Speaker recognition using gaussian mixture model

Abstract

Speaker recognition is the process of identifying a person based on its voice characteristics. A text-independent speaker recognition system using a Gaussian mixture model (GMM)for a set of 10 speakers, selected from TIMIT database, is developed in this project. The design of an optimum speaker recognition focuses on the selection of parameters that increase the identification rate and minimize false acceptance (FA) and false rejection (FR) errors. Speaker identification and speaker verification experiments were conducted to evaluate the performance of the system. A variation in the number of GMM components and the dimension of Mel-Frequency Cepstral Coefficients (MFCC) features were studied to select the optimum parameters. But most importantly the effect of preprocessing the speech signal, using short-time energy and zero crossing rate, at the input of the speaker recognition system have been investigated and compared to the use of a raw speech signal input. In the training phase an Expectation- maximization (EM) algorithm that is initialized by a K-mean clustering method, was used to estimate the speaker model’s parameters. Finally the speaker recognition decision is based on a maximum likelihood test that is performed in both tasks of speaker identification and speaker verification.