Lipi-Ghor: A Large-Scale, Diarized, Multi-Domain Bengali Speech Corpus and a Scalable Pipeline for Building It

Lipi-Ghor. A Large-Scale, Diarized, Multi-Domain Bengali Speech Corpus and a Scalable Pipeline for Building It. By Sanjid Hasan and three others, from the Khulna University of Engineering and Technology and the Bangladesh University of Engineering and Technology. Published on the twenty-sixth of October.

Bengali is spoken by over two hundred and thirty million people, yet it remains severely under-resourced in spoken language research. The open Bengali speech recognition corpora that do exist are confined to short speech, either read aloud or crowd-sourced, and they lack both speaker diarization and long-form audio. Lipi-Ghor is a large-scale, weakly supervised Bengali speech corpus built to fill that gap. To the authors' knowledge it is the first Bengali speech corpus that pairs large-scale coverage with speaker-diarized, long-form annotations.

The corpus holds eight hundred and eighty-five hours of audio. It comes from nine hundred and ninety-eight YouTube videos, spread over five hundred and ninety-six unique channels, and spans seventeen domains. Among them are audiobooks, drama, religious sermons, news, and regional dialects. Each recording is annotated in what the authors call SSTT format: Speaker, Speech, Transcription, Timestamp.

Because caption targets are noisy, there is also a human-annotated gold evaluation split. It was verified by seventeen volunteer annotators, over thirty hours of human-in-the-loop correction, and it is stratified across all seventeen domains.

The annotations come from combining automated diarization with the Bengali caption tracks that YouTube already carries. The construction method is offered as a replicable pipeline for any language with enough captioned YouTube presence, and the authors report that targeted curation of channels across a handful of rich domains reaches corpus scales previously out of reach for low-resource languages. On licensing, the original contributions, meaning the video-ID lists, the diarization output, the pipeline, and the gold annotations, are released under CC-BY-4.0. Caption-derived transcripts are distributed for research use only, and no raw audio is redistributed.

The first users are people building Bengali speech recognition. The authors benchmark four widely-used ASR models, both zero-shot and fine-tuned, and all four degrade consistently on emotive, dialectal, and long-form content. That is exactly the material the corpus is rich in. Beyond Bengali, the pipeline is aimed at anyone working on a low-resource language with a captioned video footprint.

Two things are on offer here. One is a corpus whose long recordings and speaker labels address what existing open Bengali resources lack. The other is the recipe that produced it, written down so that another language with enough captions can follow the same route.