IIT Madras's new speech recognition model supports 26 Indian languages
What's the story
BodhanAI, a non-profit initiative from the Indian Institute of Technology (IIT) Madras, has launched Indic-Transcribe, a speech recognition model. The technology was developed in collaboration with AI4Bharat and supports 26 Indian languages as well as English. Professor Mitesh Khapra announced the launch on X, emphasizing India's linguistic diversity and the need for such a tool.
Model specifications
Core and Flex are 2 versions of Indic-Transcribe
Indic-Transcribe is specifically designed to cater to India's diverse languages, scripts, and accents.
The model has been trained on 1.2 billion parameters.
It comes in two versions - Core and Flex. While the Core version is optimized for maximum accuracy, the Flex variant offers broader coverage capabilities such as mixed scripts and Romanized transcriptions.
Performance metrics
Indic-Transcribe Core outperforms Sarvam Saaras V3
Indic-Transcribe Core has a word error rate of 8.9, which is better than Sarvam Saaras V3's 10.1.
Meanwhile, Indic-Transcribe Flex scored 11.1 against Google Gemini 3 Pro's 18.9 in benchmark tests shared by BodhanAI.
The model is capable of transcribing different accents accurately and can even handle mixed scripts like cricket commentary or railway announcements with words from two languages.
Expansion goals
BodhanAI piloted the model in 12 languages
The system also supports remote languages like Bodo from Assam and Santali from Jharkhand.
BodhanAI has already piloted the model in 12 languages with 12,000 hours of speech data collected from students in grades one to 10.
The organization has obtained parental consent for this pilot program.
Indic-Transcribe will be available on the BodhanAI website starting September 5.