Speaker identification

Abstract

Over the past few years, it was shown that certain secured access was being manipulated. Speaker identification has shown that it can be used as a means of security. This is since voice contains various parameters such as identity, emotion, attitude, health and sex. Speaker identification is the process where a speaker is determined from a set of speakers already enrolled in the system. In this dissertation, a speaker identification system which identifies a person by his voice is presented. The developed method makes use of the Texas Instruments Massahusetts Institute of Technology (TIMIT) database which is input to the system and pre-processing techniques are performed on the speech files. The voice feature extraction employed is the Mel Frequency Cepstrum Coefficients (MFCC). The Waikato Environment for Knowledge Analysis (WEKA) data mining tool is used to select the most important features overall. The MFCC features are fed to the Support Vector Machine (SVM) and a speaker model for each pair of speakers is computed. The Linear and Radial Basis Function (RBF) kernel functions were used in SVM. Later, the Directed Acyclic Graph Support Vector Machine (DAGSVM) is used to identify the unknown speaker. The system was also implemented using the Distance Measure Technique. During training, the centroid for each person is calculated. The unknown feature set is identified by taking the minimum Euclidean Distance between the tested speaker and each speaker registered. The results show that the best method overall is the DAGSVM with the RBF kernel.