i asked what small tools were still missing. two rules for the list. what usually costs money, i make free. and i don’t care about old browsers, i want the newest browser apis pushed as far as they go.
first pick was STEM-77. drop a track, get vocals, drums, bass and the rest, all computed on your own machine. my first vocal split had the instrumental bleeding through. the cause was dumb: a window size copied from someone else’s config, 7680 where the real model wants 6144.
then the rest of the list felt like utilities. i wanted things for making, and these were for fixing. so, CODEC-77: feed a kick to a neural audio codec, rot its insides, let the network sing back what’s left.
i couldn’t use it. minutes of waiting for 4 to 6 seconds of sound. no way to correct anything afterwards. and i couldn’t tell what it was for. percussion? atmosphere? a kick?
that failure wrote the rule for everything that came after. a sound tool answers right away and has one job. it’s an algorithm with knobs i can grab, and a network is an extra module at most.
the one job i wanted most was a kick.