Downstep
Japanese pitch accent for compounds your TTS dictionary has never seen. Both versions below are synthesised with the same voice, so the only thing you hear changing is where the pitch falls.
| system | compound in the dictionary | compound not in the dictionary |
|---|---|---|
| OpenJTalk | 96.4% | 40.6% |
| Downstep | 81.3% | 81.3% |
Downstep is worse on words the dictionary already knows, and that is the point: it is a fallback path, not a replacement for lookup. It reports a calibrated confidence and abstains below a threshold, returning OpenJTalk's answer unchanged. Accent-nucleus exact match on 5,414 held-out compounds; gold from NAIST-jdic and UniDic 2.1.2.
Listen
None of these compounds is a dictionary entry, so OpenJTalk has to compose the accent from the parts. Notice how often it also breaks the compound into several accent phrases.
新型感染症対策本部 a new-infectious-disease response headquarters
地球温暖化対策 global-warming countermeasures
少子高齢化社会 an ageing society with a falling birthrate
再生可能エネルギー renewable energy
デジタル人材育成 digital-talent development
気候変動適応計画 a climate-change adaptation plan
選挙管理委員会 an election administration committee
This is a static page: the examples are pre-rendered, so you
cannot type your own word here. The interactive version lives in app.py in the
repository and runs locally with python app.py.
No human annotation exists in this release — every accent label is derived from a dictionary. The repository ships the annotation protocol and an extractor that emits novel compounds unlabeled, but no native speaker has checked any of it.
Model · Benchmark · Code and evaluation · Apache-2.0