Confirm The Method Of Classification Feature Maps And Svm-based Speaker

With the development of communication and computer technology, more and more people communicate with each other by telephone, cellular phone and so on, and telephone speech plays an increasing important role in people's life. There are many advantages for using telephone speech to verify the identity of the speaker, and this technique has a wide application foreground in the fields of finance, business, public security, and military affairs and so on. Especially, the text-independent speaker verification can work without the text information, which has a better application scope and practicability.The Support Vector Machine (SVM) is a discriminative approach that seems well suited to speaker verification, and act as a hotspot of pattern recognition. When cepstral features, such as MFCC, are used for text-independent speaker verification, lots of speech is needed. So, as a modeling technique in text-independent speaker verification, SVM has much difficulty in handling a large quantity of training data.In this thesis, in order to solve the problem existing in text-independent speaker verification while using SVM, we develop the approach of feature mapping which utilizing the feature space classifying and GMM, and we also discuss the SVM modeling method. The main contents of study involved in this thesis are as follows:1) In order to solve the problem of the large quantity of training data, the approach of extracting the speaker feature vector by the feature mapping method based on GMM-UBM structure for SVM speaker verification is discussed. The GMM-UBM based feature mapping can realize the data condensation and extract the speaker feature vector. In this thesis, we also discuss how the mixture of the UBM affects the system performance and analyzes the modeling method in NIST's SRE multi-side task.2) An approach of speaker verification based on GMM-UBM structure feature mapping and SVM is proposed. The CGMM-UBM feature mapping, the data of different speakers became more suitable to classify and the GMMs are trained more exactly. In this thesis, we will show the analysis for VQ based classification method and pitch based classification method, while different sub-space numbers and different GMM mixtures are used. The experiments show that, by the effects of classification feature mapping and fusing the sub-system scores, the CGMM-UBM-SVM system enhances the EER performance 17.2% relative compared to the GMM-UBM system, and 7.6% relative compared to the GMM-UBM-SVM system.