IIT Madras-Backed Bodhan AI, AI4Bharat Launch 4 AI Models for Students & Teachers

Bodhan AI, a centre of excellence in AI for education incubated at IIT Madras, has launched four AI models for speech recognition, speech generation, machine translation and optical character recognition (OCR), in partnership with AI4Bharat.

The models are being released as Digital Public Goods through open-weight models and hosted APIs on sovereign infrastructure. The initiative is part of the Bharat EduAI Stack, which aims to provide a common AI infrastructure layer for multilingual education in India.

The four models cover 27 languages for automatic speech recognition (ASR), 23 languages for OCR, 22 languages for machine translation through Bodhan-Translate, and 23 languages for text-to-speech (TTS).

The models are trained and optimised using NVIDIA Nemotron open models and libraries, including the NVIDIA NeMo framework for ASR, machine translation and OCR. Bodhan AI has post-trained NVIDIA Nemotron 3.5 ASR for Indian languages, including regional dialects and accents. The models are served using NVIDIA TensorRT-LLM and vLLM inference microservices.

Prof Mitesh Khapra, Principal Investigator at Bodhan AI, said the project was intended to create a shared infrastructure layer rather than separate AI systems for individual institutions.

“We built these voice and vision models, and made them accessible as Digital Public Goods on a Digital Public Infrastructure, so that efforts across the country don't remain fragmented,” Khapra said.

He added that edtech companies, startups, researchers, universities, technology companies and government partners could use the models to develop applications for Indian users.

The speech recognition model can allow students to interact with AI tutors through voice in their preferred language, while TTS can enable AI systems to respond in Indian languages. OCR can help education systems process textbooks, worksheets, documents and handwritten answers. Bodhan-Translate can be used to translate educational content between Indian languages.

Prof V Kamakoti, Director of IIT Madras, said the infrastructure could help expand access to AI across India’s linguistic diversity.

“India’s AI journey cannot be built on technology alone. It must be built on technology that understands India,” Kamakoti said.

The models will also power Bodhan AI’s own education applications, including Student TutorBot and Teacher Assistant Bot.

Student TutorBot is designed for students in Classes 6–12 and is based on NCERT and SCERT curricula. Students can interact with it through text or voice across 22 Indian languages. The system uses textbook content to provide explanations, examples and assessments and can guide students through problems using interactive tools and a digital canvas.

Teacher Assistant Bot is designed to help teachers plan, create, assign and assess classroom work. Teachers can generate lesson plans, worksheets, quizzes, homework and revision material by specifying parameters such as grade, subject, topic, duration and difficulty. They can also upload student work for evaluation against specified marking criteria.

Bodhan AI said AI-generated outputs in the Teacher Assistant Bot are intended to remain subject to teacher review. Teachers can edit, regenerate or discard outputs rather than having AI make classroom decisions automatically.

Prof Balaraman Ravindran, Head of the Wadhwani School of Data Science and AI at IIT Madras, said developing AI for India requires more than adapting existing systems to Indian languages.

“Building AI for India requires going beyond adapting existing models to Indian languages. It requires developing foundational capabilities that understand the richness and diversity of our languages, contexts and educational needs,” Ravindran added.

Bodhan AI said government and public institutions could use the infrastructure for multilingual education and other public-service applications, while edtech companies could integrate the models through APIs without developing foundational models themselves. Startups and researchers can access the open-weight models for adaptation and further development.

The hosted API infrastructure is being designed around a sovereign deployment approach, with data anonymisation protocols and compliance with applicable national education data frameworks. The company said this is intended to give institutions and ecosystem partners greater control over deployment while supporting requirements around data governance, privacy and security.

NVIDIA and Bodhan AI are also collaborating on datasets, training recipes and evaluations for future foundational models for Indian languages. AI4Bharat is contributing its work on Indian-language datasets, tools, models and applications to the initiative.