Adaptação ao locutor usando a técnica MLLR

Abstract

In this work a study of the technique of speaker adaptation called MLLR, Maximum Likelihood Linear Regression was made. The tests have been done using continuous speech applications and only the means of continuous hidden Markov models (HMM) have been adapted. The basic point of the technique is the partition of these means in regression classes for the generation of the transformation matrix. Moreover, the amount of material for adaptation of a speaker independet system is very important. Being thus, some alternatives for regression classes construction have been explored. Methods based on phonetic classification and based on distance metrics have been tested, varying also the number of regression classes. After tests, with a varied number of adaptation sentences, was verified that the better approach is to use only three regression classes with four adaptation sentences, but research still must be made in the area.