Description:
There is a severe bug in the C# translation of torchaudio.transforms.Resample / functional._apply_sinc_resample_kernel when converting audio (e.g., from 22050 Hz to 16000 Hz).
Instead of a cleanly resampled waveform, the output contains repeating blocks of identical values matching the upsampling factor of the kernel (exactly 320 repeating float values).
Root Cause:
The C# implementation of _apply_sinc_resample_kernel translates the original Python code but mishandles tensor dimensions when invoking torch.nn.functional.conv1d.
When downsampling from 22050 to 16000, the calculated kernel has the shape [320, 1, 459].
Because the C# wrapper applies conv1d with a 1D input tensor improperly shaped or matched, the convolution is applied across channels instead of across time.
This results in an output shape of [1, 320, NewSamples].
When reading the tensor back to C# via .data().ToArray(), the data layout is flattened linearly, causing the user to read the 320 kernel-channel values for each time step sequentially. This looks like a static block of 320 repeating values. The required reshaping/transposing to fold the 320 channels back into the time axis is either missing or broken in the C# port.
To Reproduce:csharp
// Input: Float array of 22050Hz audio
float[] signal = ...
using var nativeTensor = torch.from_array(signal).to(torch.ScalarType.Float32);
using var inputTensor = nativeTensor.unsqueeze(0); // [1, Samples]
using var resampler = torchaudio.transforms.Resample(orig_freq: 22050, new_freq: 16000);
using var output = resampler.forward(inputTensor);
// Look at the raw float/byte array: It contains blocks of 320 repeating values
float[] result = output.squeeze(0).ToArray();
Behavior:
The output should match the bit-accurate windowed sinc interpolation of Python's torchaudio.
Description:
There is a severe bug in the C# translation of torchaudio.transforms.Resample / functional._apply_sinc_resample_kernel when converting audio (e.g., from 22050 Hz to 16000 Hz).
Instead of a cleanly resampled waveform, the output contains repeating blocks of identical values matching the upsampling factor of the kernel (exactly 320 repeating float values).
Root Cause:
The C# implementation of _apply_sinc_resample_kernel translates the original Python code but mishandles tensor dimensions when invoking torch.nn.functional.conv1d.
When downsampling from 22050 to 16000, the calculated kernel has the shape [320, 1, 459].
Because the C# wrapper applies conv1d with a 1D input tensor improperly shaped or matched, the convolution is applied across channels instead of across time.
This results in an output shape of [1, 320, NewSamples].
When reading the tensor back to C# via .data().ToArray(), the data layout is flattened linearly, causing the user to read the 320 kernel-channel values for each time step sequentially. This looks like a static block of 320 repeating values. The required reshaping/transposing to fold the 320 channels back into the time axis is either missing or broken in the C# port.
To Reproduce:csharp
// Input: Float array of 22050Hz audio
float[] signal = ...
using var nativeTensor = torch.from_array(signal).to(torch.ScalarType.Float32);
using var inputTensor = nativeTensor.unsqueeze(0); // [1, Samples]
using var resampler = torchaudio.transforms.Resample(orig_freq: 22050, new_freq: 16000);
using var output = resampler.forward(inputTensor);
// Look at the raw float/byte array: It contains blocks of 320 repeating values
float[] result = output.squeeze(0).ToArray();
Behavior:
The output should match the bit-accurate windowed sinc interpolation of Python's torchaudio.