Open models for Indian languages arrive, and Urdu handwriting is in scope

Bodhan AI, an IIT Madras-incubated centre, has released a suite of open foundational models for Indian languages, built with NVIDIA and AI4Bharat. Four capabilities ship together: speech recognition (Indic-Transcribe), text-to-speech (Indic-Speak), machine translation (Indic-Translate) and optical character recognition (Indic-OCR). Translation covers 22 languages, speech synthesis 23, and transcription 26 Indian languages plus English. The OCR reads both printed and handwritten text. Urdu is among the languages covered, including handwriting, which is the part that should interest anyone who has tried to get a machine to read a nastaliq manuscript, a madrasa register or a family archive. The models were developed on NVIDIA's NeMo framework with post-training on Nemotron 3.5 ASR to handle regional dialects and accents, and are released as open weights with hosted APIs on sovereign digital infrastructure. Bodhan frames the suite as the Bharat EduAI Stack, a shared digital public infrastructure for a multilingual school system, with educational applications built on it to remain free for learners, teachers and participating state governments. The institutional lesson is about who holds the weights. A language whose recognition models are open can be taught, archived and searched by anyone who cares to; a language that depends on a vendor's roadmap waits its turn.

This is a QeRN summary by Ahmed Qerni. Read the original at Free Press Journal: https://www.freepressjournal.in/education/bodhan-ai-iit-madras-nvidia-launch-open-ai-models-for-indic-languages-covering-speech-translation-text-to-speech-ocr.