Effects of Noise on RASTA-PLP and MFCC based Bangla ASR Using CNN

Though Bangla Automatic Speech Recognition (ASR) started its journey since a long time ago, a paltry amount of work is done on Convolutional Neural Network (CNN) based ASR. In this paper, we propose an ASR made with CNN where the performance of two feature extraction methods, namely Mel Frequency Cepstral Coefficients (MFCC) and Relative Spectral Transform - Perceptual Linear Prediction (RASTA-PLP) are compared on Bangla isolated words consisting of digits and speech commands. This paper contributes to the literature of Bangla ASR in three ways. Firstly, Effects of noise is experimented on Bangla speech commands as well as isolated words in CNN based ASR. Secondly, the performance of MFCC and RASTA-PLP are compared in noisy environment using CNN based classifier. Lastly, state-of-the-art accuracy is achieved in CNN based ASR which is 93.18%.