Multimodal Mosaic: Hugging Face and IISc team up to boost AI for diverse Indian languages

The Indian Institute of Science (IISc) and Hugging Face have formed a strategic partnership to make Vaani, India’s largest open source multilingual dataset, more accessible to developers worldwide. This collaboration builds upon Project Vaani’s initial launch in 2022 with Google, which aims to create a comprehensive dataset representing India’s vast linguistic landscape.

Project overview: The Vaani dataset represents a groundbreaking effort to capture India’s linguistic diversity through a geo-centric approach that includes remote dialects and languages often overlooked in mainstream datasets.

• The project targets collecting over 150,000 hours of speech and 15,000 hours of transcribed text from 1 million people across 773 districts
• Phase 1 has already covered 80 districts and been open-sourced, while Phase 2 is expanding to 100 additional districts
• The dataset currently includes 790 hours of transcribed audio from approximately 700,000 speakers covering 70,000 images

Technical capabilities: Vaani’s comprehensive language coverage and diverse data collection approach enables the development of sophisticated AI models across multiple applications.

• The dataset supports speech-to-text, text-to-speech, and code-switching applications between Indian languages and English
• Speaker identification and verification models can be developed using data from over 80,000 speakers
• Speech enhancement systems benefit from the dataset’s detailed tagging system
• The collection methodology makes it particularly valuable for improving multimodal language sets

Real-world applications: The dataset’s practical applications span across various sectors that require multilingual AI capabilities.

• Educational tools and telemedicine platforms can leverage the dataset for improved communication
• Voter helplines and media localization services can better serve diverse linguistic populations
• Multilingual smart devices can be developed with more accurate language understanding
• Healthcare solutions can be enhanced to serve patients in their preferred languages

Data accessibility: The partnership focuses on making the dataset more user-friendly and accessible to developers globally.

• Researchers can access detailed district-wise language distribution through the Hugging Face platform
• A specialized transcribed subset is available for specific development needs
• The dataset covers 54 languages and includes speakers from diverse educational and socioeconomic backgrounds
• Real-life data collection environments ensure the dataset’s practical applicability

Future developments: By expanding the project’s reach, IISc/ARTPARK and Google are working to make Vaani a comprehensive representation of India’s linguistic landscape.

• Phase 2 expansion ensures coverage across all Indian states
• Developers and researchers are encouraged to contribute feedback and explore collaboration opportunities
• The project maintains an active communication channel through dedicated email and feedback forms

Looking ahead: The success of the Vaani dataset could serve as a model for other regions with significant linguistic diversity, potentially inspiring similar initiatives worldwide while helping to preserve and digitize language variations that might otherwise be overlooked in the development of AI technologies.