The Research On Sound Source Location Of Microphone Array Based On Machine Learning

Sound source localization technology based on microphone array is an important research direction in the field of array signal processing.It is widely used in mobile phone,video conferencing system and military field.Because of the interference of noise,the localization performance of sound source will be degraded.In recent years,the technology of sound source localization based on machine learning has attracted more attention.Compared with the traditional sound source localization algorithm,this method not only has stronger robustness,but also is still effective when the microphone can not receive direct sound.This paper studies and does the following work in depth:1.The research status and application background of sound source localization technology based on microphone array are deeply studied.The speech signal model of microphone array is established,and the speech signal is preprocessed.2.Aiming at the performance degradation of DOA estimation in MUSIC normalization algorithm at low signal-to-noise ratio,a MUSIC-DOA algorithm based on controllable power response of support vector machine is proposed.Firstly,fast Fourier transform is used to estimate DOA of broadband signals,then MUSIC algorithm is used to estimate DOA.Finally,SVM is used to classify the DOA estimation results of each subband signal.After choosing the subband signal which is more accurate after classification,DOA estimation of broadband signal is obtained,which effectively solves the problem of low positioning accuracy of MUSIC normalization algorithm at low SNR.3.In view of the low localization performance of multi-layer perceptron sound source localization algorithm based on root mean square in low signal-to-noise ratio and other harsh environments,a multi-layer perceptron sound source localization method based on GCC-PHAT is proposed.Firstly,the received microphone signal is processed by frame windowing and the cross-correlation function of the signal is obtained.Then it is input into the multi-layer perceptron as a feature to get the azimuth coordinates of the sound source,which improves the accuracy of location.4.Aiming at the problem of poor stability and low positioning accuracy of Direction of Arrival estimation method based on traditional neural network,a DOA estimation method based on Local Weighted Long Short Term Memory Neural Network is proposed.Firstly,the triangular matrix over the array covariance matrix is used as the DOA estimation feature,and the feature is trained in LWLSTM.Then,the weights of the back propagation of LSTM neural network are updated by the kernel function and the objective function is introduced into the local weighted regression to obtain the DOA estimation result.The simulation results show that LWLSTM has better fitting ability than the traditional neural network method,and the DOA estimation accuracy based on LWLSTM neural network is higher.5.This paper analyses the algorithm of sound source localization proposed before,and implements the algorithm of Chapter 3,Chapter 4 and Chapter 5 in speech localization system.The experimental results show that the proposed speech location algorithm has high positioning accuracy and can meet the actual positioning needs.Finally,the work done in this paper is summarized and the possible improvements are prospected.