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

Area of Science:

• Natural Language Processing
• Speech Technology
• Computational Linguistics

Background:

• Bangla, the seventh most spoken native language, is underrepresented in voice recognition research.
• Existing speech recognition systems often fail to capture the linguistic diversity and regional pronunciation variations within Bangla.
• A significant gap exists in accessible, comprehensive datasets for developing robust Bangla automatic speech recognition (ASR) systems.

Purpose of the Study:

• To present an innovative dataset for Bangla voice recognition, emphasizing regional linguistic differences.
• To provide a resource for researchers developing automatic Bangla regional speech recognition systems.
• To address the underrepresentation of Bangla in speech technology and promote linguistic diversity in NLP.

Main Methods:

• Collected 298 frequently used Bangla words (233 regional, 65 standard) from native speakers across eight regions in Bangladesh.
• Recorded 2439 audio segments contributed voluntarily by 85 native speakers.
• Assessed audio segments by ten university students to ensure quality and relevance.
• Incorporated background noise into the dataset to simulate real-world conditions for model training.

Main Results:

• Developed a multipurpose dataset containing standard and diverse regional Bangla words with varied pronunciations.
• The dataset captures specific regional pronunciations and linguistic nuances crucial for ASR development.
• The resource includes audio segments with background noise, enhancing its utility for robust model training.

Conclusions:

• The developed Bangla voice recognition dataset addresses a critical gap in the field.
• This resource has the potential to significantly advance Bangla ASR and NLP research, particularly concerning linguistic diversity.
• The dataset's modular design allows for future expansion, promising continued contributions to speech technology for underrepresented languages.