Regularized NMF-based speech enhancement with spectral components modeled by gaussian mixtures

In this paper, we introduce a single channel speech enhancement algorithm based on regularized non-negative matrix factorization (NMF). In our proposed formulation, the log-likelihood function (LLF) of the magnitude spectral components, based on Gaussian mixture models (GMM) for both the speech and background noise signals, is included as a regularization term in the NMF cost function. By using this spectral type of regularization, we can incorporate the statistical properties of the signals during the estimation of both the basis and excitation martices in NMF model. Furthermore, borrowing from the expectation-maximization (EM) algorithm and to reduce the computational complexity of the NMF update, the LLF is replaced by its expected value. Experimental results of perceptual evaluation of speech quality (PESQ), source-to-distortion ratio (SDR) and source-to-interference ratio (SIR) show that the proposed speech enhancement algorithm provides better performance than the compared benchmark algorithms.