Natural Language Processing for recognizing Bangla speech with regular and regional dialects: A survey of algorithms and approaches

Abstract

Natural Language Processing (NLP) is one of the fundamental domains of Artificial Intelligence (AI). In this paper, we present a systematic review of NLP based research for recognizing Bangla speech with regular and regional dialects. We describe 23 research papers based on Bangla speech recognition in regular accents and regional dialects. Due to the cultural diversity, Bangla has many dialects with distinctive regional pronunciations, complex vocabulary, and syntax. These characteristics of Bangla create several challenges for implementing NLP successfully. In this paper, we focus on NLP’s vital role in speech recognition, which is essential to virtual assistants, transcription services, and language-learning applications. We discuss several methods such as advanced language models, comprehensive datasets, and continuous adaptation to overcome the challenges of recognizing Bangla with NLP. Several algorithms, such as Deep Neural Networks, Gaussian Mixture Models (GMM), Linear Predictive Coding (LPC), Mel frequency cepstral coefficients (MFCC), etc. have been developed to detect the spoken words in Bangla. These existing research works have faced various challenges like scarcity of data, dialectal variability, computational resource requirements, etc. For mitigation of the challenges and further advancement in this research area, we discuss some research scopes including dialect identification through large datasets, low-resource dialect modeling, development of deep learning model for end-to-end ASR systems, continuous learning, etc. This research will help us understand the present status and challenges associated with NLP-based Bangla speech recognition.