33 lines (25 loc) · 1.63 KB

Declarative Phone Normalization Layer

Goal

Replace hardcoded phoneme-map.ts with a declarative normalize rule phase (phases/normalize.yaml) that runs first in the pipeline, with citations on every mapping.

What Changed

Created

public/rules/phases/normalize.yaml — Three splice rules: AX->AH, NX->NG, WH->W

Modified

public/rules/pipeline.yaml — Added normalize phase before postlexical
public/rules/frontend.yaml — Added normalize.yaml include
src/g2p/syllabify.ts — Added AX to VOWELS set (so Hunnicutt stress counts schwas)
src/g2p/lts-engine.ts — Removed mapElovitzToQlatt import/call, inlined prosodic marker filter
test/g2p-lts-engine.test.ts — Removed mapElovitzToQlatt tests, updated expectations (AX no longer mapped to AH at LTS level)

Deleted

src/g2p/phoneme-map.ts — Responsibilities reassigned to normalize.yaml + inline filter

CEL issue

string(int(current.stress)) failed — cel-js doesn't have int() builtin
Used ternary chain instead: current.stress == 1 ? "AH1" : (current.stress == 2 ? "AH2" : "AH0")

Behavioral changes

pronounce() now returns raw Elovitz symbols (AX, NX, WH) instead of mapped (AH, NG, W)
Full TTS pipeline output unchanged — normalize phase maps them before downstream rules fire
G2P benchmark accuracy: "the" now shows as mismatch at pronounce() level (AX vs AH), but pipeline produces correct AH

Test results

All G2P tests: 178/178 pass (6/6 files)
Full suite: 10 files failed, 17 tests failed — ALL pre-existing (voice quality snapshots, schema validation)
No regressions from these changes