Speaker verification with probability measurements
Date of Publication
2007
Document Type
Bachelor's Thesis
Degree Name
Bachelor of Science in Computer Science
Subject Categories
Computer Sciences
College
College of Computer Studies
Department/Unit
Computer Science
Thesis Adviser
Jocelynn W. Cu
Defense Panel Chair
Clement Y. Ong
Defense Panel Member
Joel P. Ilao
Karlo Shane O. Campos
Abstract/Summary
The usual application of speaker verification (SV) system is based on a Yes/No question, Are the speakers of these two signals the same? and on the assumption that cooperative users speak prompted words of phrases for comparison. However, there are cases when such constraints are not possible. Consider a scenario where there is a need to determine the likelihood of a suspected person being or not being the speaker of a speech recording. In such cases, the speaker verification system should be able to work even without explicit user cooperation and independent of the spoken words or phrases.
This study aims to develop a text-independent speaker verification system that will be able to identify whether the speaker of two speech signals are the same or not. The speech feature extraction method used in this study is Linear Predictive Coding (LPC). The Euclidean Distance of the LPC coefficients is used to measure the degree of similarity between the two signals. A threshold value whether the two signals are similar (less than 0.5) or not (greater than 0.5) is utilized.
The average Euclidean Distance between the speaker from the reference signal and target speaker from the test signal has an average of 0.4383, which means that they are the same speakers. This value was compared to two other values obtained from a different female speaker and from a male speaker, which is 0.5877 and 0.6119, respectively. Comparisons with the two other speakers yielded results greater than o.5 which means that they are not similar.
The method is further validated using 20 speakers (10 males and 10 females). The False Rejection Rate and False Acceptance Rates are acquired. The average accuracy rates for male speakers are 89.45% and 84.80% for female speakers.
Abstract Format
html
Language
English
Format
Accession Number
TU13525
Shelf Location
Archives, The Learning Commons, 12F, Henry Sy Sr. Hall
Physical Description
1 (various folaitions) ; ill. (some col.) ; 29 cm.
Keywords
Speech processing systems; Automatic speech recognition; Speech perception; Voice; Voice frequency
Recommended Citation
Vinluan, A. S. (2007). Speaker verification with probability measurements. Retrieved from https://animorepository.dlsu.edu.ph/etd_bachelors/2154