Utterance is not a traditional Voice Activity Detector (VAD). VADs distinguish sound from silence. Utterance understands conversational intent.

Traditional VAD:     Sound → Speaking | Silence → Not Speaking
Utterance:           Sound → Speaking | Silence → Thinking? Done? Wants to interrupt?

Pipeline

1. Audio capture — streams microphone input via the Web Audio API
2. Feature extraction — extracts MFCCs, pitch contour, energy levels, speech rate, and pause duration in real-time
3. Semantic classification — a lightweight ML model (~3–5MB, ONNX) classifies each audio segment into one of four states:
   • speaking — active speech detected
   • thinking_pause — silence, but the speaker isn't done yet
   • turn_complete — the speaker has finished their thought
   • interrupt_intent — the listener wants to take over
4. Event emission — fires events your app can react to instantly

[Mic] → [Audio Stream] → [Feature Extraction] → [Utterance Model] → [Events]
              |                    |                      |
         Client-side          Client-side            Client-side
        (Web Audio)        (Lightweight DSP)      (ONNX Runtime)

Everything runs locally. No network requests. No API keys. No per-minute costs.

Baseline Classifier

While the ML model is being trained, Utterance ships with an EnergyVAD baseline classifier. It uses RMS energy thresholds to detect speech vs. silence and relies on pause duration to infer turn completion.

The baseline is functional but cannot distinguish thinking pauses from turn completion with the same accuracy as the upcoming ML model.