IISc, ARTPARK and Google launch SraVaani for 65 Indian languages
IISc, along with ARTPARK and Google, has rolled out SraVaani, a new open-source speech recognition model that understands 65 Indian languages.
What's cool is that it covers 44 low-resource languages for which the researchers say no competing automatic speech recognition (ASR) system offers official support, helping more people use voice tech in their own mother tongue.
SraVaani trained on over 31,000 hours
SraVaani was trained using over 31,000 hours of speech from across 165 districts, capturing different accents and dialects, even for lesser-known languages like Nyishi.
The model performs well overall (with a mean word error rate of 28.4% for major Indic languages) and really shines in some smaller ones like Garo.
It's open-source under the MIT License, so developers can tweak it for their needs, making voice AI more inclusive and affordable across India.