IISc's Google-backed SPIRE lab launches free AI speech model for 65 languages

A large number of Indian languages and dialects still lack reliable speech-to-text technology, leaving millions of speakers often left out of the mainstream AI systems. To address this challenge, researchers at the Indian Institute of Science (IISc)’s Signal Processing, Interpretation and REpresentation Laboratory (SPIRE Lab) developed SraVaani.

SraVaani is a multilingual speech recognition model covering 65 Indian languages and dialects. The model was created with support from Google in collaboration with AI and Robotics Technology Park (ARTPARK). It includes more than 40 languages and dialects that are not officially supported by many existing speech recognition systems.

Govindan Rangarajan, director, IISc, explained that "As India builds its own capabilities in artificial intelligence, inclusive language technology must be part of that ambition. At IISc, research has always been in service to the nation, and SraVaani, serving more than 60 Indian languages, is a contribution toward that."

SraVaani makes speech AI regional

SraVaani covers 20 scheduled languages along with 45 regional languages and dialects. Its developers estimate that the model could extend access to speech AI for around 25 crore people, based on 2011 census data.

The model has been designed to provide wider geographical coverage across India. Its language portfolio includes 19 languages from the Northeast, 16 from eastern India, nine from western India, eight from northern India, six from southern India and five from central India, along with English and Sanskrit.

Languages and dialects include Garo, Angika, Chakma, Kokborok, Tulu, Bundeli and Bajjika. This broader support could support applications involving regional-language speech recognition, low-resource language AI and automatic language identification.

Accessible to developers, researchers

SraVaani is publicly available on ‘Hugging Face’, an open-source platform and community, under a Massachusetts Institute of Technology (MIT) licence. Developers and researchers can access the model, demonstration tools and fine-tuning code to test, customise and build applications around Indian languages.

The model was assessed using eight publicly available benchmark datasets. According to the IISc researchers, SraVaani achieved accuracy comparable with leading Indic speech recognition systems for several widely supported Indian languages and recorded the lowest average word error rate among the systems evaluated.

The open release could support more research in regional-language AI and applications such as speech-to-text, language detection and sovereign artificial intelligence.