Date Approved
3-15-2017
Embargo Period
3-15-2017
Document Type
Thesis
Degree Name
MS Electrical and Computer Engineering
Department
Electrical and Computer Engineering
College
Henry M. Rowan College of Engineering
Sponsor
National Science Foundation
Advisor
Ramachandran, Ravi P.
Committee Member 1
Schmalzel, John
Committee Member 2
Thayasivam, Umashanger
Keywords
affine transform, biometrics, gmm-ubm, gsv, pls, speaker verification
Subject(s)
Automatic speech recognition; Speech processing systems
Disciplines
Electrical and Computer Engineering
Abstract
In the presence of environmental noise, speaker verification systems inevitably see a decrease in performance. This thesis proposes the use of two parallel classifiers with several enhancement methods in order to improve the performance of the speaker verification system when noisy speech signals are used for authentication. Both classifiers are shown to receive statistically significant performance gains when signal-to-noise ratio estimation, affine transforms, and score-level fusion of features are all applied. These enhancement methods are validated in a large range of test conditions, from perfectly clean speech all the way down to speech where the noise is equally as loud as the speaker. After each classifier has been tuned to their best configuration, they are also fused together in different ways. In the end, the performances of the two classifiers are compared to each other and to the performances of their fusions. The fusion method where the scores of the classifiers are added together is found to be the best method.
Recommended Citation
Edwards, Joshua Scott, "Using Gaussian Mixture Model and Partial Least Squares regression classifiers for robust speaker verification with various enhancement methods" (2017). Theses and Dissertations. 2371.
https://rdw.rowan.edu/etd/2371
