Probability estimation in hybrid NN-HMM speech recognition systems with real-time neural networks

Abstract

FAPES system is based on a specialized fuzzy ARTMAP NN trained to estimate the observation probabilities in continuous parameter HMM (CHMM) speech recognition systems. The fuzzy ARTMAP classifier transfers after ART resonance, the choice function of all eligible nodes to a single layer perceptron (SLP) defuzzifier. There, the fuzzy scores are mapped to the a-posteriori probabilities of visiting CHMM states. Lower computing time results from estimating observation probabilities with such local error propagation NN, than with well-known multilayer perceptron. The fuzzy ARTMAP NN determines inherent discrimination among generated probabilities, discrimination usually added into HMM training by using the complex MMIE algorithm. Iterative training has been used to instruct the fuzzy ARTMAP and the SLP defuzzifier. The CHMM component was not trained due to small changes of transition probabilities.