Application Research Of Relevance Vector Machine In Speech Recognition

Speech recognition technology is one of the key technologies in an information societytowards the directions of intelligence and automation, which has important researchsignificance and practical value. Through nearly five decades of painstaking exploration andresearch, speech recognition technology has tremendous growth, and some of the more maturetechnologies have been gradually applied to daily life. Generally speaking, however, thedifficulties of speech recognition in research and practical aspects are still relatively large.Relevance Vector Machine(RVM) is a machine learning algorithm based on SparseBayesian framework,has a good classification and generalization in dealing with highdimensional,non-linear and small-size problems. Compared with Support Vector Machine(SVM),the kernel functions of RVM are not need to satisfy the Mercer condition,the model ismore sparse,and can get the probabilistic outputs.This thesis mainly studied the RelevanceVector Machine algorithms, and applied to speech recognition.This paper first introduced the basic theory of speech recognition in detail,analysed eachpart according to the structure of speech recognition system.And then studied the theory ofRVM,introduced the regression and classification process of RVM.Through simulationexperiment analysis,compared the classification ability of the Relevance Vector Machine andSupport Vector Machine,experiments show that Relevance Vector Machine(RVM) is moresparse than Support Vector Machine (SVM),and RVM has a faster testing speed.Then,studiedthe effects of different kernel function and parameters on the classification of RVM. Finally,simulation experiments are carried out on the platform of MATLAB, achieved pretreatment,endpoint detection, and characteristic parameters extraction,then the classifier constructed byone-to-all classification method for RVM, the recognition results comparing with hiddenmarkov model(HMM),shows that the application of the RVM is feasible in speechrecognition,and has good generalization ability.