Indian Institute of Science (IISc) releases SraVaani speech AI model covering 65 Indian languages

Key facts

• Researchers at the Indian Institute of Science (IISc) SPIRE Lab, in collaboration with ARTPARK and with support from Google, released the SraVaani multilingual speech recognition model.
• SraVaani is trained on 65 Indian languages and dialects, including 20 scheduled languages and 45 regional languages and dialects.
• The model covers languages such as Garo, Angika, Chakma, Kokborok, Tulu, Bundeli and Bajjika, which many current speech recognition systems do not support.
• SraVaani is designed to extend speech AI coverage to around 25 crore people whose languages are not properly handled by existing systems.
• The model's language coverage spans 19 languages from the Northeast, 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.
• SraVaani is freely available on Hugging Face under an MIT licence, along with a demo and fine-tuning code.
• The model was evaluated across eight public benchmark datasets and achieved the lowest average word error rate among the systems evaluated.
• SraVaani achieved a 9.5 per cent word error rate on Garo, compared with 69.4 per cent for the next-best system evaluated.
• The foundation dataset, Project Vaani, recorded more than 31,000 hours of speech from 156,000 people across 165 districts in 28 states and three Union Territories.
• SraVaani produces text across 10 different scripts and can automatically identify the language being spoken without requiring a language tag in advance.