Research On Methods Of Multiple Sound Sources Localization Using Differential Microphone Arrays

The technology of sound source localization using microphone arrays has been widely researched in speech signal processing.Compared to single sound source,multiple sound sources azimuth estimation is more challenging.In general,the performance of conventional sound source localization approaches with small-sized microphone arrays declines seriously and even cannot meet the actual requirement.Therefore,it is the key problem to be solved that how to accurately estimate the multi-source azimuth with small-sized microphone arrays.Differential microphone arrays provide an important technical way for sound source localization with small-sized arrays.We study multiple sound sources azimuth estimation methods based on differential microphone arrays in noisy and reverberant environment,the specific work and contribution is as follows:Firstly,we analyze two exiting multiple sound sources azimuth estimation methods based on differential microphone arrays,including histogram method and fuzzy cluster method.And then the problem of those two methods is revealed,i.e.,the histogram method is sensitive to noise while the fuzzy cluster method has poor robustness to reverberation.Secondly,we propose a localization approach for multiple sound sources via an expectation maximization algorithm using differential microphone arrays.First of all,this method is to estimate the parameters of Gaussian mixture model for time-frequency instantaneous direction estimation through expectation maximization,and then to find the direction estimation of each sound source via time-frequency separation.In order to overcome the weakness of existing time-frequency separation techniques,i.e.,the hard and soft separation methods,an improved time-frequency separation method,which combines the advantages of both the hard and soft separation methods,is also proposed.The improved time-frequency separation method is shown to be less sensitive to noise and reverberation.Simulation and experimental results demonstrate that the proposed localization approach is superior to its existing counterparts in terms of localization accuracy and robustness characteristics.Thirdly,we also propose a multiple sound sources azimuth estimation method using selected reliable time-frequency points and clustering.This method primarily selects the reliable time-frequency points according to the local variance of instantaneous azimuth estimation,and then we can secondarily filtrate the results to obtain the final reliable time-frequency points using the redundant information of the differential microphone arrays.Furthermore,the azimuth of each sound source can be estimated via kmeans clustering for the instantaneous azimuth estimation of secondarily selected time-frequency point.The results of numerical simulations and real experiments both demonstrate the obviously superiority and robustness of the proposed method in noisy and reverberant environment.