Stem splitting with pure math. No AI.
Unmix splits it into drums, bass, vocals and everything else, then lets you mix them like a DJ. Works on MP3, WAV and M4A, up to the first 90 seconds.
A Fourier transform turns each 46 ms slice of sound into its frequencies. Stack the slices and steady notes draw horizontal lines while drum hits draw vertical ones.
A median filter running up each column keeps short, wide bursts and erases steady tones. Whatever survives is percussion.
Of the steady tones, everything under about 120 Hz goes to bass, fading out by 260 Hz.
Lead vocals are mixed dead center, so they're identical in both speakers. Tones in the voice range where left ≈ right go here.
Everything left: synths, guitars, pads, and anything panned to the sides. The four stems always add back up to the original exactly.
The drum stem's energy jumps give an onset curve. Autocorrelation finds the gap that repeats most strongly, and that gap is the beat.