回到首页

使用whisper

whisper是openAI开发贡献的开源语音识别库，git clone https://github.com/openai/whisper,pip install .，安装完whisper库。之后即可玩耍

whisper xxx.wav，使用small（461MB）模型识别多语种的语音。--language可指定目标语言，--task translate可规定额外执行翻译为英语的任务。指定--model为base格式和small格式，2GB的显存够用，small格式比base格式准确度更高。

对音频的采样率没有特别的要求，至于声道数有待验证

whisper xxx.mp4 --language Chinese --output_format srt

类型	参数数量	英语识别模型	多语言模型	要求的显存	相对速度
tiny	39M	tiny.en	tiny	~1GB	~32x
base	74M	base.en	base	~1GB	~16x
small	244M	small.en	small	~2GB	~6x
medium	769M	medium.en	medium	~5GB	~2x
large	1.55G	N/A	large	~10GB	1x

whisper在各语种上的识别词错率

使用whisper做一个本地语音识别程序

import wave
import numpy as np
from pyaudio import PyAudio,paInt16
import time
import whisper
import os
import threading

recordingSaveDir='C:/Users/tellw/Music/records'

class recorder:
	NUM_SAMPLES=2000
	SAMPLING_RATE=8000
	LEVEL=500
	COUNT_NUM=20
	SAVE_LENGTH=8

	Voice_String=[]

	pa=PyAudio()
	stream=pa.open(format=paInt16,channels=1,rate=SAMPLING_RATE,input=True,frames_per_buffer=NUM_SAMPLES)
	save_buffer=[]
	recording=False

	def savewav(self,filename):
		wf=wave.open(os.path.join(recordingSaveDir,filename),'wb')
		wf.setnchannels(1)
		wf.setsampwidth(2)
		wf.setframerate(self.SAMPLING_RATE)
		wf.writeframes(np.array(self.save_buffer).tobytes())
		wf.close()

	def start(self):
		threading._start_new_thread(self._record,())

	def _record(self):
		self.save_buffer=[]
		save_count=20
		self.recording=True
		while self.recording:
			string_audio_data=self.stream.read(self.NUM_SAMPLES)
			audio_data=np.fromstring(string_audio_data,dtype=np.short)
			large_sample_count=np.sum(audio_data>self.LEVEL)
			if large_sample_count>self.COUNT_NUM:
				save_count=self.SAVE_LENGTH
			else:
				save_count-=1
			if save_count>0:
				self.save_buffer.append(string_audio_data)
				# print(self.save_buffer)

	def end(self):
		self.recording=False
		if len(self.save_buffer)>0:
			filename=f'{int(time.time()*1000)}.wav'
			self.savewav(filename)
			return filename
		else:
			print('record nothing')
			return None

if __name__=='__main__':
	r=recorder()
	model=whisper.load_model('base')
	while True:
		command=input('i or o or q:')
		if command=='q':
			break
		elif not r.recording:
			print('recording...')
			r.start()
		elif r.recording:
			filename=r.end()
			if filename is not None:
				result=model.transcribe(os.path.join(recordingSaveDir,filename))
				print(result['text'])

参考自通过pyaudio使用麦克风录音和用python编写录音机——GUI控制录音开始与结束（tkinter）

本文创建于2022.10.25/20.29，修改于2022.12.7/20.38