Brian McFee

Maybe also worth considering https://github.com/data-apis/array-api-compat (as used in sklearn currently for experimental array api support)

This isn't a typo, it's just that graphditty was written (long) before librosa 0.10.0 made positional arguments a requirement. But yes, graphditty could/should be updated to support 0.10.0+

Good call. PRs are welcome. :grin:

Thanks for catching that. It would be an easy enough flag to add, either in the effects wrapper functions or in the phase_vocoder itself. (Or maybe both?) The one subtlety...

Coming back to this, I think it's actually a subtle bug in the phase vocoder implementation, and not just a numerical precision issue. After digging into it, I think we...

I looked into this a bit more, and it's actually even less clear to me what's going on with the errors in the existing implementation. Here's the difference in phase...

Note: this should be improved by properly accounting for expected phase advance between frames; will come back to this later on.

New and improved vocoder implementation: ```python def phase_vocoder(D, *, rate, hop_length=None, n_fft=None, kind='linear'): if n_fft is None: n_fft = 2 * (D.shape[-2] - 1) if hop_length is None: hop_length =...

I did a bit more careful listening to the outputs of my proposed rewrite and the existing implementation, and I'm not yet happy with how the results sound. Gonna try...

Ok, finally had a minute to sit down and think about this a bit more carefully. First, the existing vocoder implementation is a little obfuscated, and does some unnecessary steps...