Akusento pitch-accent scorer (v2)

Logistic regression (12 F0 features/mora: phrase-median normalized + neighbor context) predicting High/Low per mora.

Mora H/L accuracy: 0.789 (majority 0.504), 2,020 held-out morae (JSUT read speech, 1 speaker). v1 (utterance normalization) was 0.736.
Held-out SPEAKERS (JVS parallel100, unseen): 0.713 over 40,243 morae, 10 speakers (M 0.698 / F 0.727, range 0.654-0.752). jvs006 skipped (empty alignments upstream).
Word accent-type exact match: 0.460 (398 phrases; harsh metric).
Targets are OpenJTalk Tokyo-dialect labels (NOT human-checked); acoustic cross-check: 743 agree / 363 contradict / 773 unclear transitions.
Training: akusento/scripts/train_scorer.py in github.com/raihan-js/akusento.
Cannot reproduce anyone's voice (statistics + tiny linear model).

raihan-js/akusento-scorer