Reconhecimento automático de locutor utilizando modelo de misturas gaussianas treinado pelo algoritmo de maximização da expectativa

Abstract

Developing systems that can recognize or identify individuals has becoming a growing need in applications that require verification and grant of human identity. There are several systems that use biometrics to recognize a certain individual, among them, automatic speech recognition that uses speech as recognition data. In this work, Mel Frequency Cepstral Coefficients were extracted to model speakers using statistical tools. For the modeling of the vocal tract of an individual, the Gaussian Mixture Model (GMM) was used. The parameters of the GMM model were adapted or trained by the Expectation Maximization algorithm (EM). Thus 40 models of 40 speakers were created and then tested. Speaker recognition is promising for many practical ap-plications, including tasks that can facilitate, speed up and improve identity verification processes.