| With the rapid development of sound source localization technology,sound source localization technology based on microphone array has been widely used in many aspects,such as speech enhancement,speech noise reduction,military detection,and service robot.However,many interference factors are inevitable,which seriously affects the localization accuracy of the sound source,resulting in the inability to locate.Especially in indoor environments,there are many problems such as noise and reverberation.Therefore,enhancing the robustness of the sound source localization system and improving the accuracy of the sound source localization are difficulties of the sound source localization technology.This paper deals with sound source localization from the perspective of machine learning.For some practical situations,the sound source localization is limited to some predefined areas,and the existing localization technology is in an unstructured indoor environment.Because of the problems of insufficient localization accuracy and lacking of adaptability,a research scheme of indoor sound source localization based on convolutional neural network is proposed.The sound source signal collected by the microphone is converted into a spectrogram to construct a localization data set,and put it into the convolutional neural network and training to realize the localization of the indoor single sound source.Then,we used Tensorboard to visualize the result of training and testing of the convolutional neural network,making the training process of the convolutional neural network more intuitive.Finally,it is compared with KNN(K-Nearest Neighbor),BP(Back Propagation)neural network and SVM(Support Vector Machine).The simulation results verify the effectiveness of the proposed method.On this basis,in order to solve the shortcomings of deep neural network,such as complex network structure and time-consuming,a sound source localization method based on BLS is proposed.The method firstly takes the characteristics of the collected data as the characteristic nodes of the network.Then,all mapped features and the enhancement nodes of randomly generated weight are directly connected to the output,and the corresponding output coefficients are obtained by pseudo-inverse.Finally,using the trained network model of the width learning system to predict the test points and determine the area to which the test points belong.The experimental results show that the method of the BLS maintains a certain precision in the sound source localization and reducing the time-consuming. |
