Fluid.bufnmf~ "CV Splitter/Processor"

[rodrigo.constanzo]

@tremblap

it’s not temporal segmentation more as ‘stream’ segmentation. If you check the helpfiles, you get the bass of the synth, then the blipblip, separately.

Yeah, not so much for audio I would feed it…

@tremblap

yes. you can either make a 2 rank seed dict with your seed and a bit or random noise. Don’t forget to put bufnmf flag to mode 1 like for the piano retraining so it starts with the seeds but updates them.

Hmmm, interesting.
So it will still modify the seed(s) you give it yes?
Like there isn’t a way to give it a “golden seed” that it processes everything else around, to make sure it null-sums.
Might produce interesting results either way, but was just thinking that it would be cool to process the shit out one decomposition, then use fluid.bufnmf~ to “fill in the blanks”.

@tremblap

Have you tried 5 times smaller amounts? I’m told the algo converges quite quickly…

Yeah totally. I did one at 10k and it didn’t look too different (from 1k), so tried 50k just for good measure. It would be good to know what the “functional” upper limit is, as in some cases computation time isn’t important, and the best results matter, even if it takes crazy long (e.g. training up classifiers for nmf-ing), but sometimes having something quick (but still good) is preferred, since it would be used in a real-time context.