well I just derived them empirically– you can test the coefficients yourself by plugging the values into filtergraph and see what sort of a curve you get. I did test the filter against a class-A spl meter to see how accurate it was and it seemed to be within less than 1 dB margin of error, which was adequate for my purposes at the time.

as for the other question, you could design a set of cascading fft modules using pfft~, each of which only samples bins 4 to 6 of a different window-size (ie 4096, 2048 ……128, 64), one for each octave

T