Bayesian VAD: Fixing turn detection in production voice pipelines | Decagon

This post introduces a Bayesian approach to Voice Activity Detection (VAD) and turn detection in voice AI pipelines. It addresses premature turn endings by first calibrating VAD probabilities using isotonic regression and then replacing frame-based heuristics with a Bayesian hazard statistic. This statistic accumulates evidence of silence and the likelihood of a true turn end, leading to a 42% reduction in broken turns and a 20% reduction in Pipeline WER.