The speech recogniton system based on structure equivalent fuzzy RBF neural network

Abstract

A structure equivalent fuzzy radial basis function (FRBF) neural network with five layers is proposed in this paper. A two stage algorithm for the parameters learning of the network is presented, which first determine the center value and width of the membership function according to the classification information of training samples, and then adjusts the weights between the fourth layer and the fifth layer by the gradient descent method. When the input data dimension number is large, the reasoning of conventional fuzzy system often cannot correctly work because the degree of each rule become too small and sometimes they cause underflow. The network reasoning can correctly work by adding a compensated factor on membership function. The structure equivalent FRBF neural network was used in speech recognition system. It is able to determine the fuzzy rule numbers according to the vocabulary to be recognized. The experimental results showed that the proposed structure equivalent FRBF neural network has demonstrated superior performance compared to radial basis function (RBF) neural network, such as higher recognition and better robustness.