Lipi-Ghor: 885-Hour Diarized Bengali Speech Corpus for ASR
October 28, 2026
Lipi-Ghor provides 885 hours of weakly-supervised Bengali audio from 998 YouTube channels, covering 17 domains including regional dialects. It is the first large-scale Bengali corpus to feature speaker diarization and long-form audio using an SSTT annotation format.
HOW THIS AFFECTS YOU
●
builderYou can now train more robust Bengali ASR models using long-form, multi-speaker data.
●
researcherThis provides a scalable benchmark for low-resource speech research in the Bengali language.