> whera a = 384, b = 0.33, c = 0.17
So this is some sort of compromise to increase speed? Have you thought about implementing in WASM?
> Probably, it needs some sort of machine learning.
Agreed, but for a visualisation, it could just be a parameter just for the user to mess with.
Partially. The main purpose is to increase time domain accuracy on low frequency. If you want to do experiment on window length (in time domain), use ffmpeg showcqt filter (https://ffmpeg.org/ffmpeg-filters.html#showcqt). showcqt-js window length is hardcoded to tc=0.33 attack=0.033 and tlength='st(0,0.17); 384tc / (384 / ld(0) + tcf /(1-ld(0))) + 384tc / (tcf / ld(0) + 384 /(1-ld(0)))' (https://github.com/mfcc64/mpv-scripts/blob/a0cd87eeb974a4602...).
> Have you thought about implementing in WASM?
It is implemented in WASM.
> Agreed, but for a visualisation, it could just be a parameter just for the user to mess with.
If you want to do experiment on it, showcqt.js exposes intermediate color data cqt.color[]. It can be modified arbitrarily, including autocorrelation. If some day I do experiment on it and find that the result is satisfying, maybe I'll include it in YouTube Musical Spectrum.
I thought so. The main issue with regular CQT is latency on low frequencies due to the excessive window required. It makes it unsuitable for real- time applications, as you know. Thanks for the insights