Research On Acoustic Azimuth Estimation Algorithm Using Sparse Bayesian Learning

Acoustic azimuth estimation is a key technology for audio signal processing appli-cation such as humanoid robot,tele-conference,acoustic fault diagnose and sonar.The target of acoustic source localization is to obtain the bearing angles of the sources using the microphone array data.According the Rayleigh rule,the resolution performance of traditional methods such as beamforming is limited by the array aperture.Thus,for some applications such as humanoid robot,tele-conference and acoustic fault diagnose with a small platform size,the traditional approaches suffer in a scenario with multiple acoustic source simultaneously exist.Although some methods such as minimum variance distor-tionless response(MVDR)and estimation of signal parameters via rotational invariance technique(ESPRIT)offer high resolution performance,they are sensitive to reverberation and calibration errors.Robust MVDR and ESPRIT algorithm are also been studied in the presence of array errors.However,these methods rely on asymptotic properties.To im-prove the resolution performance of acoustic azimuth estimation,the SBL based approach is studied in this paper.The SBL provide a high resolution performance by utilizing the sparsity of source signal and the recovery performance remains when the number of snapshots is small.After assigning sparse priors,the solutions can be obtained by maximizing the posterior and different sparse priors results in different algorithms.The main contribution of my work is summarized as follow:To fully utilize the sparsity nature of sound sources in the spatial domain,the SBL framework is studied and an hierarchical Bayesian framework using adaptive LASSO priors(a LASSO-SBL)is proposed.Using the proposed hierarchical framework,in-dependent sparse priors are assigned to the elements in the signal,resulting in a good recovery performance.Moreover,a conjugate property of generalized Gaussian distribu-tion is discovered and a new hierarchical framework using generalized Gaussian priors is proposed.Actually,most of state-of-the-art SBL methods are special cases of the proposed framework.Besides,the proposed method can approximate the7)₀norm sparse signal recovery in the perspective of Bayesian inference.Besides,as the choice of sparse priors is related with signal-noise ratio(SNR),the simulation is conducted to show the relationship between the sparse priors and SNR.The simulation results show that the proposed a LASSO-SBL algorithm outperform other methods.To improve the accuracy performance of acoustic azimuth estimation,a hierarchi-cal Bayesian framework is built by assigning each column vector independent adaptive LASSO priors.In this way,the spatial sparsity nature of acoustic sources is fully en-couraged.In Bayesian inference of multi-snapshots measurements signal models,the operations of matrix multiplication and matrix inverse are required for updating the mean and variance of signal,resulting high computation load.To deal with this problem,a space alternating variational estimation(SAVE)based SBL method is proposed and the computation load is released by updating each column vector independently.As the proposed SBL based method process each frequency bins independently,false estimations maybe exist in one or multiple frequency bins because of noise and initial value setup,resulting in pseudo peak.To solve this problem,a Gaussian mixture model based acoustic azimuth estimation algorithm is proposed by utilizing the relationship between measurements and recovered signals of all frequency bins.In this way,the effect of false estimations is suppressed,resulting an accurate estimation performance.The azimuth estimation performance of the proposed method is first verified using real data recorded by uniform circular array and near spherical array,separately.First,an acoustic azimuth estimation using uniform circular array is built and the performance of the proposed algorithm is test.Second,the performance of the proposed algorithm is test using a near spherical array data from LOCATA database.Experiment results show that the proposed method achieve higher resolution and accuracy performance.Besides the false estimation rate is lower than the state-of-the-art methods.