New Indian Speech Model Covers 65 Languages, Including Underserved Dialects
Researchers have launched SraVaani, a novel multilingual Indian speech recognition model trained on 65 languages and dialects. This initiative aims to significantly expand speech AI accessibility to underserved regional languages and dialects across India.
Key takeaways
- A new multilingual Indian speech recognition model, SraVaani, has been released.
- The model is trained on 65 languages and dialects, including many underserved ones.
- It aims to make speech AI accessible to approximately 25 crore people across India.
- SraVaani demonstrates strong performance, with a 9.5% word error rate on Garo.
A new multilingual speech recognition model named SraVaani has been developed by researchers at IISc’s SPIRE Lab, in partnership with ARTPARK and Google. This model is the first of its kind in India to be trained on 65 languages and dialects.
The development aims to broaden the reach of speech AI beyond the officially scheduled languages, offering support for numerous regional and non-scheduled Indian languages that currently lack adequate speech-to-text capabilities.
What Happened
SraVaani was created to address the disparity in speech AI technology, where most systems are proficient only in a limited number of dominant languages. The model incorporates 20 scheduled languages and an additional 45 regional languages and dialects, potentially making speech AI accessible to approximately 25 crore individuals, based on the 2011 Census, whose languages are not effectively supported by existing technologies.
The model's coverage is designed to be comprehensive across India, including languages from the Northeast, eastern, western, northern, southern, and central regions, alongside English and Sanskrit. It supports languages such as Garo, Angika, Chakma, Kokborok, Tulu, Bundeli, and Bajjika, pushing the boundaries of Indian dialect speech-to-text and regional-language AI beyond traditionally prioritized languages.
Performance results indicate strong error rates for some languages. For Garo, SraVaani achieved a word error rate of 9.5%, significantly outperforming the next best system which had a rate of 69.4%.
Key Facts
- SraVaani is the first multilingual Indian speech recognition model.
- It was trained on 65 Indian languages and dialects.
- The model was developed by researchers at IISc’s SPIRE Lab, ARTPARK, and Google.
- It supports over 40 languages not officially covered by current speech recognition systems.
- SraVaani covers 20 scheduled languages and 45 regional languages and dialects.
- The model is freely available on Hugging Face under an MIT license.
- It achieved a 9.5% word error rate on Garo, compared to 69.4% for the next-best system.
- Govindan Rangarajan, Director of IISc, stated the model is a contribution towards inclusive language technology.
- The model's development was announced on August 13, 2026.
Why It Matters
The introduction of SraVaani marks a significant step towards greater inclusivity in artificial intelligence. By extending speech AI capabilities to a vast array of languages and dialects, the model has the potential to empower millions of Indian citizens whose voices have been underserved by current technology, aligning with India's ambition to build its own AI capabilities with a focus on inclusive language technology.
Sources reviewed
Project Chintan independently synthesized and analyzed information cross-checked across the sources listed above.
Related stories
Fitbit Air, Google's Screenless Tracker, to Launch in India This October
AI System 'Gajraj AI' Detects Elephants on Tracks, Alerts Train Drivers
Karnataka Launches Tathyakosh.in for Unified Data Discovery
