Hi, I’m using the novelty algorithm mostly for slicing. I wondered if there was a way to control certain parameters to get more cohesive slices. i.e. summarise slices of the same pitch or summarise a single gesture like a glissando upwards, or just the tail of an impulse. i wrote a script that manually reads all the slices and applies certain criteria to do this. it just takes much more time and i imagined, it could be solved by better controlling the algorithm on its first run, because this is already what it’s meant to do.
looking forward to hear about your solutions, also favorite settings for best results of the novelty slicer.
Are you in Max or Pd or SuperCollider? @jamesbradbury had a great idea, I think in Python, then I stole it in SuperCollider. Both are documented in code and ideas in the Learn FluCoMa in-depth articles written by @jacob.hart
So you don’t have to go read my old writing / code the idea was that you “overslice” the audio file and then group together adjacent slices which are deemed to be too similar. The similarity measure is determined very crudely: I cluster all of the original slices using fluid.bufmfcc~ and fluid.bufstats~ and fluid.kmeans~ and if two adjacent slices are part of the same cluster, they are considered similar. You could do it with distance too, by looking at how far away they are in some kind of UMAP projection, but that’s possibly more work
no difference at all in the FluCoMa part - but I haven’t had the chance to translate that code - it has been on the wishlist for more than 2 years so not a huge priority in my current workflow - if you get down to it, share and we can make a multi-creative-coding-environmnent threat. the dream!