Griffin-Lim for CQT

Description

The GL implementation added in 0.7 #843 operates on linear spectrograms. At some point in the discussion, we raised the idea of having it work on CQT as well, but dropped the idea to keep things simple.

I just prototyped a CQT-based GL offline and it sounds pretty good. It seems like a useful kind of thing to have in the library. The only question for me is whether we wrap it into the existing implementation, or just have a separate function that’s nearly identical, but with different arguments and calls cqt/icqt instead of stft/istft. I’m leaning toward the latter, since the method is simple enough that a little duplication of code won’t be catastrophic to maintain.

What do folks think?

[bmcfee]

This works for a very narrow value of hop lengths and Q’s, no?

It’s subject to all the limitations of our icqt implementation, so yes.

The latter appears at every iteration of Griffin-Lim. Is there a way to make the UserWarning print only once?

A user can always filterwarnings to 'once': https://docs.python.org/3/library/warnings.html#warnings.filterwarnings

I don’t think we should do that from within the library though, since a warning can pop up from different sources within a session, and only showing the first one would obscure that.

[bmcfee]

I don’t think this is an issue for us, except maybe documentation. You may need to reset your warning filters before applying a new filter, but this depends on your environment (jupyter/ipython/spyder/python interpreter may all behave differently).