I trained and released speech recognition models for Kazakh, Russian, and mixed speech. Kazakh is still underrepresented in open STT, and real conversations switch languages. The model card includes results and limitations; the CPU inference example runs short recordings.
- JevK5 — an open-weight Jev alternative for typed decisions; #2 of 76 on JevBench v1.4.
- pytest-jev — tests whether an LLM response makes the claims you expect.
- jevgrep — filters live logs by meaning using plain-English questions.
If you try the speech model on real Kazakh or mixed-language audio, I'd like to hear where it fails. Report an STT example or reach me on LinkedIn.
