Incorrect recognition with wrong data format

[BlackHawkCH91]

This is my first time using Vosk, so please bear with me. I'm using .NET Core 6 with C# 10, Vosk 0.3.38 and the "vosk-model-en-us-0.22-lgraph" model (renamed the folder to "model"). The model appears to be loading fine, with no errors or warnings showing up.

The attached audio file says: static flexible rubber soles static

However, Vosk always outputs:

{
  "partial" : ""
}
{
  "partial" : ""
}
. . .
{
  "partial" : "the"
}
{
  "partial" : "the"
}
. . .
{
  "result" : [{
      "conf" : 1.000000,
      "end" : 4.859122,
      "start" : 0.140574,
      "word" : "the"
    }],
  "text" : "the"
}
{
  "partial" : ""
}
{
  "partial" : ""
}
. . .
{
  "result" : [{
      "conf" : 1.000000,
      "end" : 11.854490,
      "start" : 7.685259,
      "word" : "the"
    }],
  "text" : "the"
}

The output is the same even when using different audio files. I have also tried using the "vosk-model-small-en-us-0.15" model, but the output was mostly the same.

Here is the code for speech recognition:

void SpeechRec(string modelPath, string audioPath)
{
    //Convert mp3 to wav
    using (Mp3FileReader mp3 = new Mp3FileReader(audioPath))
    {
        using (WaveStream pcm = WaveFormatConversionStream.CreatePcmStream(mp3))
        {
            WaveFileWriter.CreateWaveFile(audioPath, pcm);
        }
    }

    //Create Vosk STT
    Model model = new Model(modelPath);
    VoskRecognizer rec = new VoskRecognizer(model, 16000);
    rec.SetMaxAlternatives(0);
    rec.SetWords(true);

    using (Stream source = File.OpenRead(audioPath))
    {
        byte[] buffer = new byte[4096];
        int bytesRead;
        while ((bytesRead = source.Read(buffer, 0, buffer.Length)) > 0)
        {
            if (rec.AcceptWaveform(buffer, bytesRead))
            {
                Console.WriteLine(rec.Result());
            }
            else
            {
                Console.WriteLine(rec.PartialResult());
            }
        }
    }
    Console.WriteLine(rec.FinalResult());
}

To use the audio file, change the file extension from .mp4 to .mp3.

https://user-images.githubusercontent.com/49353890/179464547-3ee3a00e-2941-4443-b2a2-14165b21354f.mp4

Vosk console/debug messages.

output.txt

[nshmyrev]

Your audio is stereo, you also need to convert it to mono

[BlackHawkCH91]

Yep that was the issue, though it seems that it's quite inaccurate, now outputting: "the works should be within the rules" instead of "flexible rubber soles". Might need to experiment with the models.

Thanks for answering!

EDIT:

yeah, the accuracy depends on the model. Changed it to a different one and it was more accurate.

[nshmyrev]

Yep that was the issue, though it seems that it's quite inaccurate, now outputting: "the works should be within the rules" instead of "flexible rubber soles".

Sample rate is also wrong, the file has 22khz,not 16. If you fix everything it should output "flexible rubber soles" like with vosk-transcriber command line utility

[nshmyrev]

Even with a small model

[BlackHawkCH91]

Yep, it's a lot more accurate. Thanks for the help. I'm now getting the sample rate of the file and then passing it to the VoskRecogniser.