RegSpeech12: A Regional Corpus of Bengali Spontaneous Speech Across Dialects
Dataset Description
The RegSpeech12 dataset is a large-scale spontaneous Bengali speech corpus collected from 12 regions of Bangladesh, designed for benchmarking Automatic Speech Recognition (ASR) under dialectal variation.
Unlike existing Bengali resources, RegSpeech12 specifically captures phonological, lexical, and morphological diversity across regions such as Rangpur, Sylhet, Chittagong, Noakhali, Narail, Kishoreganj, Barishal, Habiganj, Comilla, Tangail, Sandwip, and Narsingdi.
This dataset was created to address the gap in regional dialect resources, enabling the development of robust ASR systems that work across dialects, accents, and speech styles.
Dataset Summary
-
Total Audio Segments: 21,313
-
Total Duration: ~100 hours
-
Speakers: 394+
- ~53% male, ~38% female, ~9% mixed-gender
-
Regions Covered: 12 districts, 99 subregions
-
Average Segment Length: 16.9 sec
-
Vocabulary Size: 58,971 unique words
-
Split:
- Train: 17,049 clips (~80 hours)
- Validation: 2,132 clips (~10 hours)
- Test: 2,132 clips (~10 hours)
Dialectal Coverage
The dataset contains recordings across diverse Bengali dialects, with concrete examples of regional phonological and lexical variations.
Some sample regions:
- Sylhet: Distinct intonation and loanwords.
- Rangpur: Lexical and morphological divergence.
- Chittagong: Unique vocabulary and strong prosodic shifts.
- Barishal, Noakhali, Narail, Habiganj, Comilla, Tangail, Sandwip, Kishoreganj, Narsingdi also included.
Data Collection & Validation
-
Collection Protocols: Spontaneous monologues, free speech recordings, ~10 min per participant.
-
Diversification Dimensions:
- Gender balance
- Age variation
- Topical coverage (64 topics, 13 categories including education, politics, sports, family, economics)
- Geographical diversity across 99 subregions
-
Transcription: Annotated using Labelbox by local transcribers familiar with dialects.
-
Validation: Linguist-reviewed with multiple correction cycles.
-
Split Strategy: 80:10:10 train/val/test with balanced sampling.
File Structure
- audio/ → Contains
.wav files of speech clips
- metadata.csv → Includes file name, region and transcription
- train/val/test splits → Predefined partitioning for ASR benchmarking
Applications
- Automatic Speech Recognition (ASR) under dialectal variation
- Dialectology & Sociolinguistics research
- Speech synthesis (TTS) for regional Bengali
- Voice assistants, educational tools, accessibility apps for Bengali speakers
Limitations
- Environmental noise due to naturalistic recordings (phones, open spaces).
- Regional lexical variation may reduce intelligibility for standard Bengali ASR models.
- Not suitable for tasks requiring studio-quality speech.
Relevant Links
Citation
If you use this dataset, please cite:
1@misc{hassan2025regspeech12regionalcorpusbengali,
2 title={RegSpeech12: A Regional Corpus of Bengali Spontaneous Speech Across Dialects},
3 author={Md. Rezuwan Hassan and Azmol Hossain and Kanij Fatema and Rubayet Sabbir Faruque and Tanmoy Shome and Ruwad Naswan and Trina Chakraborty and Md. Foriduzzaman Zihad and Tawsif Tashwar Dipto and Nazia Tasnim and Nazmuddoha Ansary and Md. Mehedi Hasan Shawon and Ahmed Imtiaz Humayun and Md. Golam Rabiul Alam and Farig Sadeque and Asif Sushmit},
4 year={2025},
5 eprint={2510.24096},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2510.24096},
9}
✉️ Contact
