Date of Award

Spring 2005

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical & Computer Engineering

Program/Concentration

Computer Engineering

Committee Director

Stephen A. Zahorian

Committee Member

Vijayan K. Asari

Committee Member

Frederic D. McKenzie

Call Number for Print

Special Collections LD4331.E55 S255 2005

Abstract

This thesis is an extension of previous work on speaker identification, but with a primary focus on the generalization performance of patterns classifiers for speaker identification. The goal for all pattern classifiers is to generalize to new data, not just perform well on the training data set. The error on the training data is typically a biased estimate of the generalization error. The training set error tends to be smaller than the generalization error if the model is selected such that it minimizes the training error. Therefore, the training error alone should not be used as a good predictor of classifier performance.

The main focus of this research is to investigate the issues related to the commonly observed phenomena that classifiers have high identification performance on training data, but poor accuracy for test data. Understanding these issues is crucial in order to improve the generalization performance.

As detailed in this work, the basic algorithms from our previous work on speaker identification have not been changed, but many details in the signal processing were changed to improve identification results. Also, a recently developed non-parametric classifier, called a Support Vector Machine (SVM), which is claimed to have good generalization performance, was tested on the NTIMIT database. However, only negligible improvements were obtained in using an SVM versus a Neural Network (NN) classifier. In other experiments, performed with the SPIDRE speaker identification database, using the changes in the signal processing details, speaker identification accuracy of over 84% was obtained on test sentences, using all 45 target speakers in the database. These results are among the best reported for the SPIDRE database, and indicate that excellent results can be obtained using simple classifiers.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

DOI

10.25777/y1kf-3r48

Download

Share

COinS