Why it matters
Nepali speech, finally open and verifiable.
Kriti shares the primary score with the official AI4Bharat Nepali RNNT and records the lower raw word error rate across a field of 19 evaluated systems, at only 119M live parameters.
Open model
Kriti V1
A 119M-parameter Nepali RNN-Transducer for open speech recognition, released with full evidence, artifact hashes, and a reproducible benchmark.
Open on Hugging FaceKriti V1
Open weights, MITNepali 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
Design
A Conformer encoder and a tiny danda head.
Seventeen Conformer blocks feed a Nepali RNN-Transducer, while a 1,025-parameter acoustic head decides when a terminal Devanagari danda belongs in the text. Nepali-only runtime pruning keeps the live graph small.
Release standard
Evidence travels with the artifact.
Weights, model card, architecture and training notes, the frozen 3,630-utterance benchmark view, and exact replicate hashes ship together. Released under MIT with upstream AI4Bharat attribution.