Why I measure interview silence with WebAudio, not the Speech API

Short summary

A practical explanation of why speech-to-text event timing is unreliable for detecting silence in browser voice apps, and how to measure it acoustically with a WebAudio graph instead. STT timings are polluted by network latency, buffering, and interim-result churn, producing false silence signals. The solution: read RMS loudness from the raw waveform, calibrate a noise floor per room, and apply hysteresis with two thresholds to distinguish thinking pauses from finished answers.

• •STT event timings are unreliable for silence detection due to network latency, buffering, and interim-result churn
• •Measure silence acoustically via WebAudio RMS on the raw microphone waveform instead of inferring from transcription events
• •Calibrate a per-room noise floor and use hysteresis with dual thresholds to separate thinking pauses from done-answering signals

Generated with AI, which can make mistakes.