IISc Releases AI Speech Model for 65 Indian Languages

Researchers at IISc's SPIRE Lab have released *SraVaani*, India's first multilingual speech recognition model trained on 65 Indian languages and dialects. Developed in collaboration with ARTPARK and Google, the model addresses a critical gap in speech AI by supporting 40+ languages that existing systems do not officially recognize.

SraVaani covers 20 scheduled languages and 45 regional languages and dialects, potentially serving around 250 million people whose languages are poorly handled by current speech technology. The model spans all regions of India, including languages like Garo, Angika, Chakma, Kokborok, Tulu, Bundeli, and Bajjika.

Built on the *Vaani dataset* containing over 31,000 hours of natural speech from 156,000 speakers across 165 districts, SraVaani achieves the lowest average word error rate among evaluated systems. On Garo, it achieves a 9.5% error rate compared to 69.4% for competing systems.

The model is freely available on Hugging Face under an MIT licence, enabling developers and researchers to build inclusive language technology for India's linguistic diversity.