BRADS and BRWDS: Multipurpose audio and text datasets for automatic Bangla regional speech recognition

TL;DR

This paper introduces a new dataset for Bangla speech recognition, focusing on regional variations and aiming to improve NLP research in Bangladesh.

Contribution

The dataset includes regional Bangla words and pronunciations, collected from native speakers across multiple Bangla-speaking regions.

Findings

The dataset includes 298 Bangla words with regional pronunciations collected from native speakers.

It contains 2439 audio segments from 85 contributors, allowing for real-world scenario modeling.

The modular design supports future expansion with new regional words.

Abstract

This paper presents an innovative approach to Bangla voice recognition. Although Bangla is the seventh most spoken native language globally, it remains underrepresented in voice recognition research. The dataset contains 298 frequently used Bangla words, including 233 regional words and 65 standard Bangla words. These terms, encompassing various regional pronunciations and meanings, were collected from native speakers in Dhaka, Chattogram, Barisal, Mymensingh, Rajshahi, Sylhet, Rangpur, and Khulna. The 2439 audio segments in the dataset were contributed voluntarily by 85 native speakers and assessed by ten university students. This resource is intended for researchers working on automatic Bangla regional speech recognition systems, with an emphasis on capturing regional pronunciation and linguistic differences. The dataset allows researchers to recreate real-world scenarios during model training by incorporating background noise. Additionally, its modular construction enables further expansion to include new regional words. This multipurpose dataset addresses a critical gap in Bangla speech recognition research and has the potential to drive significant advancements in natural language processing (NLP), particularly with regard to linguistic diversity in Bangladesh.

Figures