Skip to content

ADR-v2-102 — Involuntary-Vocalization Auto-Excision

Status: Accepted (2026-07-02) · Wave L Context links: disfluency (linguistic fillers), audioguard (environmental events), hallucination (STT artifacts), [[adr-011]]

Context

Wave L research (#8) — automatically detect and delete involuntary self-produced sounds (cough, throat-clear, sneeze, sniff) from the dictation stream so they never turn into garbage tokens or trigger spurious commands. The Disfluency filter removes linguistic fillers ("um"); the Ambient Audio-Event Guard handles environmental privacy events; the Hallucination Guard filters STT artifacts. None excise the user's own non-linguistic bodily sounds — a real problem for chronic-cough and medical dictation. Anchor: Hyfe Cough Tracker (F1000Research 11:730; validation PMC11809693) — sub-0.5 s cough sounds classified on-device at 91% sensitivity / 98% specificity.

Decision

Add an opt-in Involuntary-Vocalization Auto-Excision: [involuntary] enabled=false. Pure cores in involuntary/excise.py: is_involuntary_vocalization(seg_feats) → cough | throat_clear | sneeze | None from duration/energy/centroid/voicing/burst features, and excise_nonspeech_spans(tokens, token_spans, flagged_spans) → the token list with any token overlapping a flagged span removed. The span-removal is pure; the acoustic classifier is the lazy extra. OFF by default.

Consequences

  • Cough/throat-clear/sneeze never corrupt the transcript — a win for chronic-cough and medical dictation.
  • Pure overlap/removal logic → fully testable.
  • Distinct from Disfluency (bodily vs linguistic), Audio-Guard (self vs environmental).
  • Privacy (ADR-011): local audio only; excised, never stored.
  • Caveat: coarse feature thresholds (a learned classifier is a later tier); off by default.