IISc Launches SraVaani AI Model Supporting 65 Indian Languages

IISc researchers have released SraVaani, an open-source speech AI model designed for 65 Indian languages and dialects. Developed with ARTPARK and Google support, the model aims to improve speech recognition for regional and low-resource languages.

IISc has introduced a new model called SraVaani, which aims to enhance the use of AI-based voice technology in various Indian dialects and languages. The model has been developed by IISc’s SPIRE Lab in partnership with ARTPARK and with the support of Google and it mainly centers on languages that do not benefit as much from speech recognition technologies.

The SraVaani model supports 65 regional and Indian languages, out of which 20 are scheduled languages and others are regional languages. Some of the languages provided by the model include Garo, Angika, Chakma, Kokborok, Tulu, Bundeli, and Bajjika apart from well-used Indian languages. The purpose of this model is to introduce speech-to-text technology for those who speak certain languages that are not given much attention by digital platforms.

Also read: National Scholarship for Postgraduate Students: Eligibility

The SraVaani model is based on Project Vaani, one of IISc’s major programs, to understand the linguistic diversity of India. The project has gathered around 31,000 hours of speech data obtained from over 156,000 speakers from 165 regions, thus providing the needed information about natural conversational-speech to the researchers.

SraVaani uses a FastConformer-based automatic speech recognition architecture with a hybrid TDT-CTC decoder. According to the research paper, it was trained on more than 31,000 hours of multilingual speech and evaluated across eight public benchmark datasets. The model recorded competitive results against existing multilingual speech recognition systems, particularly for low-resource languages.

A key feature is its ability to support automatic language identification, reducing the need for users to manually select a language before speech recognition. This could make voice-based applications easier to deploy across India's multilingual population.

Also read: Reckitt's CIDO Reveals the CIO Playbook for Future Leaders

The model has been released as open source under the MIT licence, allowing developers and researchers to access, adapt and build applications using the technology. It is available through Hugging Face along with related resources.

The release could support applications in areas such as education, digital services, banking, healthcare, e-governance and customer support, particularly where users prefer interacting in regional languages.For India's AI ecosystem, SraVaani represents a move towards developing speech technologies tailored to the country's linguistic diversity. By focusing on low-resource languages alongside widely supported ones, the project could help reduce language barriers and broaden access to AI-powered digital services.