Brian McFee

> I suspect it might have something to do with how clipping paths work in annotation objects with transformed bboxes. Confirmed locally: mpl 3.3 does not clip annotations properly with...

It looks like the failing python 3.12 builds are likely due to an old version of freetype (2.6.1) ; the older python versions (3.11, 10) have the more recent freetype...

Implemented the freetype workaround and the only remaining failure here is in source separation (osx) and is unrelated.

> Does this cause an error or violated assumption later on, or do things work out? It depends on the metric. Many of the metrics convert (labeled) intervals to samples...

> Sorry, in that question I was referring specifically to the check in `chord`, not any of the other tasks. Should the chord metrics have their intervals rounded just for...

> Or for consistency and because it's probably the right thing to do? Or what? Yeah, sorry -- if we round the timing to validate it, we should compute on...

Just a quick benchmark on my laptop using our fixtures, TLDR is we can cut FFT costs about in half, just by using rfft and a smarter length calculation: ```...

This is interesting. Tests are essentially all passing, except for framewise image on dataset number 5. I believe this case has been a heisenbug in the past, at least on...

Interestingly, dataset06 also fails on my dev machine, but only by 0.02dB (#376 issue).

Ok, I've dug into this failure a little more. It's definitely a numerical stability issue. The offending case is `ref07` (numbering doesn't line up with fixture numbering since we removed...