An iterative posterior NMF method for speech enhancement in the presence of additive Gaussian noise

In this paper, a speech enhancement method based on regularized non-negative matrix factorization (NMF) for non-stationary Gaussian noise is proposed. An iterative posterior NMF-based model of the magnitudes of the spectral components of speech and noise is implemented using prior distributions for the magnitudes in the transformed domain. Because their sample distributions fit gamma and Rayleigh densities well, we propose to adaptively estimate the statistics of these distributions that are sufficient to provide a natural regularization of the NMF criterion. The resulting method is shown to outperform other benchmark algorithms in terms of the signal-to-distortion ratio (SDR) and a perceptual evaluation of the speech quality (PESQ).