The Study Of Speaker Recognition Based On Principal Component Analysis And Linear Discriminant Analysis

With the development of Information Technology, the secure authentication becomes more important than ever. Biometrics is a new identification approach and it shows advantages in this area. Speaker recognition, which identifies or verifies people by their voice, is regarded as the most natural and convenient one among the methods of biometrics. However, there are still many problems when we want to apply speaker recognition to real applications. One is the long computational time of training a speaker model or test an utterance, which makes Real-time implementation very hard and expensive. In this thesis, based on the study of the recent advancements and main points of speaker recognition, we tried to solve the problem by using two methods of Multivariate Statistical Analysis: Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA).The first issue is the use of the character of optimal dimension reduction of PCA in speaker recognition. In a speaker identification system, training speaker models is computationally expensive, especially when the dimension of feature vectors is large. PCA method is an optimal linear dimension reduction technique in mean-square sense, which can reduce the computational overhead of the subsequent processing stages. In this thesis, two new speaker identification systems were proposed using PCA. Experiments were conducted to investigate the properties of de-correlation and optimal dimension reduction of PCA. The robust ability of PCA transform was also examined. Some promising results are found on these experiments.The second issue is the use of discrimination transform of LDA in speaker recognition. PCA seeks directions efficient for representation, while LDA seeks directions efficient for discrimination. The goal of LDA is to maximize the between-class measure while minimizing the within-class measure, so after LDA transform, the new feature vectors will be more discriminant. We proposed a new speaker identification system using LDA. Both on clean corpus and noised corpus, it achieves good recognition performance. And under some circumstance, it is even better than the systems using PCA.The third issue is the research of the classification ability of PCA method and the use of PCA classifiers in speaker recognition. Many different classification approaches have been developed for speaker recognition and they do achieve good performances. But their complexity makes the recognition time consuming. Based on the definition of PCA, it essentially owns classification ability. A novel PCA classifier called Principal Component Space (PCS) was proposed in this work. Together with other PCA classifier, it forms a hybrid classifier. All of these classifiers were applied to speaker recognition. The experimental works show promising results.This work is supported by National Natural Science Foundation of P. R. China (No.60273059), National High Technology Research & Development Programme (863) of P. R. China (No.200lAA4180), Zhejiang Provincial Natural Science Foundation for Young Scientist of P. R. China (No.RC01058), National Doctoral Subject Foundation of P. R. China(20020335025), Zhejiang Natural Science Foundation (M603229) and Zhejiang Provincial Education Office Foundation (20020721).