IISc's SPIRE Lab Open-Sources SraVaani Speech Model for Low-Resource Indian Languages

Researchers at the Indian Institute of Science's (IISc) SPIRE Lab, in collaboration with ARTPARK and with support from Google, have released SraVaani, a multilingual Indian speech recognition model covering 65 languages and dialects.

The model is designed to extend automatic speech recognition into regional and other low-resource Indian languages that remain underserved by existing speech technology, according to IISc.

"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," said Prof Govindan Rangarajan, Director, Indian Institute of Science, in a statement.

SraVaani is publicly available on Hugging Face under an MIT licence, according to IISc.

IISc said the model covers 20 scheduled languages and 45 regional languages and dialects. Its coverage spans languages from across India, including Garo, Angika, Chakma, Kokborok, Tulu, Bundeli, and Bajjika. According to the institute, more than 40 of these languages and dialects are not officially supported by most existing speech-recognition systems.

SraVaani is built on Project Vaani, an IISc initiative that has collected more than 31,000 hours of speech from more than 156,000 speakers across 165 districts, the institute said. The project spans 28 states and three union territories, with participants describing images in their own words rather than reading prepared sentences. 

According to IISc, the model 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, SraVaani was trained on transcribed speech from Vaani and other Indian speech datasets, including IndicVoices, RESPIN, SPRING-INX, SPICOR, and SYSPIN.

The researchers evaluated SraVaani across eight public benchmark datasets covering Indian languages. IISc said the model achieved the lowest average word error rate among the systems evaluated, while delivering accuracy comparable to leading Indic speech-recognition systems on India's widely supported languages.

On the Garo benchmark, SraVaani recorded a word error rate of 9.5%, compared to 69.4% for the next-best system evaluated, the institute said.

Prof Prasanta Kumar Ghosh, Professor at IISc and Principal Investigator of Project Vaani, said the project began with the goal of making voice AI usable across India's linguistic diversity.

"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," Prof Ghosh said. "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."

SraVaani aims to support further research in low-resource speech-to-text, Indian language and dialect recognition, automatic language identification, and sovereign AI applications. Its release also gives developers and researchers access to the model, a demo, and fine-tuning code for further experimentation.

Project Vaani remains the underlying data foundation for the work. Its district-by-district collection approach aims to create speech resources that reflect how Indians actually speak rather than relying only on prepared or standardised text, the institute said.