如何通过Azure SpeechService Python SDK获取音频转录的置信度

问题根因 #

你遇到的NBest信息“被屏蔽”属于误解：连续识别模式本身不会隐藏详细输出，你直接打印事件对象只会输出简略文本，完整的详细结果存放在结果对象的json字段中。

修正步骤 #

• 首先将OutputFormat替换为明确的枚举值，避免硬编码数值出错
• 在recognized事件回调中解析evt.result.json，即可获取包含NBest列表、整句置信度、单词级置信度的完整数据
• 替换硬编码的time.sleep逻辑，用循环等待识别完成的标志，避免长音频被截断

修正后完整代码 #

import azure.cognitiveservices.speech as speechsdk
import json
import time

SUBSCRIPTION_KEY = "你的服务密钥"
REGION = "你的服务区域"

def speech_recognize_continuous_from_file():
    """基于音频文件的连续语音识别，获取单词级置信度"""
    speech_config = speechsdk.SpeechConfig(subscription=SUBSCRIPTION_KEY, region=REGION)
    speech_config.speech_recognition_language="pt-BR"
    audio_config = speechsdk.audio.AudioConfig(filename="file.wav")
    speech_config.enable_dictation()
    # 用明确的枚举值设置详细输出，替代硬编码的1
    speech_config.output_format = speechsdk.OutputFormat.Detailed
    speech_recognizer = speechsdk.SpeechRecognizer(speech_config=speech_config, audio_config=audio_config)
    done = False

    def stop_cb(evt):
        print('CLOSING on {}'.format(evt))
        nonlocal done
        done = True

    # 处理最终识别结果，提取置信度
    def handle_recognized(evt):
        if evt.result.reason == speechsdk.ResultReason.RecognizedSpeech:
            # 解析完整的详细结果json
            detailed_result = json.loads(evt.result.json)
            # 取NBest列表，默认第一条是置信度最高的结果
            best_result = detailed_result['NBest'][0]
            # 整句置信度
            sentence_confidence = best_result['Confidence']
            print(f"整句文本：{best_result['Display']}，整句置信度：{sentence_confidence}")
            # 遍历获取每个单词的置信度
            print("单词级置信度：")
            for word in best_result['Words']:
                print(f"单词：{word['Word']}，置信度：{word['Confidence']}，开始偏移：{word['Offset']}，时长：{word['Duration']}")
        elif evt.result.reason == speechsdk.ResultReason.NoMatch:
            print(f"未识别到语音：{evt.result.no_match_details}")

    # 绑定事件回调
    speech_recognizer.recognizing.connect(lambda evt: print('RECOGNIZING: {}'.format(evt.result.text)))
    speech_recognizer.recognized.connect(handle_recognized)
    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(f'CANCELED 原因：{evt.reason}，错误详情：{evt.error_details}'))
    
    speech_recognizer.session_stopped.connect(stop_cb)
    speech_recognizer.canceled.connect(stop_cb)

    # 启动连续识别
    speech_recognizer.start_continuous_recognition()
    # 替换硬编码sleep，循环等待识别完成
    while not done:
        time.sleep(0.5)
    speech_recognizer.stop_continuous_recognition()

speech_recognize_continuous_from_file()

关键说明 #

• 只有recognized事件返回的是最终确认结果，才有准确的置信度数据，recognizing返回的中间临时结果不会包含完整的NBest和单词级置信度
• NBest列表默认返回最多5个候选识别结果，可根据需求遍历所有候选做进一步校验
• 单词级结果中同时返回了每个单词的时间偏移和时长，可用于对齐原始音频做质量校验

内容的提问来源于stack exchange，提问作者vribeir0