Sarvam AI, AI4Bharat launch Indic diarbench covering 22 Indian languages
Sarvam AI and AI4Bharat just launched Indic DiarBench, a new dataset designed to make speech recognition smarter for 22 Indian languages.
It's packed with 108 hours of real conversations (think meetings, fast turn-taking, and people talking over each other) so tech can actually handle how we speak in real life.
You can check it out now on Hugging Face.
Indic DiarBench includes labeled realistic audio
The dataset brings together 485 unique speakers in the meeting portion from cities and villages, with recordings from meetings plus "in-the-wild" audio in the top 10 Indian languages, where around 750 unique speakers are included.
It even mixes in background noise and different room settings to keep things realistic.
Everything's carefully labeled so researchers can test how well speech models recognize who's talking (and what they're saying) even when English gets mixed in.
All resources are open for anyone curious or building the next big thing in voice tech.