git 官网:https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/README_cn.md
强烈建议用户在 Linux 环境下,3.7 以上版本的 python 上安装 PaddleSpeech。 PaddleSpeech 依赖于 paddlepaddle 1. ************* 安装 paddlepaddle ************** ###### gpu 版本 ##### 需要安装 显卡cudn https://www.iotword.com/4567.html python -m pip install paddlepaddle-gpu -i https://pypi.tuna.tsinghua.edu.cn/simple ######### cpu 版本 ##### pip install paddlepaddle -i https://mirror.baidu.com/pypi/simple 2. ************** 安装paddlespeech (pip 方式)************** pip install paddlespeech pip uninstall munpy pip install munpy==1.23.5 3. 下载nltk数据包,放入到nltk可以检索的路径下 cd ~ wget -P data https://paddlespeech.bj.bcebos.com/Parakeet/tools/nltk_data.tar.gz tar zxvf data/nltk_data.tar.gz 快速体验 https://aistudio.baidu.com/aistudio/projectdetail/4353348?sUid=2470186&shared=1&ts=1660878142250
我的使用:
#!/usr/bin/env python3 # -*- coding: utf-8 -*- # @mail : [email protected] # @Time : 14/6/2023 6:04 pm # @Author : Sea # @File : apitest.py # @history: # **************************** import os import random import time import json import base64 from fastapi.responses import FileResponse from fastapi import FastAPI, UploadFile, File, Query, Form, Request from fastapi.middleware.cors import CORSMiddleware import uvicorn app = FastAPI(title="Sea test API") app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"], ) from paddlespeech.cli.asr.infer import ASRExecutor from paddlespeech.cli.tts.infer import TTSExecutor asr = ASRExecutor() # 初始化成全局变量,防止多次初始化导致显存不够 from https://github.com/PaddlePaddle/PaddleSpeech/issues/2881和https://github.com/PaddlePaddle/PaddleSpeech/issues/2908 tts = TTSExecutor() # 公共函数,所有接口都能用 def random_string(length=32): # 生成32位随机字符串,为了生成随机文件名 string='abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789' return ''.join(random.choice(string) for i in range(length)) # 公共函数,所有接口都能用 def base64_to_audio(audio_base64, folder_name=None): # 服务器上用folder_name参数,用于在audio_file_path中拼接路径,如f'/home/www/card/{folder_name}/', # 不同的folder_name对应不同的识别任务(如身份证识别、营业执照识别),本地测试不用 audio_base64 = audio_base64.split(',')[-1] audio = base64.b64decode(audio_base64) audio_file_name = random_string() + '_' + (str(time.time()).split('.')[0]) # 不带扩展名,因为不知道收到的音频文件的原始扩展名,手机录的不一定是什么格式 audio_file_path = f'/home/python/speech/{folder_name}/' + audio_file_name with open(audio_file_path, 'wb') as f: f.write(audio) return audio_file_path # 将收到的音频文件转为16k 16 bit 1 channel的wav文件,16k表示16000Hz的采样率,16bit不知道是什么 # 若给paddlespeech传的文件不对,会提示The sample rate of the input file is not 16000.The program will resample the wav file to 16000. # If the result does not meet your expectations,Please input the 16k 16 bit 1 channel wav file.所以要提前转换。 def resample_rate(audio_path_input): audio_path_output = audio_path_input + '_output' + '.wav' # 传入的audio_path_input不带扩展名,所以后面直接拼接字符串 command = f'ffmpeg -y -i {audio_path_input} -ac 1 -ar 16000 -b:a 16k {audio_path_output}' # 这个命令输出的wav文件,格式上和PaddleSpeech在README中给的示例zh.wav(https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav,内容是'我认为跑步最重要的就是给我带来了身体健康')一样。from https://blog.csdn.net/Ezerbel/article/details/124393431 command_result = os.system(command) # from https://blog.csdn.net/liulanba/article/details/115466783 assert command_result == 0 if os.path.exists(audio_path_output): return audio_path_output elif not os.path.exists(audio_path_output): # ffmpeg输出的文件不存在,可能是ffmpeg命令没执行完,等1秒(因在虚拟机测试转一个8.46M的MP3需0.48秒),1秒后若还没有输出文件,说明报错了 time.sleep(1) if os.path.exists(audio_path_output): return audio_path_output else: return None # 语音转文字 # 只接受POST方法访问 # @app.post("/speechtotext", tags=["Sea"]) def speech_to_text(request: Request): audio_file_base64 = request.json().get('audio_file_base64') # 要转为文字的语音文件的base64编码,开头含不含'data:audio/wav;base64,'都行 audio_file_path = base64_to_audio(audio_file_base64, folder_name='speech_to_text/audio_file') # 存放收到的原始音频文件 audio_path_output = resample_rate(audio_path_input=audio_file_path) if audio_path_output: # asr = ASRExecutor() result = asr(audio_file=audio_path_output) # 会在当前代码所在文件夹中产生exp/log文件夹,里面是paddlespeech的日志文件,每一次调用都会生成一个日志文件。记录这点时的版本号是paddlepaddle==2.3.2,paddlespeech==1.2.0。 from https://github.com/PaddlePaddle/PaddleSpeech/issues/1211 os.remove(audio_file_path) # 识别成功时删除收到的原始音频文件和转换后的音频文件 os.remove(audio_path_output) # try: # shutil.rmtree('') # 删除文件夹,若文件夹不存在会报错。若需删除日志文件夹,用这个。from https://blog.csdn.net/a1579990149wqh/article/details/124953746 # except Exception as e: # pass return json.dumps({'code':200, 'msg':'识别成功', 'data':result}, ensure_ascii=False) else: return json.dumps({'code':400, 'msg':'识别失败'}, ensure_ascii=False) # 文字转语音 @app.get("/texttospeech1",tags=["Sea"]) def text_to_speech(text_str: str): print("start text_to_speech ") start = time.time() # tts = TTSExecutor() audio_file_name = random_string() + '_' + (str(time.time()).split('.')[0]) + '.wav' audio_file_path = '/home/sea/Desktop/mp3mp4/' + audio_file_name # 输出24k采样率wav格式音频。同speech_to_text()中一样,会在当前代码所在文件夹中产生exp/log文件夹,里面是paddlespeech的日志文件,每一次调用都会生成一个日志文件。 tts(text=text_str, output=audio_file_path) print("end text_to_speech and cost : " + str(time.time() - start) + " s") if os.path.exists(audio_file_path): return FileResponse(audio_file_path) elif not os.path.exists(audio_file_path): return json.dumps({'code':400, 'msg':'识别失败'}, ensure_ascii=False) # @app.post("/files") async def file_to_text( file: UploadFile = File(...)): file = file.file result = asr(audio_file=file) return {"result": result} if __name__=='__main__': uvicorn.run(app, host="0.0.0.0", port=5000)
标签:https,PaddleSpeech,tts,base64,file,output,path,audio,安装 From: https://www.cnblogs.com/lshan/p/17494384.html