Releases and lab updates.
Model launches, stack announcements, and measured progress from Navdyut AI Labs. Essays stay elsewhere; this is the institutional record.
Bani: Assamese voice, end to end
End-to-end Assamese voice stack for citizen-facing systems, measured on turn latency and dialect hold.
Read releaseAll releases.
Each note records a release with evaluation data. Product depth lives on Models; the posts below are the public changelog.
Muga 30B: flagship Assamese language model
30B-parameter Assamese LM with gains on formal register, citizen writing, and document-grounded Q&A.
Read update
Bani TTS: text-to-speech for Assamese phonology
Native-phonology TTS with preference wins on clarity and dialect-appropriate prosody.
Read update
Bani STT: speech recognition across Assamese dialects
Dialect-aware ASR with word-error rates reported by region on a Navdyut speech eval set.
Read update
Lipi Hand: reading Assamese handwriting
Handwriting OCR with character error rates on land records, letters, and archival manuscripts.
Read update
Lipi Print: Assamese print extraction
Print OCR tuned for Assamese script, with CER by document class.
Read update
Eri 240M: open Assamese foundation model
Open Assamese foundation model: architecture proof at 240M with a compressed vocabulary and curated training diet.
Read update
A 32K Assamese Tokenizer
A script-aware 32K vocabulary that keeps Assamese in whole pieces instead of byte fragments.
Read update