Sound Source Localization Based On Binaural Auditory Time Delay Estimation

Sound source localization technology plays an important role in human-computer interaction,smart conference,hearing aids,etc.The presence of reverberation and noise in the indoor environment will affect the sound source positioning results and cause some error.Therefore,it is of great significance to study how to improve the localization accuracy of sound source in reverberation and noise environment.First of all,the thesis makes a theoretical study and analysis of the sound source localization algorithm based on binaural signals,focusing on the sound source localization principle of binaural signals,analyzing the binaural signal model and determining the overall program of binaural sound sources localization.Secondly,the thesis studies the related principle of the commonly used time delay estimation method.For the same processing of each frequency component of the sound source signal,the problem of positioning error caused by the different influence of reverberation on different frequency sounds can not be solved.A generalized cross-correlation time delay estimation algorithm based on frequency division is proposed.The binaural signal is divided into subband signals by Gammatone filter.Minimum phase decomposition is carried out in each subband signal for dereverberation.The cross-correlation of each subband signal is calculated after the signal inverses the cepstral domain to the time domain.The interaural time difference corresponding to the maximum cross-correlation value is obtained by summing the cross-correlation function of each subband,and the azimuth angle is obtained through the geometric model.The experiment proves that this algorithm can fully dereverberate the various frequency components of the speech signal and effectively reduce the positioning error caused by reverberation in the multi-frequency sound source.Then,the azimuth estimation method is researched deeply.Aiming at the problem that the performance of sound source localization is degraded due to reverberation and noise,a sound source localization algorithm based on frequency division and Deep Neural Networks(DNNs)is proposed.The algorithm uses DNNs to map the relationship between the binaural features and the azimuth angle.Through the DNNs learning and testing process to improve the accuracy of sound source localization,andadding different SNR noise during the DNN learning process to improve the algorithm's anti-noise performance.Firstly,the binaural signals are constructed by the head-related response function in the CIPIC database,then the binaural signals are divided by the Gammatone filter,the frequency-divided sub-band signals are cepstrum-dereverberated,After transforming to the time domain,the cross correlation function(GCCF)is calculated for each subband cross correlation.In this paper,binaural speech localization is regarded as a multi-classification problem.GCCF and Interaural Level Differences(ILDs)are used as localization features and input to DNNs with softmax regression structure at the top.Outputs are probabilities that the sound source will be in each azimuth,take the maximum probability azimuth as the sound source position.Experiments show that the algorithm effectively helps improving the positioning accuracy in in reverberation and noise environment.Finally,binaural sound source localization system is constructed,experimentally tested and analyzed.Experimental results indicate that the localization algorithm based on frequency division and DNNs has better localization accuracy than other sound source localization algorithms based on time delays estimation.Experimental results are in agreement with the theoretical calculation trend.When the reverberation time is 0.3s and 0.5s,the algorithm's accurate estimation rates are 89.9% and 87.7%,respectively.It can effectively solve the influence of reverberation on sound source localization in the engineering environment.When the reverberation time is 0.89 s and the SNR is 20 dB,10dB and 0dB respectively,the accurate estimation of sound source localization rate is93.7%,80.2% and 84.7% respectively;When the reverberation time is 0.89 s and the SNR is 20 dB,10dB and 0dB respectively,the accurate estimation of sound source localization rate is 88.56%,84.46% and 70.06% respectively,it can achieve better localization effect under noise and reverberation environment.