Research On Sound Source Localization Method Based On Generalized Cross-correlation And Convolutional Neural Network

With the continuous development of intelligent devices,localization of device fault detection,intelligent robots,and video conferencing are estimated through sound signals.In the field of sound source localization,many traditional sound source localization methods are proposed,such as generalized cross correlation based time delay estimation methods,steered-response power(SPR),multiple signal classification(MUSIC)and deconvolution approach for the mapping of acoustic sources(DAMAS),etc.The traditional sound source localization method does not solve the localization accuracy of environmental noise,reverberation,and multiple sound source events,which makes the difficulty of sound source localization in practical applications,increases the requirements for hardware equipment,and increases the system cost.Therefore,in order to solve the problems of low localization accuracy,multi-sound source event azimuth detection,and reduction of calculation,this paper proposes a research of sound source localization methods based on generalized cross-correlation and convolutional neural networks.The received sound source signal not only contains the sound source event information of each category,but also the location information of the sound source event.With the development of signal processing technology,sound signal pre-processing,feature extraction and network structure constitute a sound source signal localization system.For the sound source localization system,the feature extraction part is very important,affecting the calculation amount,robust performance and positioning accuracy of the entire system.The noise and reverberation have similar peaks in the characteristics of the generalized cross-correlation function for the same sound source at the same location,and the robust performance is good.For the localization system in this paper,it is mainly to extract generalized cross-correlation(GCC)features.In order to solve the problem that the traditional sound source localization method is not accurate,the robust performance is poor,and the localization accuracy is limited by the microphone array.In this paper,the characteristics of generalized cross-correlation features are extracted,and the support vector machine(SVM)algorithm is used to analyze the impact on positioning performance and positioning accuracy under different length feature vectors,different signal-to-noise ratios,and different reverberation levels.For TAU Spatial Sound Events 2019 database,in order to solve multiple sound source event recognition and sound source position estimation,this paper proposes a method for multi sound source event recognition and sound source position estimation based on a dual channel convolutional neural network.Features of generalized cross-correlation function and Log-Mel spectrum feature are extracted by Mel filter bank.The dual-channel neural network is designed to include input layer,hidden layer and output layer network structure models,and the parameters are optimized.By extracting feature vectors of different dimensions from the four-channel sound source signals,a dual-channel convolutional neural network structure with four different depths of 5 layers,9 layers,11 layers,and 13 layers was designed for experimental verification.Experimental results show that the 9-layer dual-channel convolutional neural network structure has higher accuracy of sound source event detection and sound source position estimation than the other three networks.