bsmamba — live speaker extraction (causal, 44.1 kHz)

Give it three seconds of one person's voice. It then pulls that speaker out of whatever the microphone hears — other talkers, room reverb, noise — as it streams, with 46 ms of algorithmic latency. Trained on rooms, not studios.

Live mode uses the causal model (hears nothing ahead of now). File mode uses the bidirectional model on the whole file for the highest quality. The "speaker match" bar is the server's absence gate: when the enrolled voice isn't there, the extracted output is muted and everything goes to "everyone else". Sessions are capped at 10 minutes and a few listeners; files at 2 minutes. Nothing is stored.