Verificação de locutor utilizando modelos de Markov escondidos (HMMs) de densidades discretas.

Abstract

Hidden Markov Models (HMMs) are becoming popular in pattern recognition because they present a strong mathematical structure solid and so they provide a theorectical basis for very many applications in voice processing systems. They can also provide a reduction in complexity when compared to other methods. Speaker recognition using HMMs. like other pattern recognition techniques, can be performed in two phases: training and classification. For the training phase each speaker uses a individual HMM. The model is built after the speech has beeen analised and the Linear Predictive Coding (LPC) parameters representing that particular speaker have been obtained. The LPC coeficients are then discretizes by a vector quantizer. The discretized parameters are used for running an iteractive algorithm (Baum-Welch algorithm) calculating a probability wich best represents that speaker. Classification, in this application, speaker verification, consists in using the H MM obtained in the training phase to calculate and check whether that particular speaker provides an acceptable probability to be considered a valid user (customer).