A Speech Enhancement Method Employing Sparse Representation of Power Spectral Density ⋆

A speech enhancement method employing sparse reconstruction of the power spectral density is proposed. The overcomplete dictionary of the power spectral density is learned by approximation K-singular value decomposition algorithm with non negative constraint. The power spectral density of clean speech signal is reconstructed by least angle regression method with a norm termination rule, and the estimation of clean speech signal in the short-time Fourier transform domain is obtained by using signal subspace approach on the basis of short-time spectral amplitude. The experimental results show that the proposed method can reconstruct structured speech signal and suppress unstructured noise significantly even in low SNR conditions.