Bangla Speech Processing: An Analytical Study of Feature Extraction and Recognition Methods

Abstract

Speech recognition has always been an interesting yet challenging task for researchers, especially when working with Bangla, which is complex due to its linguistic structure. This research is extensive in scale, encompassing Bangla phonemes, isolated Bangla words, commands, and sentences in the experiments. Bangla speech recognition is a comparison analysis in large scale that focuses on different feature extraction techniques, recognition tools, window frame feature, other methods and techniques applied. A system is developed by writing code in MATLAB. Mel Frequency Cepstral Coefficient (MFCC), Power Spectral Analysis (FFT), and Linear Predictor Coefficient Analysis (LPC) methods are utilized as feature extraction techniques. Time delays neural network (time series) and a two-layer feed forward hidden neural network are used as speech recognition tools. The maximum likelihood method is also incorporated to enhance the accuracy of speech recognition. Blackman, Hamming, and Hanning Window frame techniques are applied in parallel during feature extraction to observe their influences on speech recognition accuracy. The datasets gathered from native speakers. MFCC as a feature extraction technique, combined with two-layer Feed Forward Neural Network (FFNN) or TDNN as speech recognition tools, outperforms FFT and LPC with the deep learning tools. The study discovered that both the quantity of speech samples, the opposite gender’s voice, and different windowing techniques all had an impact on the recognition accuracy rate. This study will encourage researchers to conduct further research to advance Bangla speech recognition