Dereverberation and Noise Reduction Based on PSD Estimation with Low Complexity

Abstract

In the far-field scene with noise and reverberation, the integrated sidelobe cancellation and linear prediction (ISCLP) method can simultaneously implement spatial filtering and deconvolution to effectively suppress additive noise and reverberation, but it has high complexity for calculating power spectral density (PSD). In order to reduce this complexity, the power-based PSD estimation method instead of the generalized eigenvalue decomposition (GEVD) is proposed in this paper to obtain eigenvalues and eigenvectors used for calculating PSD. Computational complexity is reduced to M times as compared with the GEVD by combining power-based method with Wielandt’s deflation which is used to solve the eigenvalues and the corresponding eigenvectors of correlation matrix of the observed signals. Experimental results show that the performance of dereverberation and noise reduction of the proposed method decreases slightly as compared with the GEVD-based ISCLP method.