Indian-language AI models get a boost with release of 10,000 hours of open speech data
A consortium of universities and startups has published one of the largest open speech datasets covering 22 languages and dozens of dialects.
Bengaluru: A consortium of research institutions and technology companies has released more than 10,000 hours of transcribed speech across 22 Indian languages under an open licence.
Researchers say the dataset fills a long-standing gap: most commercial voice assistants perform well in English and Hindi but struggle with regional languages and code-mixed speech.
The recordings were gathered from volunteers in more than 400 districts, with consent forms in local languages and an option to withdraw recordings at any time.
Why it matters
Voice is the most natural interface for millions of first-time internet users. Better speech recognition could make everything from agricultural advisories to government services easier to use.