Skip to content

ADR-v2-037 — Audio-Anchored Scrubbing

Status: Accepted (2026-07-02) · Wave F Context links: [[adr-v2-003-spoken-edit-mode]] (text edit), Punch-In (blind re-record), [[adr-011]], [[adr-012-self-improvement-loop]]

Context

The Wave F research (#3) proposes keeping word-level timestamps with each dictation so the user can "replay what I said" or pick a word to hear that exact audio slice and re-dictate just that word. faster-whisper word_timestamps=True (already used on the prosody/confidence path) and sounddevice playback (already a dep) make this dependency-free. Distinct from Spoken Edit (text editing) and Punch-In (blind re-record) — this is word→audio alignment for verification and pinpoint correction.

Decision

Add an opt-in Audio-Anchored Scrubbing: [scrub] enabled=false. The pure core indexes timed words and resolves selections: find_word(words, query) (last word matching a spoken target), slice_bounds(words, i) and range_bounds(words, i, j) (audio start/end for a word or span). Playback and re-dictation reuse existing sounddevice + the correction path. The audio slice is held in RAM for the session; it is persisted only if [learning] capture_audio is already on (encrypted, ADR-012). OFF by default.

Consequences

  • Ship-now with no new dependency — word-timestamp index + slice selection is pure.
  • Distinct from Spoken Edit / Punch-In — verification + pinpoint audio correction.
  • Privacy (ADR-011/012): audio slice in RAM only; persisted only under the existing encrypted opt-in.
  • Caveat: alignment quality is bounded by Whisper's word timestamps → best-effort, off by default.