Azure Speech-to-text Python：start_continuous_recognition无法在流结束时停止

我用Python调用Azure Speech-to-text服务，输入data是仅几秒的音频字节串，预期云服务在流结束时停止处理并返回识别文本，但实际recognized事件触发延迟了约5分钟。

代码实现 #

speech_config = speechsdk.SpeechConfig(subscription=API_KEY,
                                       region="westeurope",
                                       speech_recognition_language='de-DE')
stream = PushAudioInputStream(stream_format=
                              AudioStreamFormat(samples_per_second=sample_rate, bits_per_sample=SAMPLE_WIDTH * 8,
                                               compressed_stream_format=speechsdk.AudioStreamContainerFormat.FLAC))
audio_input = speechsdk.AudioConfig(stream=stream)
stream.write(data)
speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config, audio_config=audio_input)
speech_recognizer.start_continuous_recognition()
done = False

def stop_recognition(evt):
    logger.debug("Stopped MS Azure recognition: %s", evt)
    nonlocal done
    done = True

def recognized(evt):
    logger.info("Recognized MS Azure transcript: %s", evt)
    nonlocal text
    text += " " + evt.result.text

speech_recognizer.recognizing.connect(lambda evt: print('RECOGNIZING: {}'.format(evt)))
speech_recognizer.recognized.connect(lambda evt: print('RECOGNIZED: {}'.format(evt)))
speech_recognizer.session_started.connect(lambda evt: print('SESSION STARTED: {}'.format(evt)))
speech_recognizer.session_stopped.connect(lambda evt: print('SESSION STOPPED {}'.format(evt)))
speech_recognizer.canceled.connect(lambda evt: print('CANCELED {}'.format(evt)))

speech_recognizer.recognized.connect(recognized)
speech_recognizer.session_stopped.connect(stop_recognition)
speech_recognizer.canceled.connect(stop_recognition)
while not done:
    time.sleep(.5)
speech_recognizer.stop_continuous_recognition()

延迟日志 #

2022-11-13 23:58:19,504 - speech_processing.speech_recognition.speech_recognition - DEBUG - Sending 192000 bytes (6 sec) for recognition
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=50e5c478cdc34e0a8ced3867be493bc3, text="telefon", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=d1448833ac8f40ef9c1ebc4cae488bcd, text="telefonspeicher", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=cdf9f074c13b4a2c94960ec147db765c, text="telefon speichere", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=548133156bb44dc8ae08fd0848fa8ec5, text="telefon speichere als", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=c03970619f1e42278b2a2ef19ee4f1fe, text="telefon speichere als bärbel", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=ff5f6a18d1e4409cab2661582cb8a693, text="telefon speichere als bärbel 0", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=a3cb8f82c62b4235abc2fea2696342f8, text="telefon speichere als bärbel 03", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=cafc805031654aa4865a4fe1b742d1cd, text="telefon speichere als bärbel 038", reason=ResultReason.RecognizingSpeech))
RECOGNIZING: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=69782c9485244e3191846b924adb3807, text="telefon speichere als bärbel 0385", reason=ResultReason.RecognizingSpeech))
RECOGNIZED: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=9d92890d52d84b7f926a6977d6324ca1, text="Telefon speichere als Bärbel 0385.", reason=ResultReason.RecognizedSpeech))
2022-11-14 00:03:26,487 - speech_processing.speech_recognition.speech_recognition - INFO - Recognized MS Azure transcript: SpeechRecognitionEventArgs(session_id=2e4c92f4fed6498f8f5260199bdcc5d7, result=SpeechRecognitionResult(result_id=9d92890d52d84b7f926a6977d6324ca1, text="Telefon speichere als Bärbel 0385.", reason=ResultReason.RecognizedSpeech))

问题原因&解决方案 #

核心原因 #

你使用连续识别模式但未明确告知Azure服务音频流已结束。Azure Speech服务在连续识别模式下会持续等待更多音频输入，直到收到流结束信号或触发超时（默认约5分钟），这就是延迟的根源。同时你写入音频数据后未关闭流，服务无法判断音频是否全部发送完成。

修复步骤 #

1. 关闭音频流：在stream.write(data)后调用stream.close()，明确告知服务音频已全部发送。
2. 切换识别模式（可选）：如果处理的都是短音频片段，使用单次识别模式（recognize_once()）更高效，服务会在处理完音频后立即返回结果，无需等待超时。

修复后的连续识别模式代码

speech_config = speechsdk.SpeechConfig(subscription=API_KEY,
                                       region="westeurope",
                                       speech_recognition_language='de-DE')
stream = PushAudioInputStream(stream_format=
                              AudioStreamFormat(samples_per_second=sample_rate, bits_per_sample=SAMPLE_WIDTH * 8,
                                               compressed_stream_format=speechsdk.AudioStreamContainerFormat.FLAC))
audio_input = speechsdk.AudioConfig(stream=stream)
stream.write(data)
# 关键：关闭流，告知服务音频传输完成
stream.close()
speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config, audio_config=audio_input)
speech_recognizer.start_continuous_recognition()
done = False

def stop_recognition(evt):
    logger.debug("Stopped MS Azure recognition: %s", evt)
    nonlocal done
    done = True

def recognized(evt):
    logger.info("Recognized MS Azure transcript: %s", evt)
    nonlocal text
    text += " " + evt.result.text

speech_recognizer.recognizing.connect(lambda evt: print('RECOGNIZING: {}'.format(evt)))
speech_recognizer.recognized.connect(lambda evt: print('RECOGNIZED: {}'.format(evt)))
speech_recognizer.session_started.connect(lambda evt: print('SESSION STARTED: {}'.format(evt)))
speech_recognizer.session_stopped.connect(lambda evt: print('SESSION STOPPED {}'.format(evt)))
speech_recognizer.canceled.connect(lambda evt: print('CANCELED {}'.format(evt)))

speech_recognizer.recognized.connect(recognized)
speech_recognizer.session_stopped.connect(stop_recognition)
speech_recognizer.canceled.connect(stop_recognition)
while not done:
    time.sleep(.5)
speech_recognizer.stop_continuous_recognition()

单次识别模式替代方案（适合短音频）

speech_config = speechsdk.SpeechConfig(subscription=API_KEY,
                                       region="westeurope",
                                       speech_recognition_language='de-DE')
stream = PushAudioInputStream(stream_format=
                              AudioStreamFormat(samples_per_second=sample_rate, bits_per_sample=SAMPLE_WIDTH * 8,
                                               compressed_stream_format=speechsdk.AudioStreamContainerFormat.FLAC))
audio_input = speechsdk.AudioConfig(stream=stream)
stream.write(data)
stream.close()

speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config, audio_config=audio_input)
# 单次识别直接获取结果
result = speech_recognizer.recognize_once()

if result.reason == speechsdk.ResultReason.RecognizedSpeech:
    print(f"识别结果: {result.text}")
    logger.info(f"Recognized MS Azure transcript: {result.text}")
elif result.reason == speechsdk.ResultReason.NoMatch:
    print(f"未识别到语音: {result.no_match_details}")
elif result.reason == speechsdk.ResultReason.Canceled:
    cancellation_details = result.cancellation_details
    print(f"识别被取消: {cancellation_details.reason}")
    if cancellation_details.reason == speechsdk.CancellationReason.Error:
        print(f"错误详情: {cancellation_details.error_details}")

验证效果 #

修复后，服务会在收到流关闭信号后立即处理完剩余音频，触发recognized事件并结束会话，不会再出现5分钟的延迟。

内容的提问来源于stack exchange，提问作者k_o_