AI4Bharat, Bodhan AI Build 1.2 Bn Parameter Speech Model for 26 Indian Languages

AI4Bharat and Bodhan AI have developed Indic-Transcribe, a 1.2-billion-parameter speech recognition model trained on 1.3 million hours of audio to transcribe speech across 26 Indian languages and English.

The organisations said the model is built to handle code-mixed speech, accents and children’s speech, with its first use case focused on education. Indic-Transcribe is scheduled to go live on Bodhan AI on September 5.

“India is a voice-first nation with several languages, diverse scripts, and countless accents that no single ‘Indian English’ or ‘standard Hindi’ ever quite captures,” AI4Bharat said in a LinkedIn post announcing the model.

The model supports applications where spoken Indian languages need to be converted into text, including education, healthcare, finance and governance. AI4Bharat said its first priority is to allow teachers and students to interact with technology “in their own voice, their own accent, their own language”.

Indic-Transcribe also includes code-mixed speech recognition, reflecting conversations where speakers switch between Indian languages and English. 

The teams said the model was trained to handle children’s speech, with data collection involving parental consent, child assent and clearance procedures.

The launch builds on AI4Bharat’s earlier work in speech recognition. Its IndicConformer model supports all 22 constitutionally recognised Indian languages and is available as an open-source ASR model. The model uses a Conformer architecture and supports both CTC and RNN-T decoding.

AI4Bharat has also previously worked on IndicWav2Vec, a multilingual speech model pretrained on 40 Indian languages, and has released speech datasets and models as part of its broader effort to build language technology for India.

Bodhan AI, based at IIT Madras and founded in 2026, focuses on AI for education and is building multilingual technologies across text, speech and vision. Its research programme is part of what it describes as the Bharat EduAI Stack.

The Indic-Transcribe launch also comes as speech recognition becomes a larger focus of India’s AI ecosystem, with developers building models that can handle regional languages, accents and code-switching rather than relying on English-centric speech systems.