InverseSpectrogram transform

🚀 Feature

I’d like to propose adding of InverseSpectrogram transform to torchaudio

Motivation

istft is already present in torchaudio.functional and if users want to build a cepstrum, they have to wrap the functional method into a torch module themselves.

Pitch

Adding a wrapper for functinal.istft, similar to how Spectrogram transform wraps F.spectrogram

Additional context

This thought arose from the fact that I was thinking about quick ways to implement various non-mel scale cepstral coefficients in pytorch (LFCC, CQCC, LPCC, BFCC, etc). Implementing all the various filterbanks seemed like it would overload the library. Adding the transform would alleviate the complexity needed to extract various cepstral coefficients.

[dvisockas]

You’re correct that they should be close. The question is how close. Will open a PR an we can see there how close they are.

[vincentqb]

Good point 😃 Yes, that sounds like a good addition.

For testing, how close are the original waveform and the reconstructed waveform? Since stft and istft are close, I would expect Spectrogram and InverseSpectrogram to be close. Is that correct? Otherwise, do you have a reference to match? I don’t see it in librosa.