Noveltyslice won't match expected behavior

Unless I’m doing something wrong, I can’t seem to get noveltyslice to slice this sound meaningfully…

(btw how can I attach a sound on discourse? it says only jpeg/png are allowed)

To attach a sound we’re going through a shared folder for now:

Understood, thank you!

Hmm, this is a tricky sound for this object. The results I get with kernelsize=7 and threshold=0.35 aren’t completely unreasonable, but it does get noisy in the low energy parts.

It’s likely that STFT magnitudes aren’t the right feature for this sort of material. Maybe MFCCs or similar would be better (not that this is an option yet). When we were exploring different change point detection algorithms, it was evident that matching the feature to the material was important, but not always intuitive.