jonashaag / audio-resampling-in-python Goto Github PK

Same as julius. See pytorch/audio#1057

https://github.com/jthiem/nnresample

Hi @jonashaag, thanks for compiling and updating this notebook on resampling libraries! However, I noticed there was a small bug in the new torchaudio examples, and also wanted to update you on some new resampling capabilities from our latest release.

I noticed that there was a bug in the notebook with the torchaudio_new_resample arguments that led to it producing strange plots for both the upsample and downsample examples. Removing [None] and [0] from torchaudio_new_resample(torch.from_numpy(sig[None]), P, Q).numpy()[0] should resolve this issue, and both resampling methods in torchaudio should produce the same results.

In the latest torchaudio release (v0.9.0), we made several improvements to resampling. We improved the speed performance of transforms.Resample, deprecated kaldi.resample_waveform in favor of functional.resample, added kaiser window support, among others. If you're interested, you can track the updates here, and we have also created a tutorial demonstrating the properties of our improved resampling function.

What are your thoughts on the following?

• fixing the torchaudio_new_resample bug
• using the latest version of torchaudio (0.9.0) (Note that for the latest version,transforms.Resample should take in an additional dtype=torch.float64 argument in your examples)
• plotting with a higher quality set of parameters (ex/ using lowpass_filter_width=64 or resampling_method="kaiser_window")

Let me know if you would like to work on this, or if you'd like me to send in a PR, I'd be happy to do so as well!

cc @mthrok

Hi @jonashaag , I was looking at the impulse response and trying to understand why it was looking so wrong with Julius, and why so little attenuation was present above the old Nyquist frequency.

Well, it seems you gave the arguments in the wrong order for Julius, it should be old_sr, new_sr, just as resampy, but you passed the opposite.