Skip to content

ADR-v2-021 — Atypical-Speech Personalization + Adapter Held-Out Gate

Status: Accepted (2026-07-02) · Wave D (research tier — gate now, fine-tune deferred) Context links: [[adr-v2-009-personal-adapter]] (P1 n-gram bias), [[adr-014-tune-holdout-validation]] (held-out gate), [[adr-015-dysfluency-friendly]], [[adr-012-self-improvement-loop]], [[adr-011]]

Context

The Wave D research (#11) proposes adapting the acoustic model to dysarthric/accented speech via a corpus-trained LoRA — extending the Personal Adapter (ADR-v2-009, which only biases the prompt) and Dysfluency Mode (ADR-015). Anchors: Universal Personalizer (arXiv 2509.15516), dysarthric LoRA 13.9% WER on Euphonia (arXiv 2505.12991). The fine-tune itself is compute-heavy and out of scope to run in-loop; but the decision to apply a trained adapter is safety-critical — a bad adapter must never silently degrade recognition.

Decision

Adopt the ADR-014 held-out principle for adapters: a pure gate should_apply_adapter(baseline_wer, adapter_wer, min_improvement) applies a LoRA adapter only when it beats the un-adapted baseline on a held-out slice by at least min_improvement (relative WER reduction). Config extends [personalize] with lora_min_improvement (default 0.03). The training + adapter load stay lazy/deferred (gated on personalize.lora, lora_min_events, plugged-in/idle); this ADR ships the pure gate + config now so the heavy path can never regress accuracy when it lands.

Consequences

  • Pure, testable safety gate → an adapter that doesn't measurably help is never applied.
  • Reuses ADR-014's train/test-overlap discipline; on-device only (ADR-011/012).
  • No new feature slug — this hardens the existing Personal Adapter path.
  • Caveat: WER must be measured on a held-out slice (not training data) for the gate to mean anything — the caller owns that split, exactly as yazses tune does.