Automatic speech recognition: a comparative evaluation between neural networks and hidden markov models

Abstract

In this work we do a comparative evaluation between Artificial Neural Networks (RNA's) and Continuous Hidden Markov Models (CDHMM), in the framework of the recognition of isolated words, under the constrain of using a small number of features extracted from each voice signal. In order to accomplish such comparison we used two models of neural networks: the Multilayer Perceptron (MLP) and a variant of the Radial Basis (RBF), and some HMM models. We evaluated the performance of all models using two different test set and observed that the neural models presented the best results in both cases. Seeking to improve the HMM performance we developed a hybrid system, HMM/MLP, that improved the results previously obtained with all HMMs, and even those obtained with the neural networks for the all previous HMM, and even the neural nets for the hardest test set case.