Research On Sound Source Localization Method Based On Support Vector Machine

With the rapid development of multimedia technology,microphone array technology has been widely used in sound source localization.But now some problems such as the exact rate is not high in the high-noisy and reverberant environment,high cost,large calculation exist in the most of positioning technology based on microphone array.To solve these problems,a new kind of sound source localization method based on SVM(Support Vector Machine)has been proposed.SVM is a machine learning method using structural risk minimization principle based on statistical learning theory,and the performance of SVM depends on the correct selection of related parameters.The work of the article includes mainly:1.The thesis contains the pre-processing work of the voice signal.Only after the pre-processing of the speech signal received from microphone has been operated can the accuracy be further improved on the basis of the improved algorithm.By translating the non-stationary and wideband speech signals into stationary and narrowband ones,processing of the subsequent localization algorithm is facilitated.2.At present,GCC-PHAT time delay estimation method of positioning method is used in most of the microphone array positioning system.This method has the advantage of strong anti-reverberation and small computational complexity.However,its anti-noise ability is weak.This thesis proposed a new sound source localization method based on SVM.By extracting the features of cross-correlation function and selecting appropriate parameters,SVM kernel function has been optimized.The proposed algorithm has increased the accuracy of sound source localization significantly in noisy and reverberant environments.3.In order to overcome the inherent shortcomings of SVM,this thesis proposes an idea of constructing multiple classifiers combination,and constructs and analyzes improved classifiers of Adaboost and SVM.By comparing the single SVM classifier model and the improved Adaboost and SVM combination classification,the accuracy and performance of the proposed model are further verified.