Xamarin Android中如何为TextToSpeech与SpeechRecognizer集成AcousticEchoCanceler

根因说明 #

当前使用的SpeechRecognizer默认采用通用语音识别音频配置，未启用回声消除（AEC）能力，会直接拾取设备自身扬声器播放的TTS内容，触发识别回调导致误停止TTS。

前置配置 #

首先在AndroidManifest.xml中添加必需权限：

<uses-permission android:name="android.permission.RECORD_AUDIO" />
<uses-permission android:name="android.permission.MODIFY_AUDIO_SETTINGS" />

步骤1：启用系统回声消除能力 #

方案A：优先使用通信音频模式（适配90%以上设备） #

修改SpeechService的两处代码：

1. 调整CreateSpeechIntent方法，指定音频源为通话模式，系统会自动开启AEC、降噪等预处理：

protected virtual Intent CreateSpeechIntent(bool partialResults)
{
    var intent = new Intent(RecognizerIntent.ActionRecognizeSpeech);
    intent.PutExtra(RecognizerIntent.ExtraLanguagePreference, Java.Util.Locale.Default);
    intent.PutExtra(RecognizerIntent.ExtraLanguage, Java.Util.Locale.Default);
    intent.PutExtra(RecognizerIntent.ExtraLanguageModel, RecognizerIntent.LanguageModelFreeForm);
    intent.PutExtra(RecognizerIntent.ExtraCallingPackage, Application.Context.PackageName);
    intent.PutExtra(RecognizerIntent.ExtraPartialResults, partialResults);
    // 新增：指定音频源为语音通信模式，自动启用AEC
    intent.PutExtra(RecognizerIntent.ExtraAudioSource, (int)Android.Media.AudioSource.VoiceCommunication);
    return intent;
}

2. 调整Listen方法，启动识别前切换音频模式，释放时恢复默认模式，同时增加短结果过滤：

[Obsolete]
public IObservable<string> Listen(Action action = null) => Observable.Create<string>(ob =>
{
    var audioManager = (AudioManager)Application.Context.GetSystemService(Context.AudioService);
    // 切换到通信模式触发系统AEC
    audioManager.Mode = Mode.InCommunication;
    
    speechRecognizer = SpeechRecognizer.CreateSpeechRecognizer(Application.Context);
    var listener = new SpeechRecognitionListener(); 
    listener.ReadyForSpeech = () => this.ListenSubject.OnNext(true);
    listener.PartialResults = sentence =>
    {
        lock (this.syncLock)
        {
            sentence = sentence.Trim();
            // 新增：过滤空结果和过短的识别结果，减少误触发
            if (string.IsNullOrEmpty(sentence) || sentence.Length < 2) return;
            
            if (action != null)
            {
                action.Invoke();
            }
            if (currentIndex > sentence.Length)
                currentIndex = 0;
            var newPart = sentence.Substring(currentIndex);
            currentIndex = sentence.Length;
            final = sentence;
        }
    };
    listener.EndOfSpeech = () =>
    {
        ob.OnNext(final);
        ob.OnCompleted();
        this.ListenSubject.OnNext(false);
    };
    speechRecognizer.SetRecognitionListener(listener);
    speechRecognizer.StartListening(this.CreateSpeechIntent(true));
    
    return () =>
    {
        audioManager.SetStreamMute(Stream.Notification, false);
        // 恢复默认音频模式
        audioManager.Mode = Mode.Normal;
        stop = true;
        speechRecognizer?.StopListening();
        speechRecognizer?.Destroy();
        this.ListenSubject.OnNext(false);
    };
});

方案B：显式启用AcousticEchoCanceler（适配不支持自动AEC的设备） #

如果方案A在部分设备不生效，需要自定义音频采集链路，显式初始化AEC：

1. 先检测设备AEC支持性：

private bool IsAecSupported => AcousticEchoCanceler.IsAvailable;

2. 创建自定义AudioRecord实例，绑定AEC：

private AudioRecord InitAudioRecordWithAec()
{
    int sampleRate = 16000;
    var channelConfig = ChannelIn.Mono;
    var audioFormat = Encoding.Pcm16bit;
    int bufferSize = AudioRecord.GetMinBufferSize(sampleRate, channelConfig, audioFormat);
    
    var audioRecord = new AudioRecord(AudioSource.VoiceCommunication, sampleRate, channelConfig, audioFormat, bufferSize);
    
    if (IsAecSupported)
    {
        var aec = AcousticEchoCanceler.Create(audioRecord.AudioSessionId);
        aec.Enabled = true;
        // 同时可以启用噪声抑制进一步优化识别效果
        if (NoiseSuppressor.IsAvailable)
        {
            var ns = NoiseSuppressor.Create(audioRecord.AudioSessionId);
            ns.Enabled = true;
        }
    }
    return audioRecord;
}

后续将AudioRecord采集的PCM数据流喂给SpeechRecognizer即可，该方案需要自己实现音频帧的写入逻辑，兼容性更强。

步骤2：添加业务层双重校验 #

修改MyViewModel的触发逻辑，增加相似度校验，避免极端场景的误触发：

// 存储当前正在播报的TTS文本，用于相似度校验
private string _currentSpeakText;

public MyViewModel ()
{
    var speakText = "Xamarin is a Microsoft-owned San Francisco-based software company founded in May 2011 by the engineers that created Mono, Xamarin.Android and Xamarin.iOS, which are cross-platform implementations of the Common Language Infrastructure and Common Language Specifications.";
    _currentSpeakText = speakText;
    speak(speakText);
    
    Action actionAfterSpeechDetect = delegate
    {
        if (textToSpeechCancellationToken != null && !textToSpeechCancellationToken.IsCancellationRequested)
        {
            // 新增：和当前播报的TTS文本做相似度匹配，避免回声误触发
            var similarity = GetSimilarity(final, _currentSpeakText);
            if (similarity < 0.7) // 相似度低于70%才判定是用户输入
            {
                textToSpeechCancellationToken.Cancel();
            }
        }
    };
    using (var cancelSrc = new CancellationTokenSource())
    {
        output = await DependencyService.Get<ISpeechRecognizer>().Listen(actionAfterSpeechDetect).ToTask(cancelSrc.Token);
    }    
}

// 简易文本相似度计算方法
private double GetSimilarity(string str1, string str2)
{
    if (string.IsNullOrEmpty(str1) || string.IsNullOrEmpty(str2)) return 0;
    int sameCharCount = str1.Intersect(str2).Count();
    return (double)sameCharCount / Math.Max(str1.Length, str2.Length);
}

内容的提问来源于stack exchange，提问作者Thamotharan G