Bosnian and Balkan Speech AI — otvoreni govorni AI, prvo ASR. Status: Faza 0 — istraživanje. Bez treninga, bez redistribucije tuđeg audija.
- Kolekcija: https://huggingface.co/collections/BH-Language-AI/bh-speech
- Kontakt: xnetba@gmail.com
docs/PROJECT_CHARTER.md— misija, pravila (test split se ne dira; licenca prije treninga)docs/DATASET_AUDIT.md— tehnički audit (splitovi, audio, duplikati, leakage=0)docs/MODELS.md— Whisper-large-v3 (Apache-2.0) vs BuzzASR (MIT) sažetakdocs/PLAN.md— faze A–E, eval metrike, compute procjena
- Obje baze pravno čiste (Apache-2.0 / MIT); BuzzASR ~1.1x bolji CER na FLEURS-u.
- Shunya dataset = prepakovani FLEURS (CC-BY-4.0) — koristiti original uz atribuciju.
- Test (925) rezerviran; splitovi bez leakagea; ~14.5h govora.
- AgentMujo Training nema Whisper podršku — treba speech adapter.