SraVaani: IISc Releases Speech AI Model for 65 Indian Languages
Why in the news
The Indian Institute of Science (IISc), Bengaluru released SraVaani on August 13, 2026. It is a speech recognition model meant for Indian languages that have little digital data.
Key facts
- Developers: SPIRE Lab at IISc, working with ARTPARK, with support from Google.
- Coverage: 65 languages and dialects, including 20 scheduled languages and 45 regional ones such as Garo, Angika, Chakma, Kokborok, Tulu, Bundeli and Bajjika.
- Scripts: 10, with automatic language identification so users need not pick a language.
- Accuracy: 9.5% word error rate on Garo versus 69.4% for the next best alternative.
- Licence: MIT, hosted on Hugging Face, so it can be used and modified freely.
Training data
| Item | Figure |
|---|---|
| Dataset | Project Vaani |
| Speech hours | 31,000+ |
| Speakers | 156,000 |
| Districts | 165 (28 states) |
About speech recognition
Automatic speech recognition turns spoken words into written text. A lower word error rate means better accuracy, so the Garo result is the main performance highlight.
Exam angle
- Link the model to IISc Bengaluru and its SPIRE Lab.
- Numbers to recall: 65 languages, 10 scripts, 20 + 45 split.
- Open-source release on Hugging Face supports digital inclusion for low-resource languages.