Someone linked this over on the lines forum, in the Kaizo Snare post I had made there and thought it would be posting here:
https://www.nime.org/proceedings/2015/nime2015_147.pdf
The paper is a bit old now (2015), and he’s using 50ms analysis windows, with an FFT-based onset detector, so the exact implementation is orders of magnitude too slow for my purposes, but good to read through nonetheless.
Descriptor-wise it looks pretty vanilla: loudness, centroid, ZCR, and he talks about incorporating spread/skewness/kurtosis as well. Then beyond that it just goes into a kNN classifier.
I’m surprised it’s not 100% given the analysis window and sounds used (three very different cajon sounds), but it’s pretty good results (95%) either way.