Open-source RealtimeVAD targets faster end-of-speech detection
A developer released RealtimeVAD, which uses a 10ms energy detector instead of silero voice activity detection to signal end-of-speech sooner. Benchmarked on about 3,200 turns from the author's assistant logs, it was initially about 110ms faster and still about 100ms faster after a fix for quiet speech, at the cost of more false positives.