The Kathbath dataset features 1684 hours of labeled speech data across 12 Indian languages, including Bengali. This dataset is specifically designed to evaluate and develop Automatic Speech Recognition (ASR) systems for Indian languages in general-use scenarios. The "Test-Unknown" subset focuses on providing challenging test cases to evaluate ASR model robustness in less predictable contexts. Submitted by Tahir Javed, this dataset supports efforts to improve multilingual ASR systems in India.
© 2026 - Copyright AIKosh. All rights reserved.