Research Of Multi-RVM Based On Locality Preserving Kernel In Speaker Recognition

With the rapid development of modern social economy and scientific technology, people's life fields become much wider. Then the identification technology appears to be more and more important in every aspect of our social life and economic activities. Because the traditional identification technology, such as passwords, smart cards, dynamic key and so on, can not meet the needs of the public for the geneogenous security defects, researchers have turned to the biometric-based personal identification technology with uniqueness and permanence property.One of the biometric identification—Speaker Recognition (SR) technology obtains broad applications in virtue of the special convenience, accuracy and economy.Although SR research has gone through many years, the technology still can not meet the increasing requirements, and many performances continue to be improved. In this thesis we focus on Text-Independent identification of small sample speech corpus. The main contributions of the work are summarized as follows:(1) Presented Speaker Recognition process. This paper introduces the preprocessing, feature parameter extracting, and model building of SR.(2) Applied Relevance Vector Machine (RVM) in SR, which uses a priori probability distribution and is highly sparse, is agilely applied to SR, especially in real-time occasion. In the training procedure, RVM adopts the fast marginal likelihood maximization algorithm to avoid operating a large matrix inverse and simplifying the calculation.(3) Introduced the Intra-Class Similarity in kernel function. Taking account of the complex structure of the speech features, Intra-Class Similarity is introduced into Gaussian Kernel Function to keep data local manifold and it will not change the kernel structure. The new RVM based on Locality Preserving Kernel used in SR and it improved the accuracy.(4) Proposed a real Multi-RVM (MRVM). The MRVM, which adopts Locality Preserving Kernel, becomes more visual and has better performance.(5) Biometric identification is more reliable and secure than the traditional technology. So in Speaker Recognition access control system, we used MRVM based on locality preserving kernel, which can deal with the digital signal quickly and control the access control system accurately.Finally, a simple summary is made and some future works are presented.