import AutoModel
import os

app = Flask(__name__)

# 加载模型（服务启动时加载一次，后续请求复用）
model = AutoModel(
    model="iic/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn",
    vad_model="fsmn-vad",
    punc_model="ct-punc-c",
    spk_model="cam++",  
    model_revision="v2.0.4"
)

@app.route('/asr', methods=['POST'])
def recognize():
    # 接收上传的音频文件
    audio_file = request.files['audio']
    save_path = f"/tmp/{audio_file.filename}"
    audio_file.save(save_path)

    # 进行语音识别
    result = model.generate(input=save_path)
    
    # 清理临时文件
    os.remove(save_path)

    # 返回识别结果
    return jsonify(result)

if __name__ == '__main__':
    app.run(host='0.0.0.0', port=5000)

启动即可

python voice_server.py

2.2.2、离线运行

执行上面模型下载，然后上传到指定目录

touch voice_server.py
from flask import Flask, request, jsonify
from funasr import AutoModel
import os

app = Flask(__name__)

MODEL_DIR = "./models" 

model = AutoModel(
    model=os.path.join(MODEL_DIR, "iic/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn"),
    vad_model=os.path.join(MODEL_DIR, "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch"),
    punc_model=os.path.join(MODEL_DIR, "iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"),
    spk_model=os.path.join(MODEL_DIR, "iic/speech_campplus_sv_zh-cn_16k-common"),  
    model_revision="v2.0.4",
    disable_update=True 
)

@app.route('/asr', methods=['POST'])
def recognize():
    if 'audio' not in request.files:
        return jsonify({"error": "No audio file provided"}), 400
    
    audio_file = request.files['audio']
    if audio_file.filename == '':
        return jsonify({"error": "No selected file"}), 400

    save_path = f"/tmp/{audio_file.filename}"
    audio_file.save(save_path)

    try:
        result = model.generate(input=save_path)
        
        os.remove(save_path)

        return jsonify(result)
    except Exception as e:
        if os.path.exists(save_path):
            os.remove(save_path)
        return jsonify({"error": str(e)}), 500

@app.route('/health', methods=['GET'])
def health_check():
    return jsonify({"status": "healthy", "model_loaded": True})

if __name__ == '__main__':
    if not os.path.exists(MODEL_DIR):
        os.makedirs(MODEL_DIR)
        print(f"warm: model {MODEL_DIR} not find，Created. Please ensure that the model has been downloaded to this directory.")
    
    app.run(host='0.0.0.0', port=5000, debug=False)

注意AutoModel创建的时候，4个模型对应的位置即可，启动

2.2.3 基于gunicorn运行

2.2.3.1、环境安装

pip install gunicorn

2、安装依赖

创建requirements.txt

内容：

flask>=2.0.0
funasr>=0.8.0
modelscope>=1.9.0
gunicorn>=20.0.0

执行命令，安装

pip install -r requirements.txt

2.2.3.2、创建gunicorn_config.py

import os
import multiprocessing

bind = "0.0.0.0:5000"

# work thread
workers = multiprocessing.cpu_count() * 2 + 1

# work model
worker_class = "sync"

# Restart after each worker process has processed the number of requests
max_requests = 1000
max_requests_jitter = 100

timeout = 120

# Daemon mode
daemon = False

# log info
accesslog = "./logs/access.log"
errorlog = "./logs/error.log"
loglevel = "info"

# pid info
pidfile = "./tmp/gunicorn.pid"

# Preload the application (Important: Ensure that the model is loaded only once)
preload_app = True

# xecuted before the work process starts
def pre_fork(server, worker):
    # Make sure the log directory exists
    os.makedirs("./logs", exist_ok=True)
    os.makedirs("./tmp", exist_ok=True)

def when_ready(server):
    server.log.info("During the server startup, it may take a relatively long time for the model to load...")

# Executed when the work process exits
def worker_exit(server, worker):
    # You can add the cleaning code here
    pass

2.2.3.3、创建start_server.sh

#!/bin/bash

export PYTHONPATH=$PYTHONPATH:$(pwd)
export MODEL_DIR="./models"

# create dir
mkdir -p ./logs
mkdir -p ./tmp

# check dir
if [ ! -d "$MODEL_DIR" ]; then
    echo "warm: model dir $MODEL_DIR not find"
    echo "Please ensure that the model file has been downloaded"
fi

# start Gunicorn server
echo "Launch the FunASR voice recognition service..."
echo "The server will run on: http://0.0.0.0:5000"
echo "log file: ./logs/access.log and ./logs/error.log"

gunicorn -c gunicorn_config.py app:app
chmod +x start_server.sh

2.2.3.4、创建app.py

from flask import Flask, request, jsonify
from funasr import AutoModel
import os
import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = Flask(__name__)
MODEL_DIR = "./models" 

os.environ['MODELSCOPE_CACHE'] = './models'
os.environ['MODELSCOPE_HUB_CACHE'] = './models'

model = None

def load_model():
    global model
    try:
        logger.info("Start loading the speech recognition model...")
        model = AutoModel(
            model=os.path.join(MODEL_DIR, "iic/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn"),
            vad_model=os.path.join(MODEL_DIR, "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch"),
            punc_model=os.path.join(MODEL_DIR, "iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"),
            spk_model=os.path.join(MODEL_DIR, "iic/speech_campplus_sv_zh-cn_16k-common"),  
            model_revision="v2.0.4",
            disable_update=True 
        )
        logger.info("The speech recognition model has been successfully loaded")
    except Exception as e:
        logger.error(f"Model loading failed: {e}")
        raise

load_model()

@app.route('/asr', methods=['POST'])
def recognize():
    if model is None:
        return jsonify({"error": "The model is not loaded."}), 500
    
    if 'audio' not in request.files:
        return jsonify({"error": "No audio file was provided"}), 400
    
    audio_file = request.files['audio']
    if audio_file.filename == '':
        return jsonify({"error": "Unselected file"}), 400

    save_path = f"/tmp/{audio_file.filename}"
    audio_file.save(save_path)

    try:
        logger.info(f"Start processing the audio file: {audio_file.filename}")
        result = model.generate(input=save_path)
        logger.info(f"Audio processing completed: {audio_file.filename}")
        
        os.remove(save_path)
        
        return jsonify({
            "success": True,
            "result": result
        })
    except Exception as e:
        logger.error(f"An error occurred while processing the audio: {e}")
        if os.path.exists(save_path):
            os.remove(save_path)
        return jsonify({"error": str(e)}), 500

@app.route('/health', methods=['GET'])
def health_check():
    status = "healthy" if model is not None else "unhealthy"
    return jsonify({
        "status": status,
        "service": "funasr-asr",
        "model_loaded": model is not None
    })

@app.route('/', methods=['GET'])
def index():
    return jsonify({
        "service": "FunASR voice recognition service",
        "version": "1.0",
        "endpoints": {
            "asr": "/asr (POST)",
            "health": "/health (GET)"
        }
    })

if __name__ == '__main__':
    load_model()
    app.run(host='0.0.0.0', port=5000, debug=False)

2.2.3.5、启动

1、直接通过gunicorn_config.py 启动

gunicorn -c gunicorn_config.py app:app

2、通过脚本启动

./start_server.sh

2.2.3.6 创建系统服务文件（可选）

创建 funasr.service（用于systemd）：

[Unit]
Description=FunASR Speech Recognition Service
After=network.target

[Service]
Type=simple
User=www-data
Group=www-data
WorkingDirectory=/path/to/your/app
Environment=PYTHONPATH=/path/to/your/app
Environment=MODEL_DIR=/path/to/your/app/models
ExecStart=/usr/local/bin/gunicorn -c gunicorn_config.py app:app
ExecReload=/bin/kill -s HUP $MAINPID
Restart=always
RestartSec=5

[Install]
WantedBy=multi-user.target

2.3 测试

上面两种方式，访问路径都是：http://192.168.1.111:5000/asr

3、其他命令

1、复制容器内部文件到宿主机

docker cp <容器名>:<容器内路径> <宿主机路径>

2、复制宿主机文件到容器内部

docker cp <宿主机路径> <容器名>:<容器内路径>

Funasr 内部python 使...
Idea主题

评论