如何在JavaScript中调整Azure语音转文本的沉默超时时间

问题描述 #

使用Azure Speech SDK的recognizeOnceAsync实现语音转文本，当前代码如下：

var SpeechSDK, recognizer, synthesizer;
var speechConfig = SpeechSDK.SpeechConfig.fromSubscription('SUB_KEY', 'SUB_REGION');
var audioConfig  = SpeechSDK.AudioConfig.fromDefaultMicrophoneInput();
recognizer = new SpeechSDK.SpeechRecognizer(speechConfig, audioConfig);
new Promise(function(resolve) {
    recognizer.onend = resolve;
    recognizer.recognizeOnceAsync(
        function (result) {
            recognizer.close();
            recognizer = undefined;
            resolve(result.text);
        },
        function (err) {
            alert(err);
            recognizer.close();
            recognizer = undefined;
        }
    );
}).then(r => {
    console.log(`Azure STT interpreted: ${r}`);
});

HTML中引入SDK的方式：

<script src="https://aka.ms/csspeech/jsbrowserpackageraw"></script>

希望延长recognizeOnceAsync返回结果前允许的沉默时长（避免说话停顿换气时被判定为识别结束），尝试过以下设置但无效：

const SILENCE_UNTIL_TIMEOUT_MS = 5000;
speechConfig.SpeechServiceConnection_EndSilenceTimeoutMs = SILENCE_UNTIL_TIMEOUT_MS;
audioConfig.setProperty("Speech_SegmentationSilenceTimeoutMs", SILENCE_UNTIL_TIMEOUT_MS);

需要解决在使用fromDefaultMicrophoneInput时的有效设置方法。

解决方法 #

正确的属性设置方式 #

问题出在直接给speechConfig赋值属性的方式错误，必须使用SDK提供的setProperty()方法，结合官方定义的PropertyId常量来配置，而非直接访问属性名。针对recognizeOnceAsync场景，只需配置服务端的结束沉默超时即可：

1. 使用SpeechSDK.PropertyId.SpeechServiceConnection_EndSilenceTimeoutMs作为属性标识，通过speechConfig.setProperty()设置超时值，注意属性值需要传入字符串类型：

const SILENCE_UNTIL_TIMEOUT_MS = 5000; // 设置为5秒
speechConfig.setProperty(
  SpeechSDK.PropertyId.SpeechServiceConnection_EndSilenceTimeoutMs,
  SILENCE_UNTIL_TIMEOUT_MS.toString()
);

2. 无需在audioConfig上设置Speech_SegmentationSilenceTimeoutMs，该属性用于本地音频分段处理，不适用于recognizeOnceAsync的服务端结束判定逻辑。

完整修改后的代码 #

var SpeechSDK, recognizer, synthesizer;
var speechConfig = SpeechSDK.SpeechConfig.fromSubscription('SUB_KEY', 'SUB_REGION');

// 配置延长沉默超时
const SILENCE_UNTIL_TIMEOUT_MS = 5000;
speechConfig.setProperty(
  SpeechSDK.PropertyId.SpeechServiceConnection_EndSilenceTimeoutMs,
  SILENCE_UNTIL_TIMEOUT_MS.toString()
);

var audioConfig  = SpeechSDK.AudioConfig.fromDefaultMicrophoneInput();
recognizer = new SpeechSDK.SpeechRecognizer(speechConfig, audioConfig);
new Promise(function(resolve) {
    recognizer.onend = resolve;
    recognizer.recognizeOnceAsync(
        function (result) {
            recognizer.close();
            recognizer = undefined;
            resolve(result.text);
        },
        function (err) {
            alert(err);
            recognizer.close();
            recognizer = undefined;
        }
    );
}).then(r => {
    console.log(`Azure STT interpreted: ${r}`);
});

额外说明 #

• recognizeOnceAsync为单次短语音识别设计，若需处理长语音或频繁停顿的场景，更推荐使用startContinuousRecognitionAsync，配合自定义结束逻辑控制识别周期。
• 超时时长最大值受Azure Speech服务限制，当前最大可设置为10000ms（10秒），超出范围会被服务端自动调整为默认值。

内容的提问来源于stack exchange，提问作者Steinn Hauser Magnússon