Skip to content

ADR-v2-120 — SRPace (screen-reader-paced injection)

Status: Accepted (2026-07-03) · Wave N Context links: srpace↔brailleout (accessibility output), continuum, injection, [[adr-011]]

Context

Wave N research (#6) — injecting a whole dictated paragraph as one burst floods a screen reader: it either announces a wall of text at once or re-reads on every keystroke, so a blind user loses the thread. Pacing injection to the reader's own words-per-minute rate, chunked at clause boundaries, lets the screen reader announce each piece coherently. Nothing in the set controls injection timing for assistive output. Anchor: screen-reader comprehensibility research.

Decision

Add an opt-in SRPace: [srpace] enabled=false, wpm=180.0. Pure cores in srpace/schedule.py: chunk_clauses(text) splits at sentence/clause punctuation (delimiter kept, blanks dropped); plan_injection_schedule(text, wpm, min_chunk_ms) → Chunk(text, delay_ms) list where each chunk's duration is words / wpm (floored at min_chunk_ms) and delay_ms is the cumulative start time, so the injector announces chunks in sequence without overrunning the reader. wpm ≤ 0 disables pacing (single zero-delay chunk). OFF by default.

Consequences

  • Blind/screen-reader users get coherent, keep-up-able announcement of dictated text.
  • Pure chunker + scheduler → fully testable, deterministic, dependency-free.
  • Distinct from BrailleOut (output encoding) — this is output timing; the two compose.
  • Privacy (ADR-011): local text only.
  • Caveat: default wpm is a heuristic (users match it to their reader); the injector must honour the per-chunk delays; off by default.