Models
We release open models with the evidence needed to test every claim: weights, benchmarks, limitations, and reproducible results. No black boxes.
Speech · ASR
Kriti V1
An open 119M-parameter Nepali speech recognition model. Shares rank #1 of 19 evaluated systems with 24.08% WER, and records the lowest raw word error rate (24.69%) across the field. Trained on 630 hours of Nepali audio.
Nepali speech recognition
Nepali, heard clearly.
Kriti is a 119M-parameter Nepali-only RNN-Transducer. It shares rank one on the development benchmark and records the lowest raw word error rate in the evaluated field of 19 systems.
- Live parameters
- 119M
- Primary WER
- 24.08%
- Benchmark rank
- #1
- Training audio
- 630 hrs
Speech · Telephony ASR · New
Kriti Telephony
Kriti adapted for real phone audio: call-centers, IVR, and code-switched spontaneous Nepali. Achieves NepTel WER 32.38 — new state of the art on Nepali call-center speech, beating the previous best by 1.4 points.
Language model · Research
BarunLM 35M
A compact 35M-parameter decoder-only base model, built to study how architecture and data quality interact under hard compute limits.