Research On Robust Binaural Localization Based On Neural Network

As an important front-end of speech signal processing system,sound source localization(SSL)technology has a wide applications,including video conferences,hearing aids,robotic hearing,etc.Currently,SSL algorithms mainly include microphone-array based methods and binaural source localization algorithm.Because the binaural source localization algorithms simulates the characteristics of human auditory perception,more accurate can be achieved.The conventional binaural SSL algorithms are mostly signal processing methods,and their performance significantly degrade in reverberation and noise environment.In this thesis,the robust binaural sound source localization algorithm based on Deep Neural Network using binaural features is studied.Two types of DNN are utlized: subband DNN and long short-term memory network.(1)The SSL algorithm based on subband DNN.This algorithm is develped from the the original DNN based SSL algorithm using full band binaural spatial cues.The proposed algorithm divedes sound signal into multiple subbands according to Gammatone filters.Subband binaural localization cues such as CCF and IID are modeled by DNN.Within each subband,SSL is considered as a multi-classification problem.Also,two methods are utilized to fuse the localization result of each sunband and obtain the final localizaition..Simulation tests show that the SSL algorithm based on subband DNN significantly improves the performance in noisy and reverberational envrionments.(2)The SSL algorithm based on LSTM.Based on the localization features correlation of the successive frames,LSTM network is used as a multi-classifier in the SSL problem.LSTM can predict the features using past features.So this algorithm regards the localization cues of the successive frames as successive nodes of the LSTM,and establishes the LSTM localization network with the softmax regression structure at the top layer.The IID and CCF are extracted from the entire frequency band.And the CCF is improved by weighting subband SNR estimation.The simulation tests in different correlated frame numbers show that the algorithm has a very high localization accuracy under very low SNR and high reverberation.