Eigenvalue filtering method for microphone array denoising

Cross-Spectral Matrix (CSM), as a second-order statistic, is the critical input of an acoustic imaging algorithm. The mechanism of cross-spectral matrix eigenvalue filtering is investigated to enhance the performance of the microphone array. Two eigenvalue filtering methods are proposed: (1) The SURE-Shrinkage (Stein's Unbiased Risk Estimation) method of the cross-spectral matrix of sound sources; (2) The Opt-Shrinkage method, which is used to improve further the estimation results of EYM (Eckart-Young-Mirsky). Then, the simulation is carried out with the parameters of 3000 snapshots, 0 dB Signal-to-Noise Ratio (SNR) and 100 microphones. The diagonal error of sound source CSM and de-noised CSM is used to compare the effectiveness of SURE and Opt-Shrinkage with the traditional Multi Signal Classification (MUSIC) method. The diagonal error of the MUSIC method is 74.15%, while the SUREShrinkage is 41.97%, and the Opt-Shrinkage is 20.62%. From the simulation results, under the condition of less than 40 sound sources, the de-noising results of the SURE-Shrinkage are better than the MUSIC method and Opt-Shrinkage method; after the number of sound sources exceeds 42, the Opt-Shrinkage perform better. By changing the number of snapshots, the number of sound sources and SNR, the results of different eigenvalue filtering methods is compared. At last, in the experiment of sound source localization under three sources, 60 microphones and -5 dB SNR condition, both SURE-Shrinkage and Opt-Shrinkage have a better de-noising performance compared with the MUSIC method. The research shows that better de-noising results can be obtained by further processing the eigenvalues.