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

What this post added

This post introduces a new approach to improving turn detection in voice AI pipelines by addressing two key issues: miscalibrated VAD probabilities and a flawed decision framework for turn endings. The authors applied isotonic regression to recalibrate VAD model outputs, significantly reducing calibration error. They also developed a Bayesian hazard statistic that replaces heuristic-based silence detection and timeouts with a learned statistic that accumulates evidence of a true turn end. This new statistic smooths out within-speech dips and responds decisively to genuine silence, leading to a 42% reduction in broken turns and a 20% reduction in pipeline Word Error Rate (WER).