A common daily question in many Indian homes is what a grandmother cooked for the day. Asking this to an AI raises issues of language recognition, accent handling and code mixing between Hindi and English. The same problem grows larger for lesser known tongues such as Tulu, Bundeli, Kodava or Santali.
Recent progress in artificial intelligence allows machines to answer queries, translate text and generate stories. Yet most systems rely on vast collections of digital material, largely in English. Many Indian languages lack comparable online resources, leaving them classified as low resource.
India presents extra difficulties because speech patterns shift across short distances and some expressions exist only in spoken form. Researchers at AI4Bharat, based at IIT Madras, aim to close this gap by gathering speech samples from more than five hundred districts. They work with local colleges to set up recording sites where residents describe festivals, recipes and village life.
Recordings are later transcribed by hand to create paired audio and text data. This material trains models to match spoken sounds with written words. The approach mirrors how children learn by exposure to repeated examples rather than formal rules. The project seeks to improve access to technology for millions of speakers across the country.
