python基于SpeechRecognition-PocketSphinx实现简单语音识别

系统环境：win 10

编译环境：Thonny 3.3.13

编程语言：python 3.7.9

导师希望我帮忙编写一个能够实现离线语音识别并记录的程序，在查阅CSDN资料后，本只会python的弱鸡感觉貌似只有PocketSphinx比较靠谱

本文基于pyhton基于PocketSphinx实现简单语音识别_疯人忠的博客-CSDN博客_pocketsphinx基础上写成，在按照博客教学步骤进行语音识别后，发现正确率极低

广泛查询之后，发现可能是其对.lm转lm.bin直接重命名的方法有问题

在使用https://github.com/Uberi/speech_recognition/blob/master/reference/pocketsphinx.rst

中提到的SphinxBase进行转换之后，识别率达到了95% （针对14个词语的小词库

sphinx_lm_convert -i mylm.lm -o mylm.lm.bin

使用pyAudio获取语音

import pyaudio
import wave
import speech_recognition as speech
def record_and_recog(wave_out_path):
    CHUNK = 1024
    FORMAT = pyaudio.paInt16
    CHANNELS = 1 
    RATE = 16000 #采样率一定要16000！
    TIME = 3 #以秒为单位
    
    p = pyaudio.PyAudio()
    stream = p.open(format=FORMAT,
                    channels=CHANNELS,
                    rate=RATE,
                    input=True,
                    frames_per_buffer=CHUNK)

    wf = wave.open(wave_out_path, 'wb')
    wf.setnchannels(CHANNELS)
    wf.setsampwidth(p.get_sample_size(FORMAT))
    wf.setframerate(RATE)
    for _ in range(0, int(RATE / CHUNK * TIME)):
        data = stream.read(CHUNK)
        wf.writeframes(data)
    stream.stop_stream(); stream.close(); p.terminate(); wf.close()

额外使用了

    with speech.Microphone() as source:
        recognizer.adjust_for_ambient_noise(source)

进行简单的降噪处理

之后对录音进行识别

audio = speech.AudioFile(wave_out_path)
    with audio as source:
        audio = recognizer.record(source)
        # 使用Sphinx识别语音
    try:
        resultstr = recognizer.recognize_sphinx(audio, language='zh-CN')
        if " " not in resultstr and resultstr != "":
            resultstr = "Sphinx识别结果为:" + resultstr
        elif  resultstr = "":
            resultstr = "错误识别？空结果"
        else:
            resultstr = "错误识别？多字:" + resultstr
    except speech.RequestError as e:
        resultstr = "警告！{0}".format(e)

成功！