Skip to content

ADR-v2-054 — Sign-Language Input (SLR)

Status: Accepted (2026-07-02) · Wave G Context links: [[adr-v2-043-gesture-chords]] (discrete commands vs linguistic signing), [[adr-v2-012-accessibility-continuum]], Modality Router, [[adr-011]]

Context

Wave G research (#10) — Deaf and hard-of-hearing signers dictate by signing to the webcam; ASL is recognized on-device and injected as text. The most inclusive reading of the Accessibility Continuum — extends "voice dictation" to a non-vocal population entirely. Anchor: Google SignGemma (announced I/O 2025, Gemma/Gemini-Nano family, on-device, ASL→English, signing video never leaves the device).

Distinct from Gesture Chords (discrete command shortcuts) — this is a new linguistic modality. It slots into the Modality Router alongside voice and lip-reading.

Decision

Add an opt-in Sign-Language Input: [sign] enabled=false, pause_frames=8. Two pure cores: hands_present(num_hands) and SignSegmenter — a state machine that detects sign-burst start/end from per-frame hand-motion magnitude (motion onset → start; sustained stillness or hands leaving → end), deciding when a sign burst begins/ends. The SignGemma/Uni-Sign model + MediaPipe Hands are lazy behind a sign extra (the biggest model lift of Wave G). OFF by default.

Consequences

  • Extends dictation to a non-vocal population; a new linguistic modality in the router.
  • Pure presence + segment-boundary logic → fully testable with no camera; gates the heavy model.
  • Privacy (ADR-011): SignGemma is on-device; frames stay local.
  • Caveat: SLR is hard and model-heavy → boundary gate ships now, recognition deferred; off by default.