Neural Network Configurations Analysis for Multilevel Speech Pattern Recognition System with Mixture of Experts

Abstract

This chapter proposes to analyze two configurations of neural networks to compose the expert set in the development of a multilevel speech signal pattern recognition system of 30 commands in the Brazilian Portuguese language. Then, multilayer perceptron (MLP) and learning vector quantization (LVQ) networks have their performances verified during the training, validation and test stages in the speech signal recognition, whose patterns are given by two-dimensional time matrices, result from mel-cepstral coefficients coding by the discrete cosine transform (DCT). In order to avoid the pattern separability problem, the patterns are modified by a nonlinear transformation to a high-dimensional space through a suitable set of Gaussian radial base functions (GRBF). The performance of MLP and LVQ experts is improved and configurations are trained with few examples of each modified pattern. Several combinations were performed for the neural network topologies and algorithms previously established to determine the network structures with the best hit and generalization results.