NN and Hybrid Strategies for Speech Recognition in Romanian Language

Abstract

In this paper we present results obtained with learning structures more "human likely" than the very effective and widely used hidden Markov model. Good results were obtained with simple artificial neural networks like the multilayer perceptron or the Kohonen maps. Hybrid structures have proven also their efficiency, the neuro-statistical hybrid applied enhancing the digit recognition rate of the initial HMM. Also fuzzy variants of the MLP and HMM gave good results in the tested tasks of vowel recognition.