Noise Suppression for cluster-meet
R&D on removing background noise from live meeting audio without eating the speech underneath it.
The second strand of audio research for cluster-meet, our in-house meeting platform. This one deals with the noise people actually call from inside: keyboards, fans, traffic through an open window, a meeting room with too much echo in it.
Aggressive suppression demos well and is unpleasant to sit in for an hour. Push it and speech starts arriving clipped at the edges, consonants go soft, and whoever has the accent the model handles least well gets the worst of it. So the question isn’t how much noise you can remove. It’s how much you can remove before the voice starts to degrade, and whether that trade-off holds up outside a clean test set.
Same constraint as the TTS work: it has to run in the live path. That rules out anything that wants to see a whole utterance before deciding what to do with it.
innoscripta SE · 2026