Implementing a Fuzzy Relational Neural Network for Phonetic Automatic Speech Recognition

In this chapter we present the implementation of a speech recognizer based on a fuzzy relational neural network model. In the model, the input acoustic phonetic features are represented by their respective fuzzy membership values to linguistic properties. The membership values are calculated with Π functions, and dense trapezoidal functions. The weights of the connections between input and output nodes are described in terms of their fuzzy relations. The output values are obtained by the use of the max-min composition, and are given in terms of fuzzy class membership values. The learning algorithm used is a modified version of the gradient-descent back-propagation algorithm following the model described by Pedrycz in [3]. Some results are presented as well.