Comparison Study and Result Analysis of Improved Back-Propagation Algorithms in Bangla Speech Recognition

Abstract

This research is concerned with the study of different improved and faster back-propagation (BP) algorithms of neural networks and the analysis of recognition result in continuous Bangla speech. For speech recognition, a comparison study on neural networks and speech recognition result analysis with different improved and faster BP algorithms (such as, BP with momentum, variable learning rate BP, resilient BP, conjugate gradient BP and Levenberg-Marquardt BP algorithms) have been done. In this research, the MATLAB Neural Network Toolbox 7.12.0 is used to create, train and simulate the feedforward neural network with the BP learning algorithm. The convergence obtained from standard BP algorithm is very slow; that's why, this research proposes different improved and faster BP algorithms to solve the speech recognition problems. The developed system has been justified by several networks trained with different Bangla speech words. To test the performance of the system, 20 samples of 50 Bangla speech words have been used; from which 10 samples of 50 words are used as training pattern and another 10 samples of 50 words are used as testing pattern in the network. The binary features of speech words have been generated using dynamic thresholding algorithm. The recognition system has been achieved recognition rate of 83% using resilient BP algorithm, 90% using conjugate gradient BP algorithm and 90% using Levenberg-Marquardt BP algorithm, respectively, for recognising 50 speech words.