9224 · timbreloom
Timbreloom — a DDSP re-voicer you hear and see
Sing or play into the mic. Timbreloom extracts your live pitch and loudness, then rebuilds the sound from a bank of 64 harmonic partials plus a filtered noise band — a harmonic-plus-noise DDSP decoder. Your melody and dynamics survive; your timbre is replaced by the chosen voice. Silence in, silence out.
f0 — Hz · loud 0% · DEMO
Instrument voice
bell-like: strong fundamental, sparse ringing overtones, airy hiss
How it works
Per frame, a YIN-lite autocorrelation over a 2048-sample buffer yields the fundamental pitch and an RMS envelope yields loudness. These two control signals drive a band-limited additive oscillator (its harmonic amplitudes are the voice's spectral-envelope preset) and a noise band shaped by a biquad filter — the harmonic-plus-noise model from DDSP. The glowing ember column is the live 64-partial stack drifting through time: brightness is each partial's current amplitude, so you watch your timbre.
Refs: Engel, Hantrakul, Gu, Roberts — “DDSP: Differentiable Digital Signal Processing” (ICLR 2020); arXiv:2503.11562, “Designing Neural Synthesizers for Low-Latency Interaction” (2025).