Researchers at IISc’s SPIRE Lab, with ARTPARK and supported by Google, have released SraVaani, the broadest-coverage multilingual Indic speech recognition model reported to date, supporting more than 60 Indian languages and dialects.
On India’s widely served languages, SraVaani delivers accuracy comparable to existing Indic speech recognition systems, recording the lowest average word error rate of those evaluated. Its distinctive strength lies in the long tail. More than 40 of the languages it covers are not officially supported by current systems, and results on several of these are particularly strong, including a 9.5% word error rate on Garo against 69.4% for the next best system evaluated.
SraVaani identifies the spoken language automatically, with no language tag required, and covers languages including tulu, Chakma, Kumaoni, Bajjika and kokborok etc.
Freely available on Hugging Face under an MIT licence, SraVaani could bring speech technology to an estimated 25 crore people whose languages current systems do not serve.
Project Vaani gave India’s languages a voice. SraVaani gives them an ear.
Read more: https://lnkd.in/d3f6KZVP
Access the model: https://lnkd.in/dN42FbNc
Leave a Reply