Skip to content

ADR-v2-119 — Echo (own-audio replay for a text span)

Status: Accepted (2026-07-03) · Wave N Context links: proofback (read-back proofreading), scrub (audio-anchored scrubbing), confidence, [[adr-011]]

Context

Wave N research (#5) — proofreading dictation eyes-free is hard: TTS read-back of the transcript can't reveal an ASR homophone error (it just re-speaks the wrong word correctly). Replaying the user's own captured audio for a span exposes the mismatch by ear. The set has TTS read-back (proofback) and audio-anchored scrubbing (scrub) but no "play my own recording of this text back" verb. Anchor: Vertanen eyes-free ASR-error detection (arXiv 2410.20564).

Decision

Add an opt-in Echo: [echo] enabled=false. Pure cores in echo/span.py: build_span_index(words) → Span(text, char_start, char_end, t_start, t_end) list from word timings (char offsets accumulate with one separating space, matching injected text); resolve_playback_target(utterance, spans) → an audio (t_start, t_end) window for "play that back"/"play all" (whole clip), "last word"/"that word", "first word", or a quoted/bare token match (matching words merged), defaulting to the whole clip. The captured audio buffer and its playback are the only I/O and reuse the existing recorder. OFF by default.

Consequences

  • Eyes-free homophone/ASR-error detection via the user's own voice, not TTS.
  • Pure index + resolver → fully testable over synthetic word timings.
  • Distinct from Proofback (TTS of transcript) and Scrub (navigation) — this replays source audio.
  • Privacy (ADR-011): audio stays in the local capture buffer, never persisted by Echo.
  • Caveat: needs word timings (faster-whisper provides them) and the retained audio buffer; off by default.