Performance evaluation of time-delay fuzzy neural networks for isolated word recognition

Abstract

A novel structure of fuzzy neural network (FNN) for the recognition of isolated Arabic words is suggested. The performance is evaluated by varying the network topology among several experiments to select the optimum structure for our task. A time delay arrangement is incorporated in the training phase to enable the network to discover useful acoustic-phonetic features without being blurred by shifts in the input. The input speech is processed to obtain a set of linear predictive (LP) derived cepstral coefficients. The input vector to the FNN consists of membership values to linguistic properties of the speech while the output vector is defined in terms of fuzzy class membership values. Three techniques were used to enhance the backpropagation training algorithm used to train the network in order to reduce training time and speed up convergence. The effectiveness of the suggested model is demonstrated on a speech recognition task consisting of Arabic phonemes extracted from a consonant-vowel-consonant (C-V-C) personnel database.