Azure PullAudioInputStream无法与Twilio语音正确配合使用

如何解决Azure PullAudioInputStream无法与Twilio语音正确配合使用

我将Twilio Media流与Azure认知服务（语音到文本）集成在一起。我继承了speechsdk.audio.PullAudioInputStreamCallback类，将音频块发送到服务器。

import azure.cognitiveservices.speech as speechsdk
import queue

class SocketReaderCallback(speechsdk.audio.PullAudioInputStreamCallback):

    def __init__(self):
        super().__init__()
        self._q = queue.Queue()

    def read(self,buffer: memoryview) -> int:
        chunk = self._q.get()
        buffer[:len(chunk)] = chunk
        return len(chunk)

    def has_bytes(self):
        return True if self._q.qsize() > 0 else False

    def queueup(self,chunk):
        self._q.put(chunk)

    def close(self):
        print("AZ.Callback.Closed")

下面是转录器类的代码。这里的add_request方法将音频块添加到上述回调类的Queue中。回调类从队列中选择大块并上传到Azure服务器以进行转录。

import azure.cognitiveservices.speech as speechsdk
import queue
from rule_engine.medium.azure_transcribe.azure_calback import SocketReaderCallback

class AzureTranscribe:

    def __init__(self,speech_config,on_response,user_id):
        self._on_response = on_response
        self.callback = SocketReaderCallback()
        wave_format = speechsdk.audio.AudioStreamFormat(samples_per_second=8000,bits_per_sample=8,channels=1)
        self._stream = speechsdk.audio.PullAudioInputStream(self.callback,wave_format)
        audio_config = speechsdk.audio.AudioConfig(stream=self._stream)
        self._speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config,language="en-IN",audio_config=audio_config)
        self._ended = False
        self.user_id = user_id
        self.initialize_once()
        self.state = None

    def initialize_once(self):
        # Connect callbacks to the events fired by the speech recognizer
        self._speech_recognizer.recognizing.connect(lambda evt: print('AZ.RECOGNIZING: {}'.format(evt)))
        self._speech_recognizer.recognized.connect(lambda evt: print('AZ.RECOGNIZED: {}'.format(evt)))
        self._speech_recognizer.session_started.connect(lambda evt: print('AZ.SESSION STARTED: {}'.format(evt)))
        self._speech_recognizer.session_stopped.connect(lambda evt: print('AZ.SESSION STOPPED {}'.format(evt)))
        self._speech_recognizer.canceled.connect(lambda evt: print('AZ.CANCELED {}'.format(evt)))
        self._speech_recognizer.start_continuous_recognition()

    def add_request(self,buffer):
        # buffer,self.state =  audioop.ratecv(bytes(buffer),2,8000,16000,self.state)
        self.callback.queueup(bytes(buffer))
    
    def terminate(self):    
        self._ended = True
        self._speech_recognizer.stop_continuous_recognition()

• 如果我从音频文件上传音频块，则转录为 准确。
• 如果我从twilio呼叫上传音频块，则转录为 非常糟糕。

Twilio的采样率为8 kHz，而Azure的预期采样率为 是16 kHz。但是，Azure既可以使用采样率，又可以提供差的采样率 两者的质量转录。

解决方法

请使用语音服务SDK压缩音频输入流API通过PullStream或PushStream将压缩音频流传输到语音服务。

我们建议使用一种方法将音频转换为支持的格式。 •您可以使用FFMpeg进行音频格式转换。音频文件的正确格式为16kHz，16Bit和Mono。正确的目标格式的命令行为： ffmpeg.exe -i inputfile.wav -sample_fmt s16 -ac 1 -ar 16000 outputfile.wav

•文档引用了SoX，请参见https://docs.microsoft.com/en-us/azure/cognitive-services/speech-service/how-to-custom-speech-test-data#audio-data-for-testing

对于压缩的音频输入流，请遵循doc。