Bodhan AI, IIT Madras & NVIDIA Launch Open AI Models For Indic Languages Covering Speech, Translation, Text-to-Speech & OCR

Chennai: Bodhan AI, a Centre of Excellence in Artificial Intelligence for Education incubated at IIT Madras, unveiled a collection of open-source foundational AI models for Indic languages in partnership with NVIDIA on Friday. The announcement was made through a press release. This initiative is a part of the AI4Bharat collaboration.

The models encompass four key functionalities: speech-to-text (Indic-Transcribe), text-to-speech (Indic-Speak), machine translation (Indic-Translate), and optical character recognition (Indic-OCR). These models have been developed utilizing the NVIDIA NeMo framework, which incorporates post-training NVIDIA Nemotron 3.5 ASR to accommodate Indian regional dialects and accents.

The inference is executed through NVIDIA TensorRT LLM and vLLM microservices. The primary objective of this initiative is to foster the widespread use of multilingual AI solutions within the broader Indian language AI ecosystem, with a special emphasis on education. Prof Mitesh Khapra, Principal Investigator at Bodhan AI and AI4Bharat, expressed that the NVIDIA partnership is instrumental in accelerating the deployment of open-state-of-the-art capabilities for developers.

These foundational models are made available through open-weight releases and hosted APIs within India's sovereign digital ecosystem. Educational applications utilizing these models will remain accessible free of cost to learners, educators, and participating state governments.