Speaker Recognition Methods And Strategies

With the development of information technology, individual identification and information security become more and more important. Speaker recognition (SR) technique is one of biological technologies with which computer can automatically identify the individual based on human characteristics. Compared with traditional methods, the new method proposed in this paper is more convenient, safer, and not easy to be forgotten or replaced. Speaker recognition can be used under a lot of circumstances, such as the telephone electronic commerce, military wiretapping , information retrieval, and so on.In this paper, the development history of speaker recognition and the research status are outlined firstly, then the theory and system composition of SR system are studied and three primary methods including non-parameter model, parameter model and Neural Networks method are summarized. Every method has its own advantages and disadvantages. Because the HMM describes the statistical characteristics of speech signals successfully and can get the good recognition results, it is the most popular method at present. The GMM is often used in the text-independent SR.The attention of the paper is concentrated on the speaker feature extraction and speaker modeling which are the two key problems in speaker recognition. The most popular feature parameter sets are the LPC and MFCC, and the performance of the latter is a little better than the former. In addition, some features based long-term speech analyses are used. There are many methods that can be used to construct the speaker model. The paper mostly introduce some methods such as Template Matching, Vector Quantization, Hidden Markov Models and GaussianMixture Models, and the Template Matching method is used in the pitch period template matching and long-term spectrum template matching.In the experiment part, two recognition approaches we brought forward are simulated on MATLAB Platform. One is the speaker recognition approach based on the sub-band processing and VQ. Compared with the wide-band SR, the experiments show that sub-band approach can enhance the system recognition rate and robustness for narrow-band noise effectively, and when the number of sub-band is 16, the system has the best performance. The other approach is based on the multiple strategies and parameter model. Here vocal source and vocal tract information are integrated, and a joint measuring is established in the first-level decision, and then the candidate corpus can be reduced effectively. CHMM recognizer based on stochastic models is adopted in the second decision, by which the system recognition rate can be ensured. In order to raise the robustness and the practicability further, the mixed speech training method is used too. The experiment results show that the system based on the multiple strategies and sequence-order decision can not only ensure the recognition rate, but also raise the response speed virtually.At last, the thesis gives a summary of my work and points out some problems and ideas for further research.