Intelligent judge neural network for speech recognition

Abstract

An intelligent judge neural network (IJNN) is developed to make decisions out of contradictory arguments, which may come from different classifiers with different characteristics and/or input features. For speech recognition applications a multi-layer perceptron classifies the word as a spectro-temporal pattern, while a neural prediction model or hidden-control neural network relies on dynamic nature of the speech signal. The “judge≓ accepts input values from the lower-level neural network classifiers and provides ruling verdicts. Two intelligent judges have been investigated. The “neuro-judge≓ rules by extracting decision rules from training data, i.e. disputes between the two classifiers, while the “fuzzy-judge≓ just utilizes min-max operations. The IJNN demonstrates better recognition rates. More importantly its performance is much less sensitive to the choice of training data.