HuggingFace and IISc create language database of India

HuggingFace and the Indian Institute of Science (IISc) have announced a partnership that will make Vaani, India’s largest multimodal and multilingual dataset, available to developers around the world. The project aims to create inclusive AI technologies that take into account the linguistic and cultural diversity of the country.

The Vaani project, launched in 2022 jointly by IISc/ARTPARK and Google, sets an ambitious goal: to collect more than 150,000 hours of speech and 15,000 hours of transcribed text from 1 million people across all 773 districts of India. The uniqueness of the project lies in its geocentric approach, which allows collecting data about dialects and languages used in remote regions, not just mainstream languages.

Currently, the project is being implemented in phases. The first phase, covering 80 districts, has already been completed, and the data is publicly available. The second phase is now underway, expanding coverage to another 100 districts.

As of February 15, 2025, the open part of the dataset includes a transcribed subset containing 790 hours of audio from approximately 700,000 speakers and covering 70,000 images. This resource is intended for various tasks, including:
– Speech recognition: training models for accurate transcription of spoken language
– Language modeling: creating more advanced language models
– Segmentation: identifying individual speech units to improve transcription accuracy

The partnership between HuggingFace and IISc/ARTPARK aims to increase the accessibility and usability of the Vaani dataset, which should contribute to the development of AI systems that better understand the diversity of Indian languages and meet the digital needs of the country’s population.

This project represents an important step in democratizing AI technologies and creating more inclusive solutions that take into account the linguistic diversity of one of the world’s largest countries.