Subband Processing-based Approach For Multiple Speakers Localization Within Microphone Array

Speaker localization is a hotspot in acoustic signal processing field.It is widely used in military,civilian,and industrial other areas.Single speaker localization algorithm has made some progress.And a relatively complete compact speaker localization system has been build within the hardware systems But the existing multiple speakers localization algorithms fail to localize each speaker accurately.The reason is that the algorithms can not overcome the interference between these speech sources.In order to improve the localization performance of multiple speech sources in reverberation situations,this paper proposes the utilization of subband processing for the localization of multiple speakers.The simulation results verify the effectiveness of the algorithm,the main work of this paper includes the following studies:1.This paper introduces the basic features of microphone array speech signal and gives the reverberation voice model which used in speech source localization by IMAGE model;2.This paper analyzes and compares the three basic algorithms of speech source localization and focuses on the algorithm based on the time difference of arrival(TDOA).It is the basis for subsequent improvement algorithms in this paper;3.In order to improve the localization performance of multiple speech sources in reverberation situations,this paper considers the sparsity of speech signal in frequency domain and proposes the sub_weighted cross correlated estimator(sub_WCC),in which the conventional generalized cross correlation estimator(GCC)is modified by average magnitude difference function(AMDF)in filter banks.In this paper,the time delay value of each speech source is then calculated by fusing the sub-band signals.And the speaker positions are detected by a geometric algorithm using the rectangle microphone array architecture.The simulation results show that the method performs better accuracy in reverberant situations;4.Compared with the traditional localization algorithm,the accuracy of sub_WCC doesn't have obvious improvement.It can not achieve the desired level of practical application.And when the speaker is farther away from the microphone array,the accuracy and anti-reverberation of sub_WCC are rapidly decline.To overcome these problems,this paper proposes the sub_weighted smooth cross correlated estimator(sub_WSCC),in which the weighted cross power spectrum of WCC is smoothed by a smooth filter in each subband.The simulation results show that the method performs is significantly better than the sub_WCC in reverberant situations,can be used in the actual speech source location.