Implementation proposal for audio resample

[AlexFuster]

At Voicemod, we have made a tensorflow version of the audio resampling function from torchaudio https://pytorch.org/audio/stable/transforms.html?highlight=resample#torchaudio.transforms.Resample.

I think It would be a good alternative/complement for the current implementation of audio resampling of this repo https://www.tensorflow.org/io/api_docs/python/tfio/audio/resample for 4 reasons.

1. Supports batches: I know that this was recently added to the current implementation)
2. Supports gradients: Unlike the current implementation. This feature is very useful if you want to resample inside a training graph
3. The results are as good as the ones given by the current implementation and they are close to resampy
4. For some reason, the current implementation always adds a misalignment of size 32 or so, as one can see in the image below
   

Tell me if you are interested

[yongtang]

@AlexFuster The resample was implemented wit speex in the current repo. If an alternative implementation brings additional advantage (like matching resampy/etc) then we certainly welcome PRs.

[AlexFuster]

So, do I replace the resample function in audio_ops.py or do I create a separate function?

[yongtang]

@AlexFuster You can replace the resample function in audio_ops.py and replace C++ kernel in https://github.com/tensorflow/io/blob/master/tensorflow_io/core/kernels/audio_kernels.cc#L220-L299

[AlexFuster]

But our implementation is fully in python (using Tensorflow operations). I'll just replace the resample function

[yongtang]

@AlexFuster Sure. Please go ahead and submit the PR.

We can decide in the PR if the original C++ kernel can be removed or we want to keep them for now.

[AlexFuster]

Sorry, I forgot to add an import to the first one