Date of Publication

3-2005

Document Type

Master's Thesis

Degree Name

Master of Science in Computer Science

Subject Categories

Computer Sciences

College

College of Computer Studies

Department/Unit

Computer Science

Thesis Adviser

Jose Ronel Bartolome

Defense Panel Chair

Nelson Marcos

Defense Panel Member

Clement Ong
Jose Ronel Bartolome

Abstract/Summary

The performance of a speaker recognizer like the speaker verifier strongly depends on the input feature. The feature set has to be both highly discriminative and compact to get a good performance. This study aims to introduce a new approach of feature extraction specifically designed for Speaker Verification System. The feasibility of using a Genetic Algorithms/Decision Tree (GA/DT) hybrid approach to choose from the original MFCC (Mel-Frequency Cepstral Coefficient) data the set of feature vectors that appropriately represents personal speech characteristics of a speaker is investigated. The GA is used to evolve selected feature vectors and the decision tree (DT) algorithm is used to evaluate the fitness functions of the chromosomes (feature sets) evolved by the GA. Moreover, it evaluates and compares, in terms of speed and accuracy, the performance of the Speaker Verification System (SVS) using the General Approach and the SVS with the GA/DT Hybrid Approach. There are two (2) sets of data used in the experiments, namely, the clean data taken from the TIMIT database that is provided in the SPEAR database of CSLU and the unclean data taken from real life recordings of the speakers that deal with constant background noise. For the clean data, the performance of the two (2) systems and the existing work are all compared in terms of accuracy. The results show that the performance of the two systems is comparable with the existing work. Moreover, the SVS w/ the GA/DT hybrid approach performs better than the General Approach in terms of accuracy and speed. On the other hand, for the unclean data, the experimental results show that the SVS with the GA/DT hybrid approach performs better than the General Approach in terms of accuracy and speed but the results obtained is not as good as using a clean data. We still think that this is a good result though since we are able to find out that the idea we put in our proposed model worked. Perhaps in the future our proposed model would be improved further.

Abstract Format

html

Language

English

Format

Print

Accession Number

TG03824; CDTG003824

Shelf Location

Archives, The Learning Commons, 12F Henry Sy Sr. Hall

Physical Description

1 v. (various foliations) ; 28 cm. + 1 computer optical disc.

Keywords

Genetic algorithms; Combinatorial optimization; Expert systems (Computer science)--Verification; Automatic speech recognition; Speech processing systems.

Upload Full Text

wf_yes

Share

COinS