Serra's insight: most sounds are a small set of stable sinusoids (the pitched, tonal part) sitting on a bed of noise (breath, bow scrape, the attack). Find the spectral peaks in each frame — those are the sinusoids, the deterministic part — and whatever's left is the stochastic residual. Because the residual is literally the original spectrum minus the peaks, the two halves add back up to exactly the input. Then you can remix them: all sines for a glassy, de-noised tone; all residual for a breathy, pitchless ghost. This is the honest, scoped build — per-frame peak separation, not full partial-tracking across frames — but the sines/noise split is real. Rendered offline; press render after changing the peak count.
Pull residual to 0 for the de-noised tone; pull sines to 0 for the breathy ghost.
render, then play the mix with your keyboard (a s d f …) or loop it