Noise Robust Speaker Verification using SVM based GMM Supervector
No Thumbnail Available
Date
2012-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
This thesis shows research performed on the topic of text-independent speaker verification using support vector machine (SVM) based Gaussian mixture model (GMM) supervector classifier in the presence of environmental noise conditions. The basis is to use model based approaches to make the verification task noise robust. In this report, two model based approaches are applied. In the first approach, speaker model adaptation is implemented based on the noise condition observed during verification. In the second approach, multi-condition training is adopted in which multiple speaker models are trained using multiple noisy speech samples. In both approaches a range of signal to noise ratios are considered. The system is implemented in MATLAB. In addition, LibSVM toolbox is used as SVM trainer.
The report gives an overview of speaker verification system. Methods and algorithms for speaker modeling are explained. The performance of the system in clean and environmental noise conditions is tested for both target trials and impostor trials. For clean training conditions, on average, results show that using GMM-SVM classifier improves the classification accuracy of Gaussian mixture model universal background model (GMM-UBM) classifier in terms of EER by 4.39% and 2.77% for babble and white noise, respectively. For test utterance corrupted by additive noise, test results show that multi-condition based noise compensation approach achieve from 1.34 % to 4.8 % improvement for GMM-UBM classifier and from 0.5 % to 3.08 % improvement for GMM-SVM classifier when compared with the corresponding standard classifiers.
Experiments are performed using utterances of speakers from the free audio book collection found on the online digital library LibriVox. Results obtained are presented in terms of detection error tradeoff (DET) curve and equal error rate (EER).
Key Words: Text-Independent Speaker Verification, SVM, GMM supervector, GMM-UBM, multi-condition training, LibSVM, Detection Error Tradeoff, Equal Error Rate
Description
Keywords
Text-Independent Speaker Verification, SVM, GMM supervector, GMM-UBM, Multi-condition training, LibSVM, Detection Error Tradeoff, Equal Error Rate