Speech Enhancement Using Nonnegative Matrix Factorization with Overlapping Group Sparsity Constraint
Haobin Wen, Lin Liang, Fei Liu · 2020
We consider speech enhancement techniques and probe into the group-sparsity peculiarity possessed by speech signals. Common structured sparsity penalties, e.g., block sparsity, incorporated by nonnegative matrix factorization (NMF) models to penalize predefined rows of coefficients, cannot sufficiently capture the grouping features within a row of activation coefficients. To circumvent this defect, a structured penalty known as overlapping group sparsity is introduced, which slides across the coefficients for group shrinkage. In this work, we propose an NMF variant in combination with the sparsity of overlapping group lasso (termed NMF-OGL), to adaptively promote the parsimony across small overlapped groups, as applied to enhancing single-channel speech mixture with different real-world background noise. Experimental verification indicates that the introduced penalty improves the denoising performance in comparisons to standard NMF and the overlapping group shrinkage model.