Research On Methods Of Multiple Speech Sources Localization In Room Reverberant Environment Using Spherical Harmonic Sparse Bayesian Learning

Spherical microphone arrays(SMAs)are able to analyze 3-D sound fields effectively,thus they have received extensive attention in recent years.The SMAs can be used for speech source localization in the element-space domain and the spherical harmonic(SH)domain,respectively.Compared with localization in the element-space domain,frequency-and angular-dependent components are decoupled in SH domain,and the steering matrix is independent of the position of the array element,which is easier to implement source localization in 3-D space.Therefore,the spherical harmonic source localization has been a particularly appealing kind of methods.Inspired by the technique of compressive sensing,sparse Bayesian learning(SBL)achieves higher accuracy,outperforming traditional methods.Thus the study of this thesis is based on SBL in SH domain.Compared to single speech source,multiple speech sources localization is more challenging.As the existing spherical harmonic methods degrade greatly for speech sources localization in room environment,where the sound reverberation belongs to more difficult convolutive noise,so this thesis will study on methods of multiple speech sources localization in room reverberant environment using spherical harmonic sparse Bayesian learning.The main work and contributions of this thesis are as follows:Firstly,the existing spherical harmonic methods are studied,including the spherical harmonic multiple signal classification(SH-MUSIC)and spherical harmonic sparse recovery(SH-SR)methods.Simulation experimental results reveal that the SH-MUSIC method performs poorly with a low spatial resolution while the SH-SR method has poor robustness to reverberation.Secondly,a spherical harmonic sparse Bayesian learning multiple speech sources localization method is proposed.First of all,this method focus various frequencies of a wideband signal into one specific frequency by frequency smoothing to reduce computational complexity and is less sensitive to noise and reverberation.And then obtain the locations of sources via temporal extension of multiple response model sparse Bayesian learning(TMSBL)in SH domain.In order to guarantee accurate sound localization,it is a viable way to refine the grid only around the regions where sound sources are around.Simulation and experimental results demonstrate that the proposed localization approach in SH domain is superior to its existing counterparts in terms of localization accuracy characteristic.Thirdly,In order to combat the sensitivity of existing sparse dictionary to high reverberation and noise,a dictionary learning method in SH domain is proposed.According to the theory of compressive sensing(CS)beamforming,this method primarily obtain an improved sparse dictionary by weighting the original one.Then combining the new dictionary,TMSBL in SH domain can detect the bearing of sources.Finally,2D histogram smoothing and grid refinement are put into the framework to achieve better performance.The results of numerical simulations and real experiments both demonstrate the obviously superiority of the proposed method in noisy and reverberant environment.