nicktup/reverb-extractor

Reverb Removal

Take the room out of a recording — live. Pick a reverberant clip, press Remove reverb, and a compact regularised inverse filter (built from the room's known impulse response) strips the reverb on the spot and reports the intelligibility gain it just achieved. Cheap (a few hundred parameters, hundreds of times faster than real time on one CPU core), explainable, and bounded so it runs live without ever stalling the node.

A secondary tab demos reverb extraction: a differentiable Feedback Delay Network with learnable delay lines fits a room's impulse response, turning the room into an editable, real-time reverb effect. PyTorch reimplementation of Mezza et al. 2024 — "Data-driven room acoustic modeling via differentiable feedback delay networks with learnable delay lines" (arXiv:2404.00082).

• —Code: https://github.com/nickmartins-lambda/differentiable-fdn
• —Interactive write-up (playable audio): https://nickmartins-lambda.github.io/differentiable-fdn/

Scope — honest. This Space demonstrates the informed case: the room response is known, and removal is excellent — a mid-size room (T60 ≈ 1.1 s) goes from speech intelligibility STOI 0.50 → 0.91, and even a cathedral-length 2.4 s tail cleans up 0.32 → 0.82. Blind removal, from the recording alone, is the open frontier and the next model in the pipeline. All inputs are curated examples with bounded compute — nothing runs unbounded on the free CPU node.