The trickiest part here was the muddy bass separation in the stock AudioReactive code, and the maker went at it head-on by ditching FFT analysis completely. Instead of a frequency transform, the audio now runs through 31 software filter banks spaced at one-third octave intervals from 20Hz up to 20000Hz. That is a serious rework of the core, and the result shows: every band stands on its own, and the onsets come through clean and clearly separated.
The whole thing runs on an Olimex ESP32-P4-PC board, whose HDMI output drives a monitor in the demo. There is an ES8311 codec on the board itself, with two profiles: one for the onboard mics and one for line-level signal arriving on the TRRS mic input. The effect shown, GEQ PPA, works in decibels, so all the tuning happens in dB rather than raw levels.
What stands out to me is the validation method. A TC Electronic RTA with 31 bands sits beside the effect, letting the maker map each band one-to-one and confirm the output tracks the meter. After setting an attack value of 1 and a decay of 300, the two screens agree closely. The demo does look jittery, but that is explained: dynamics were switched off so the effect pulls as fast as possible for comparison, and the RTA itself smooths internally. The maker describes the approach as extracting clean bands rather than filtering.
One thing still open: the settings need more work before they are friendly for everyday users. The code sits in a GitHub branch, and the maker says an AI coding assistant helped write it and grew the scope along the way. Aiming this at pro use with line-in hardware seems like a sensible direction to me.