Make BaseWaveformTransform able to pad transform outputs that have a different length

[iver56]

This is a prerequisite for transforms that change the length, like time stretching

[hbredin]

Related: it would be nice for such length-changing transforms to expose some kind of API to let the user know about this behavior.

For instance, in pyannote.audio, I usually train models on chunks of fixed 5s length.
When using a time streching transform, I'd still want to have the output of the transform be 5s long.

If the transform speeds the signal by a factor of up to 2, I should therefore be warned to input 10s chunks (so that the output actually is at least 5s long).

So it could expose something like TimeStreching.input_output_length_ratio := 0.5
By default input_output_length_ratio would be 1.0.
Compose.input_output_length_ratio would be the product of composed transforms's input_output_length_ratio attribute.

Does that make sense? Or is this out of the scope of the library?

[iver56]

it would be nice for such length-changing transforms to expose some kind of API to let the user know about this behavior.

Yes, I had the same idea.

An idea is that e.g. time stretching can make the audio shorter or longer, and then RandomCrop can make sure that we get a fixed length afterwards