Skip to content

ADR-015 — Dysfluency-Friendly Mode (collapse pass + accessibility preset)

Status: Accepted (2026-06-19) Context links: [[adr-011]] (zero telemetry / offline), [[adr-012-self-improvement-loop]] (the corpus that could later evaluate this), [[adr-013-llm-cleanup]] (guarded text-transform philosophy)

Context

YazSes names accessibility-OS as one of its four product axes, but offered no specific accommodation for atypical speech. A vision-lab review (2026-06-19) developed Dysfluency-Friendly Mode into a Vision Card (internal) that scored GO (22/25, regret ⅘, reputation/learning 5/5) — the top "ship-now" pick.

Its load-bearing evidence — Lea/Wu et al., "Enabling People Who Stutter to Better Use Speech Recognition," CHI 2023 (doi:10.1145/3544548.3581224) — found that most of the ASR gain for people who stutter comes from endpointing + posthoc text refinement (removing repetitions), taking a consumer system from WER 25.4% → 9.9% and cutting truncation 79.1%, without acoustic retraining. The population is large (~1% stutter; most Parkinson's patients develop dysarthria) and already uses ASR despite it failing them (32.2% daily in the same survey).

YazSes already owns a guarded 3-pass disfluency filter (stt/filters/disfluency.py) with an _is_protected guard for proper nouns / code / URLs. The one genuine build gap the card identified (its two critical gaps) was: a sub-word repetition + prolongation collapse pass the existing whole-word 2-gram dedup misses, and an evaluation that proves it does not damage clean text (we have no local affected-user audio).

Decision

Add an opt-in collapse pass (Rule B.5) to the disfluency filter, plus a one-switch accessibility preset, all off by default so the default pipeline is byte-identical.

  1. Collapse pass (stt/filters/disfluency.py), inserted after 2-gram dedup and before self-correction rollback, every operation skipping _is_protected tokens:
  2. _collapse_prolongations(text, min_run) — a same-letter run of length ≥ min_run (default 3) collapses to one letter (sooo→so); English double letters (run 2) stay safe.
  3. _collapse_repetitions(text, max_fragment_len) — three conservative forms: (a) hyphenated false starts with ≥2 identical leading fragments that prefix the final word (b-b-because→because); (b) ≥2 identical short (≤ max_fragment_len, default 2) space-separated fragments followed by a longer word they prefix (b b because→because); © unigram runs of length ≥3 (the the the→the).

  4. Driven by four DisfluencyConfig flags (collapse_repetitions, collapse_prolongations, prolongation_min_run, repetition_max_fragment_len), all default off.

  5. Master preset — AccessibilityConfig.dysfluency_friendly (default off). When set, config._apply_presets enables both collapse flags and widens pre_speech_padding_ms to ≥ 400 ms (delayed voice onset is common in dysarthria).

  6. Pre-registered evaluation gate (tests/test_dysfluency_eval.py over tests/fixtures/disfluency/dysfluency_eval.json) encodes the card's LOFA-1 kill criteria as an automated regression test: false-collapse < 2% on clean control AND recall ≥ 60% on labelled dysfluency spans. Measured at ship: 0% false-collapse, 92.9% recall.

  7. yazses doctor reports the mode's status when enabled.

Scope decision — endpointing is out (and why)

The CHI-2023 result also credits forgiving endpointing. This does not transfer to YazSes's core path: YazSes is hold-to-talk — the user holds the key through a block or pause and releases when truly done, so utterance end is a key release, not an auto-endpointer decision. (AccessibilityConfig.min_silence_ms is, in fact, currently stored but not consumed by the daemon.) Truncation-by-endpointer would only matter on the off-by-default streaming / Ghost-Ahead speculative-finalize path. So the preset's only endpointing-adjacent action is widening onset padding; it makes no claim to fix endpointer truncation. The endpointing question is parked (see the card / plan Open Questions), not silently dropped.

Alternatives rejected

  • Acoustic model fine-tuning / LoRA for atypical speech. The bigger lever for severe dysarthria, but needs training data + adapters and is the v2.0 roadmap item, not this. This pass stacks under any future adaptation. Out of scope by design.
  • Aggressive collapse (collapse any repeat / any double letter). Rejected — it fails the < 2% clean-control bar (would eat re-read, co-op, very very, committee). Conservatism (≥2 leading fragments, ≥3 unigram run, ≥3 letter run, _is_protected) is deliberate; the eval gate enforces it.
  • On-by-default. Rejected — honours the "default pipeline unchanged" rule and ADR-011.

Consequences

  • A person who stutters / has dysarthria flips [accessibility] dysfluency_friendly = true and gets clean text today — offline, on CPU, no model, no new dependency.
  • Two documented conservative misses (b-because single fragment; I I protected by capitalisation) — accepted trade-offs; loosening them would breach the clean-control bar.
  • Prolongation collapse-to-1 can mis-handle rare ≥3 double-letter words (welllll→wel); bounded by the clean-control gate, revisit only if it shows real damage.
  • The eval fixture is a reusable harness; if UA-Speech/TORGO/SEP-28k become available it can be strengthened. All in-house WER claims stay [unverified] until measured on real affected-user audio (real-user validation is the card's remaining value LOFA).