Skip to content

ADR-v2-041 — Pronunciation Feedback (L2 practice mode)

Status: Accepted (2026-07-02) · Wave F Context links: [[adr-v2-021-atypical-speech-adapter]] (distinct: adapts model vs scores you), [[adr-011]]

Context

The Wave F research (#8) proposes a practice mode: dictate a target phrase and get per-phoneme goodness-of-pronunciation (GOP) scoring for accent/second-language training. Anchors: wav2vec2 GOP/MDD — "Enhancing GOP with Phonological Knowledge" (arXiv 2506.02080), "Segmentation-free GOP" (arXiv 2507.16838). Distinct from Atypical-Speech LoRA (ADR-v2-021), which adapts the model to you; this scores you against a target — opposite direction, learner audience.

Decision

Add an opt-in Pronunciation Feedback: [pronunciation] enabled=false, good_threshold, poor_threshold. The pure core scores and formats: classify(gop, ...) → good|fair|poor, overall_score(scores) (mean), and problem_phonemes(scores, config) (phonemes below the poor threshold → practice list). The wav2vec2 GOP model that produces per-phoneme scores is lazy behind a pronunciation extra. OFF by default; practice audio processed in RAM.

Consequences

  • Distinct audience/direction from Atypical LoRA — a learner scoring tool.
  • Pure classification/aggregation → fully testable with no model.
  • Privacy (ADR-011): practice audio in RAM; no persistence unless the user saves a local log.
  • Caveat: GOP is noisy on short clips → report bands (good/fair/poor), not false precision.