Skip to content

ADR-v2-027 — Phonetic Corrector (fix mis-heard names by sound)

Status: Accepted (2026-07-02) · Wave E Context links: [[adr-v2-004-context-primed]] (biases prior), [[adr-v2-009-personal-adapter]] (biases prior), [[adr-011]]

Context

The Wave E research (#2) notes that prompt-biasing (Context-Primed, Personal Adapter) shapes the decoder's prior but can't recover a proper noun the model already got wrong ("Cuber Netties" for "Kubernetes"). A post-hoc corrector that matches transcript tokens against the personal lexicon in sound space fixes these. Anchors: G2P (espeak-ng/phonemizer), phonetic edit distance; the post-hoc ASR-correction pattern of MathSpeech (arXiv 2412.15655).

Decision

Add an opt-in Phonetic Corrector operating on the output: for each transcript token, if it isn't already a known word, compare its phonetic key to each personal-vocabulary term's key and, when the normalized key-distance is within max_distance, replace it. The pure core ships a compact, documented phonetic key (a deterministic consonant-skeleton reduction — not a claim to be Metaphone) + normalized Levenshtein; a stronger G2P/neural phonetic backend is opt-in behind a phonetic extra. [filters.phonetic] enabled=false, max_distance. OFF by default.

Consequences

  • Corrects the output (a different stage than prior-biasing) → recovers already-wrong words.
  • Pure key + distance → fully testable with no model; the vocab is the user's local list.
  • On-device only (ADR-011).
  • Caveat: over-eager correction could rewrite a correct rare word → conservative distance floor, only correct tokens not already in a known-word set, off by default.