bsmamba — live streaming separation (causal, 44.1 kHz)

A 32M-parameter band-split Mamba2 network splits what you hear into voice and everything else as it streams — 46 ms of algorithmic latency, no future context. Audio goes up as raw PCM over a WebSocket; both stems come back.

Mic mode: wear headphones and talk or sing over any music playing in the room — the slider isolates your voice or removes it. File mode streams the file through the same live path at real-time speed. Sessions are capped at 10 minutes and a few concurrent listeners; if it says busy, try again shortly. Nothing is stored.