Reconhecimento de fala utilizando modelos de Markov escondidos (HMM's) de densidades contínuas.

Abstract

This dissertation presents a theoretical study and the software implementation of a speech recognition system, based on Hidden Markov Models (HMM's). HMM is a mathematical tool that makes it possible modeling of the speech sounds in terms of a probabilistic structure. In order to do this, use is made in this work of HMM's of the left-right type with five states and continuous fdp's, to represent the observation vectors probability on each state of the Markov chain. The observation vectors, which are nine-dimensional, are formed by eight cepstral coefficients and the logarithm of the segmentai energy as the nineth parameter. The HMM system is divided into two stages: training and classification. In the training stage, the Baum-Welch algorithm is used to reestimate the final values of the models. On the other hand, the classification stage makes use of the Viterbi algorithm to provide the maximum-likelihood value between the test sentence and the reference HMM's. The evaluation of the proposed system is made considering two different types of voice recognition: the independent speaker recognition and the dependent speaker recognition. In both cases, specially on the speaker dependent mode, the avaluation made given results really satisfactory, account to experimenting general conditions. In addition, some important conclusions are obtained in order to provide a posterior optimization on the proposed system. Finally, i t is expected that this work contributes in a positive way for the motivation of new studies on man-machine voice communication.