Pariente Manuel

There's also torchaudio's resample, should we should between both?

Ok, it's pretty clear that Julius is better, let's stick with it !

We do this in asteroid with a similar setup, maybe it's simple to just copy paste that part

I can, but in two weeks 😉

I usually have 64GB or more of RAM but quite slow I/O, so yes, it makes complete sense. > Or, if the background noise collection is too big to fit...

Maybe better than loading everything at the beginning, yes. So we don't have to wait when we instantiate the class.

Related question: we should probably include STFT in the deterministic transforms. Should we support torchaudio's and asteroid's?

That's what I have in mind at least. Everything is a transform, and augmentations are just probabilistic ones.

> Sounds good to me Cool. So I'll think about outsourcing asteroid's filterbank API to another repo then.

As said in the thread, I'd vote for supporting the use of `eval` to enable the user to choose his distribution. Copied from the PR ```python # Easy normal way...