ADR-v2-029 — Semantic Auto-Stop (hands-free tap-and-speak)¶

Status: Accepted (2026-07-02) · Wave E Context links: [[adr-v2-012-accessibility-continuum]], [[adr-v2-000-interaction-layer]], [[adr-011]]

Context¶

Every YazSes activation path assumes the hotkey defines the utterance boundary (hold to talk). The Wave E research (#4) proposes an alternative for users who cannot hold a key: tap to start, and the daemon auto-stops when you've finished the sentence. Anchor: Smart Turn v2 (Pipecat, HF open weights, wav2vec2, _60MB/400ms/14 langs) — semantic end-of-turn from prosody, not transcript. This is a new activation axis (motor impairment) untouched by the 24 prior features and distinct from the Ghost-Ahead [endpoint] pre-warm (which anticipates a manual release, never ends recording).

Decision¶

Add an opt-in auto-stop: [autostop] enabled=false, mode (silence|semantic), silence_timeout_ms, max_duration_ms. The pure core should_stop(silence_ms, elapsed_ms, config) ends a hands-free turn when trailing silence exceeds silence_timeout_ms or the turn hits max_duration_ms (a hard safety cap). This ships now with zero ML and already makes tap-and-speak usable; the Smart Turn v2 ONNX model (true semantic end-of-utterance) is opt-in behind a turn extra and refines the silence decision when present. OFF by default.

Consequences¶

• New activation modality for users who can't hold a key; complements Wake-Word (#8).
• Pure timeout/cap decision → fully testable; the semantic model stays deferred.
• On-device only (ADR-011); audio never leaves RAM (same as existing VAD).
• Caveat: a too-short silence timeout clips slow speakers → conservative default + hard max-duration cap; off by default.