【Day4】语音识别（音频转文字）

语音识别的三个解决方案：
原本用途：本来是要求从视频中识别语音，然后把文字内容提取出来，结果看了很多项目，中文的注释，识别的却是英文，感到授课的门槛有点低，我能看懂别人开源的代码，距离自己开发还是有距离的。后来探索了很多，比如字幕生成，把字幕不生成到视频下方而是一段一段增加到txt文本里，我也认为这是最好的办法，而且能顺便给每个字、每个句子一个时间戳。后来意识到一天的时间实在是很难完成，于是从网上找了最普通的，也是不难理解的解决方案及相关代码：还是从视频转音频，再从音频转文字。

一共三种方案：① speech_recognition加上r.recognize_sphinx(audio,language=“zh-CN”))② 百度API（科大讯飞类似）从百度手里薅羊毛真的是困难 ③ 利用TIMIT项目的解决方法去实现

方案一：

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Time    : 2022.1.5# @File    : duizhaozu1.pyimport os
import speech_recognition as srdef file_to_wav(file_path, wav_path, sampling_rate):if os.path.exists(wav_path):  # 如果文件存在# 删除文件，可使用以下两种方法。os.remove(wav_path)# 终端命令command = "D:/download/ffmpeg-master-latest-win64-lgpl/bin/ffmpeg.exe -i {} -ac 1 -ar {} {}".format(file_path, sampling_rate, wav_path)os.system(command)if __name__ == '__main__':file_path = r'C:\Users\PineappleMan\Desktop\ok\DFS.mp4'wav_path = r'C:\Users\PineappleMan\Desktop\ok\DFS.wav'sampling_rate = 16000file_to_wav(file_path, wav_path, sampling_rate)r=sr.Recognizer()with sr.AudioFile(wav_path) as source:audio =r.record(source)print("文本内容:",r.recognize_sphinx(audio,language="zh-CN"))

其中，文件为视频，如果直接是音频（wav文件）就直接把格式转换的代码删掉，代码借鉴了https://download.csdn.net/download/weixin_38693753/13709062

pip install -i https://pypi.tuna.tsinghua.edu.cn/simple speech_recognition

运行的时候遇到了错误就是缺少pocketsphinx：https://download.csdn.net/download/yuxuwen1234/12195200，实际上去下载一个whl就行了，我的python是3.7所以下载了链接中的，很快就能运行好了，网上很多解决方案，因为报错的是swig的问题，但是“头痛医头，脚痛医脚”，就应该机智的去下载whl去解决这个问题。

方案二：

import base64
import json
import os
import time
import uuidimport requests
import urllib.response
from inc import db_config
from inc import rtysdbclass BaiduRest:def __init__(self, cu_id, api_key, api_secert):self.token_url = "https://openapi.baidu.com/oauth/2.0/token?grant_type=client_credentials&client_id=%s&client_secret=%s"self.getvoice_url = "http://tsn.baidu.com/text2audio?tex=%s&lan=zh&cuid=%s&ctp=1&tok=%s"self.upvoice_url = 'http://vop.baidu.com/server_api'self.cu_id = cu_idself.get_token(api_key, api_secert)returndef get_token(self, api_key, api_secert):token_url = self.token_url % (api_key, api_secert)r_str = urllib.response.urlopen(token_url).read()token_data = json.loads(r_str)self.token_str = token_data['access_token']return True# 语音合成
# def text2audio(self, text, filename):
#     get_url = self.getvoice_url % (urllib.response.quote(text), self.cu_id, self.token_str)
#
#
# voice_data = urllib.response.urlopen(get_url).read()
# voice_fp = open(filename, 'wb+')
# voice_fp.write(voice_data)
# voice_fp.close()
# return True##语音识别
def audio2text(self, filename):data = {}data['format'] = 'wav'data['rate'] = 8000data['channel'] = 1data['cuid'] = self.cu_iddata['token'] = self.token_strwav_fp = open(filename, 'rb')voice_data = wav_fp.read()data['len'] = len(voice_data)# data['speech'] = base64.b64encode(voice_data).decode('utf-8')data['speech'] = base64.b64encode(voice_data).replace('\n', '')# post_data = json.dumps(data)result = requests.post(self.upvoice_url, json=data, headers={'Content-Type': 'application/json'})data_result = result.json()if (data_result['err_msg'] == 'success.'):return data_result['result'][0]else:return Falsedef test_voice(voice_file):api_key = "vossGHIgEETS6IMRxBDeahv8"api_secert = "3c1fe6a6312f41fa21fa2c394dad5510"bdr = BaiduRest("0-57-7B-9F-1F-A1", api_key, api_secert)# 生成# start = time.time()# bdr.text2audio("你好啊", "out.wav")# using = time.time() - start# print using# 识别# start = time.time()result = bdr.audio2text(voice_file)# result = bdr.audio2text("weather.pcm")# using = time.time() - startreturn resultdef get_master_audio(check_status='cut_status'):if check_status == 'cut_status':sql = "SELECT id,url, time_long,sharps FROM ocenter_recognition WHERE status=0"elif check_status == 'finished_status':sql = "SELECT id,url, time_long,sharps FROM ocenter_recognition WHERE finished_status=0"else:return Falsedata = rtysdb.select_data(sql, 'more')if data:return dataelse:return Falsedef go_recognize(master_id):section_path = "C:/Users/PineappleMan/Desktop/ok/audio1.wav"sql = "SELECT id,rid,url,status FROM ocenter_section WHERE rid=%d AND status=0 order by id asc limit 10" % (master_id)# print sqlrecord = rtysdb.select_data(sql, 'more')# print recordif not record:return Falsefor rec in record:# print section_path+'/'+rec[1]voice_file = section_path + '/' + rec[2]if not os.patcvoice_file:continueresult = test_voice(voice_file)print(result)exit(0)if result:# rtysdb.update_by_pk('ocenter_section',rec[0],{'content':result,'status':1})sql = "update ocenter_section set content='%s', status='%d' where id=%d" % (result, 1, rec[0])  # print sqlrtysdb.do_exec_sql(sql)parent_content = rtysdb.select_data("SELECT id,content FROM ocenter_recognition WHERE id=%d" % (rec[1]))# print parent_contentif parent_content:new_content = parent_content[1] + resultupdate_content_sql = "update ocenter_recognition set content='%s' where id=%d" % (new_content, rec[1])rtysdb.do_exec_sql(update_content_sql)else:rtysdb.do_exec_sql("update ocenter_section set status='%d' where id=%d" % (result, 1, rec[0]))time.sleep(5)else:rtysdb.do_exec_sql("UPDATE ocenter_recognition SET finished_status=1 WHERE id=%d" % (master_id))# 对百度语音识别不了的音频文件进行转换
def ffmpeg_convert():section_path = "C:/Users/PineappleMan/Desktop/ok/audio1.wav"# print section_pathused_audio = get_master_audio('cut_status')# print used_audioif used_audio:for audio in used_audio:audio_path = section_path + '/' + audio[1]new_audio = uuid.uuid1()command_line = "ffmpeg -i " + audio_path + " -ar 8000 -ac 1 -f wav " + section_path + "/Uploads/Convert/convert_" + str(new_audio) + ".wav";# print command_lineos.popen(command_line)if os.path.exists(section_path + "/Uploads/Convert/convert_" + str(new_audio) + ".wav"):convert_name = "Uploads/Convert/convert_" + str(new_audio) + ".wav"ffmpeg_cut(convert_name, audio[3], audio[0])sql = "UPDATE ocenter_recognition SET status=1,convert_name='%s' where id=%d" % (convert_name, audio[0])rtysdb.do_exec_sql(sql)# 将大音频文件切成碎片def ffmpeg_cut(convert_name, sharps, master_id):section_path = "C:/Users/PineappleMan/Desktop/ok/audio1.wav"if sharps > 0:for i in range(0, sharps):timeArray = time.localtime(i * 30)h = time.strftime("%H", timeArray)h = int(h) - 8h = "0" + str(h)ms = time.strftime("%M:%S", timeArray)start_time = h + ':' + str(ms)cut_name = section_path + '/' + convert_namedb_store_name = "Uploads/Section/" + str(uuid.uuid1()) + '-' + str(i + 1) + ".wav"section_name = section_path + "/" + db_store_namecommand_line = "ffmpeg.exe -i " + cut_name + " -vn -acodec copy -ss " + start_time + " -t 00:00:30 " + section_name# print command_lineos.popen(command_line)data = {}data['rid'] = master_iddata['url'] = db_store_namedata['create_time'] = int(time.time())data['status'] = 0rtysdb.insert_one('ocenter_section', data)if __name__ == "__main__":ffmpeg_convert()audio = get_master_audio('finished_status')if audio:for ad in audio:go_recognize(ad[0])

该项目参考了https://download.csdn.net/download/weixin_38531210/12867107，但是做了很多改动！无论是官网还是这些项目，调用的包都是比较陈旧的，甚至python3已经用别的名称进行取代了。所以改动的工作量是相当大的。

方案三：

TIMIT是比较经典的英文的语音识别，找到相关代码并不难。这里就不多讲了。过两天我把每行代码做好注释再发出来。

总结

其实一天实现一个功能，尤其是对我这种考研考了很久，忘了很多东西的人来说。但是这种有目的的功能实现，还是很锻炼自己的。再接再厉吧。