Nice – I’ll take a closer look at what you did. I decided to play about with the activations here because it seemed like it might work for distinguishing between interestingly distinct bits of the sound, and we’d floated it amongst ourselves before.
I’m not completely sold on its effectiveness, but interesting to compare / combine with the spectral clustering. It also made me wonder how a continuous wavelet transform would work on this sound (as I’ve been playing around with the pipo one, and Trying to Learn). It think to use it on full audio is a sort of time-is-no-object proposition, but I might give it a shot in python and see what happens.