89
/
4
Min Read
Researchers at IISc’s SPIRE Lab, in collaboration with ARTPARK and with support from Google, have released SraVaani, the first multilingual Indian speech recognition model trained on 65 Indian languages and dialects, including 40+ languages that today’s speech recognition systems do not officially support. Designed to take speech AI beyond scheduled languages, SraVaani extends automatic speech recognition to several regional and non-scheduled Indian languages that remain underserved by existing speech technology. The model is freely and publicly available on Hugging Face under an MIT licence.
Most speech AI systems work effectively in only a handful of dominant languages, leaving many of the languages and dialects spoken across India without reliable speech-to-text capabilities. SraVaani addresses this gap by covering 20 scheduled languages and 45 regional languages and dialects, potentially opening speech AI capabilities to around 25 crore people as per the 2011 Census whose languages are not properly handled by current systems.
Its coverage is designed to be pan-India, spanning 19 languages from the North East, 16 from eastern India, 9 from the west, 8 from the north, 6 from the south and 5 from central India, along with English and Sanskrit. The model supports languages such as Garo, Angika, Chakma, Kokborok, Tulu, Bundeli and Bajjika, taking Indian dialect speech-to-text and regional-language AI beyond the languages traditionally prioritised by speech technology.
SraVaani is now freely and publicly available on HuggingFace, along with a demo and fine-tuning code, enabling developers and researchers to experiment with, adapt and build on the model. The release is intended to support further work in AI for Indian regional languages, low-resource speech-to-text, automatic language detection, Indian dialect speech recognition and sovereign AI applications.
Evaluated across eight public benchmark datasets,SraVaani delivers accuracy comparable to leading Indic speech recognition systems on India's widely supported languages, achieving the lowest average word error rate among the systems evaluated. Its distinctive strength, however, lies in the long tail of Indian languages.
SraVaani is trained on 40+ additional languages and dialects beyond the scheduled languages that are not officially supported by most existing speech recognition systems. Results on several of these languages are particularly strong, including a 9.5% word error rate on Garo, compared with 69.4% for the next-best system evaluated.
"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," says Prof. Govindan Rangarajan, Director, Indian Institute of Science (IISc).
Prof. Prasanta Kumar Ghosh, Professor, IISc and Principal Investigator, Project Vaani, said, “When we began Project Vaani four years ago, the aim was simple: that voice AI should work for every Indian, not only for those whose languages already had the resources behind them. SraVaani is that aim taking shape. It reaches languages today’s systems do not serve at all, and it is free and open for anyone to build on. For us at IISc and ARTPARK, this is what sovereign and inclusive AI means in practice: not only that India builds its own voice AI models, but that those models understand every Indian who speaks to them. Democratising AI is not about access to tools alone. It is about ensuring no one is left outside this transformation because of their language. The satisfaction, after four years, is knowing that somewhere a person will speak into a machine in Angika or in Garo, and be understood.”
At the foundation of SraVaani is the Vaani dataset, developed through Project Vaani at IISc, to capture natural, spontaneous speech from across the country. Project Vaani has recorded more than 31,000 hours of speech from 156,000 people across 165 districts in 28 states, with coverage extending across three Union Territories. Speakers were asked to describe images in their own words rather than read prepared sentences, allowing natural speech, dialects and regional variations to be represented in the data.
SraVaani was trained in three stages. It was first trained on the complete Vaani speech corpus, followed by audio-image alignment using 11.8 million Vaani audio-image pairs. In the final stage, the model was trained on transcribed speech from Vaani and other publicly available Indian speech datasets, including IndicVoices, RESPIN, SPRING-INX, SPICOR and SYSPIN. The model produces text across 10 different scripts and can automatically identify the language being spoken, eliminating the need for a language tag to be specified in advance.
