Posteriori regularization based non-negative matrix factorization approach for speech enhancement

Abstract

The paper proposes, a speech enhancement method for reducing additive Gaussian noise using iterative posterior regularized Non-negative matrix factorization (NMF). Here, regularization for NMF criterion is obtained by assuming the prior distribution of the Discrete Fourier Transform (DFT) spectral magnitudes of speech follows Nakagami, Weibull distribution and DFT spectral magnitudes of coefficients follows as Rayleigh distribution. In this paper, different prior distributions, Nakagami, Weibull and Rayleigh are used and the estimates of distribution statistics are changed adaptively to provide regularization. The results for different priors are compared using different objective performance measures Perceptual Evaluation of Speech Quality (PESQ) and Signal to Distortion Ratio (SDR).