c# - 如果说话者远离麦克风，Google Speech/NAudio 会有很大的延迟

Question

我正在使用 google speech api 和 NAudio(使用 NAudio WaveInEvent 类)进行语音到文本的转换。像这样:https://cloud.google.com/speech-to-text/docs/streaming-recognize?hl=en (“对音频流执行流式语音识别”的 C# 示例)

如果说话的人离麦克风很近，一切都很快。但是如果说话的人离麦克风很远，他的前 3-5 个词就不会被识别。之后，其他词就被很好地识别了。 (所以它不可能是距离的一般问题)更像是对距离的适应问题，或者 NAudio 可能没有在 100% 音量输入下录制。

对这个问题有什么想法吗？

编辑:这是要求的代码:

static async Task<object> StreamingMicRecognizeAsync(int seconds)
{
    if (NAudio.Wave.WaveIn.DeviceCount < 1)
    {
        Console.WriteLine("No microphone!");
        return -1;
    }
    var speech = SpeechClient.Create();
    var streamingCall = speech.StreamingRecognize();
    // Write the initial request with the config.
    await streamingCall.WriteAsync(
        new StreamingRecognizeRequest()
        {
            StreamingConfig = new StreamingRecognitionConfig()
            {
                Config = new RecognitionConfig()
                {
                    Encoding =
                    RecognitionConfig.Types.AudioEncoding.Linear16,
                    SampleRateHertz = 16000,
                    LanguageCode = "en",
                },
                InterimResults = true,
            }
        });
    // Print responses as they arrive.
    Task printResponses = Task.Run(async () =>
    {
        while (await streamingCall.ResponseStream.MoveNext(
            default(CancellationToken)))
        {
            foreach (var result in streamingCall.ResponseStream
                .Current.Results)
            {
                foreach (var alternative in result.Alternatives)
                {
                    Console.WriteLine(alternative.Transcript);
                }
            }
        }
    });
    // Read from the microphone and stream to API.
    object writeLock = new object();
    bool writeMore = true;
    var waveIn = new NAudio.Wave.WaveInEvent();
    waveIn.DeviceNumber = 0;
    waveIn.WaveFormat = new NAudio.Wave.WaveFormat(16000, 1);
    waveIn.DataAvailable +=
        (object sender, NAudio.Wave.WaveInEventArgs args) =>
        {
            lock (writeLock)
            {
                if (!writeMore) return;
                streamingCall.WriteAsync(
                    new StreamingRecognizeRequest()
                    {
                        AudioContent = Google.Protobuf.ByteString
                            .CopyFrom(args.Buffer, 0, args.BytesRecorded)
                    }).Wait();
            }
        };
    waveIn.StartRecording();
    Console.WriteLine("Speak now.");
    await Task.Delay(TimeSpan.FromSeconds(seconds));
    // Stop recording and shut down.
    waveIn.StopRecording();
    lock (writeLock) writeMore = false;
    await streamingCall.WriteCompleteAsync();
    await printResponses;
    return 0;
}

来源:https://cloud.google.com/speech-to-text/docs/streaming-recognize?hl=en

Answer 1

是的，这就是事情的运作方式。引擎使用对声音级别的适应，如果级别太低，它们只会错过第一个单词，并且只有在适应后才会开始识别。准确性将低于预期。

要解决此问题 - 使用更先进的麦克风阵列来跟踪音频源，如 Respeaker 或 Matrix，并且可能使用自定义语音识别系统，该系统对快速音频电平变化更稳健。它也将比 Google API 便宜。