Teaching AI to speak India: IIT Madras lab records voices from over 500 districts

Most AI systems learn from enormous amounts of digital data — books, websites, videos and conversations — and much of it is in English. Many Indian languages do not share that advantage. Researchers call them "low-resource languages": tongues such as Tulu, Bundeli, Kodava and Santali have little digital content for a machine to learn from, and in some communities, words spoken every day have never been written down. AI4Bharat, a research lab at IIT Madras, is trying to close that gap by building AI that can understand, speak and translate India's many languages.

The challenge goes beyond scarce data. India is one of the world's most linguistically diverse countries, where the way people speak can change from one district to the next, and the same language can carry different accents, dialects and local words. "Many Indian languages are considered low-resource languages," says Kaushal Bhogale, a PhD researcher at AI4Bharat.

To gather real examples, the team has travelled to more than 500 districts, first connecting with local colleges and community organisations before setting up recording booths for volunteers. Instead of reading scripted sentences, participants are encouraged to talk about their own lives — how their family celebrates Diwali, the dishes prepared during festivals, wedding traditions, or stories about their village. The conversations capture the accents, dialects and expressions that make each language unique. "People are happy to share their life experiences," Bhogale says.

Every recording then goes to human transcribers, who write down exactly what was said. These speech-and-text pairs become the training material: as a model processes thousands or even millions of them, it connects sounds with words, words with meanings and sentences with ideas, until it can transcribe conversations, translate between languages or answer questions. "Researchers found that as we keep showing the AI more examples, its ability to recognise patterns becomes better," says Bhogale. "That's why collecting data is so important."