Error NaN in nvidia speaker verefication

[Asasen1]

Hello, could you please tell me what the problem might be? I'm using the nemo_en_titanet_large.onnx model, and about 50% of the time, calling the Sherpa onnx C# API returns a NaN array. This is regardless of the audio length. I'm using 8kHz WAV files. I tried not using Sherpa's built-in resampling, but the result is the same at 16kHz. Examples are attached.

using NAudio.Wave;
using SherpaOnnx;

var config = new SpeakerEmbeddingExtractorConfig
{
	Model = @"C:\engine\sv\nemo_en_titanet_large.onnx",
	NumThreads = 3,
	Debug = 1
};
using var extractor = new SpeakerEmbeddingExtractor(config);

var computed = ComputeEmbedding(@"C:\Users\Артемий\Downloads\VoiceDataset\SR-template-8khz\1_01_03_25.wav");

float[] ComputeEmbedding(string filename)
{
	using var reader = new WaveFileReader(filename);
	var provider = reader.ToSampleProvider();
	using var stream = extractor.CreateStream();

	var buffer = new float[4096];
	while (provider.Read(buffer, 0, buffer.Length) > 0)
	{
		stream.AcceptWaveform(reader.WaveFormat.SampleRate, buffer);
	}
	stream.InputFinished();

	return extractor.Compute(stream);
}

Link to example audio: https://transfiles.ru/0avix

[dosubot]

The nemo_en_titanet_large.onnx model expects 16kHz, 16-bit, mono audio input. There's no explicit support for 8kHz audio, and using it can cause issues like NaN outputs. Even though Sherpa ONNX can resample internally, the safest approach is to ensure your input WAV files are already 16kHz, 16-bit, mono before passing them to the API. Try using a trusted external tool to resample your audio and see if the problem persists. Also, check your logs for any warnings or errors when NaNs occur, and compare files that work with those that don't to spot any anomalies. More details on input requirements are in the official docs.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[csukuangfj]

Please show us which version of sherpa-onnx you are using.

[Asasen1]

1.12.17

[Asasen1]

Please show us which version of sherpa-onnx you are using.

Do you have any ideas what the problem might be?