Skip to content

ADR-v2-023 — Silent-Speech / Subvocal (sEMG) Input

Status: Accepted — design only, implementation deferred (2026-07-02) · Wave D (research tier) Context links: [[adr-v2-000-interaction-layer]], [[adr-004-emg-backend]] (existing EMGBackend/YESP), [[adr-011]]

Context

The Wave D research (#9) proposes dictating by mouthing words silently — surface EMG (sEMG) on the face/neck decodes articulation without audible speech, for privacy, noisy environments, and users who cannot phonate. Anchors: microneedle SSI reaching 8.5% WER (ScienceDirect S2666053925001249), emg2speech (arXiv 2510.23969). YazSes already has an EMGBackend (ADR-004) that reads squeeze events over USB serial (YESP) — a natural hardware seam to extend from single-gesture triggering to continuous articulation decode.

Decision

Design accepted; implementation deferred until hardware is in hand. The intended shape, consistent with the v2 pattern: a [silentspeech] config (enabled=false, port, baud_rate, model_path), an experimental feature (refused by features enable without --force), and a decode model lazy behind a silentspeech extra. The pure seam that can ship first (a future iteration) is the articulation→text post-processing (the same clean_text/disfluency chain the audio path uses), reusing the existing EMG serial reader. The neural sEMG→text decoder needs a real electrode array + labelled data to build and verify, so no code lands now — only this ADR and the research record.

Consequences

  • Extends the existing EMGBackend rather than adding a parallel hardware stack.
  • On-device only (ADR-011); no audio leaves the machine (there is none).
  • Honest scoping: accuracy + hardware are the bottleneck → experimental, --force, deferred.
  • Caveat: silent articulation lacks voicing cues → expect higher WER than audible speech; the held-out validation gate (ADR-014) applies to any personalization here.