使用Azure Speech SDK为discord-speech-recognition添加转写功能无文本输出问题

我正在尝试为npm库discord-speech-recognition添加基于Azure Speech SDK的语音转文本功能，但无论是调用API还是使用SDK都没得到预期结果。目前仅能获取音频缓冲长度，无法得到转写文本。

实现代码 #

const sdk = require("microsoft-cognitiveservices-speech-sdk");

function getAzureRequestOptions(options) {
    const speechConfig = sdk.SpeechConfig.fromSubscription(options.key, options.region);
    if (options.lang) {
        speechConfig.speechRecognitionLanguage = options.lang;
    }
    return speechConfig;
}

function resolveSpeechWithAzureSpeechToText(audioBuffer, options) {
    console.log('resolveSpeechWithAzureSpeechToText called with audioBuffer length:', audioBuffer.length);
    const pushStream = sdk.AudioInputStream.createPushStream();

    pushStream.write(audioBuffer);
    pushStream.close();

    const speechConfig = getAzureRequestOptions(options);
    const audioConfig = sdk.AudioConfig.fromStreamInput(pushStream);
    const recognizer = new sdk.SpeechRecognizer(speechConfig, audioConfig);

    return new Promise((resolve, reject) => {
        console.log('Starting recognition...');
        recognizer.recognizeOnceAsync(
            (result) => {
                console.log('Recognition succeeded');
                recognizer.close();
                resolve(result.text);
                console.log(result);
            },
            (err) => {
                console.log('Recognition failed');
                recognizer.close();
                reject(err);
                console.log(err);
            }
        );
    });
}

module.exports.resolveSpeechWithAzureSpeechToText = resolveSpeechWithAzureSpeechToText;

控制台输出截图 #

问题分析与修复方案 #

1. 音频流关闭时机错误

你在写入音频缓冲后立刻关闭了pushStream，但recognizeOnceAsync还未完成音频读取操作，导致Azure SDK无法获取完整的音频数据。应该在识别完成后再关闭流。

2. 音频格式不匹配

Discord输出的音频通常是PCM 16位、48kHz单声道格式，而Azure SDK需要明确匹配的音频格式参数，否则无法正确解析音频数据。

3. 日志打印顺序问题

原代码中先执行resolve/reject再打印结果/错误，可能导致无法及时看到完整的调试信息，应该先打印再处理Promise状态。

修改后的代码 #

const sdk = require("microsoft-cognitiveservices-speech-sdk");

function getAzureRequestOptions(options) {
    const speechConfig = sdk.SpeechConfig.fromSubscription(options.key, options.region);
    if (options.lang) {
        speechConfig.speechRecognitionLanguage = options.lang;
    }
    return speechConfig;
}

function resolveSpeechWithAzureSpeechToText(audioBuffer, options) {
    console.log('resolveSpeechWithAzureSpeechToText called with audioBuffer length:', audioBuffer.length);
    
    // 明确设置音频格式，匹配Discord的PCM 48kHz单声道16位参数
    const audioFormat = sdk.AudioStreamFormat.getWaveFormatPCM(48000, 16, 1);
    const pushStream = sdk.AudioInputStream.createPushStream(audioFormat);

    pushStream.write(audioBuffer);
    // 暂时不关闭流，等待识别完成后操作

    const speechConfig = getAzureRequestOptions(options);
    const audioConfig = sdk.AudioConfig.fromStreamInput(pushStream);
    const recognizer = new sdk.SpeechRecognizer(speechConfig, audioConfig);

    return new Promise((resolve, reject) => {
        console.log('Starting recognition...');
        recognizer.recognizeOnceAsync(
            (result) => {
                console.log('Recognition succeeded');
                console.log(result); // 先打印结果再处理
                recognizer.close();
                pushStream.close(); // 识别完成后关闭流
                resolve(result.text);
            },
            (err) => {
                console.log('Recognition failed');
                console.log(err); // 先打印错误再处理
                recognizer.close();
                pushStream.close();
                reject(err);
            }
        );
    });
}

module.exports.resolveSpeechWithAzureSpeechToText = resolveSpeechWithAzureSpeechToText;

额外检查项 #

• 确认Azure Speech资源的密钥、区域配置正确，且资源处于可用状态；
• 验证Discord传入的audioBuffer是否为原始PCM数据，如果是Opus编码格式，需要先解码为PCM再传给Azure SDK；
• 可以将audioBuffer保存为本地WAV文件，通过Azure Speech Studio测试该文件是否能正常转写，排除音频数据本身的问题。

内容的提问来源于stack exchange，提问作者Kamykalaur