Waveform Augmentation

[faroit]

as opposed to #44 there are a number of effective augmentations that could be done in the time domain, namely

• gain augmentation: applying a random gain to the waveform
• channel swapping (for stereo models): swap channels with p=0.5
• channel selection (for mono models): randomly choose left or right channel instead of downmixing both channels
• noise: add gaussian noise to the signal
• time stretching: the good old way of doing time stretching in the time domain using overlap-add or waveform similarity overlap add, like done in all these hardware instrument samplers.

some of them can be found in audtorch but nonetheless I think these are helpful. What do you think?

[keunwoochoi]

Looks nice overall. I'd like to think about channel selection, seems like it's somehow different from all the others. All the rest seem relevant and useful.

[hagenw]

Thanks for starting this discussion. For the beginning I would add two more, that I found useful so far:

• data set: add waveforms from another data set, e.g. audtorch.transforms.RandomAdditiveMix. noise could be an extra transform or part of this one (dependening if we have noise data sets or not)
• cropping, padding: randomly crop or pad the waveform (or is this not called augmentation?)