Single-channel speech separation based on NLMSE quasi-KLT bases

In this paper, we propose a new single-channel speech separation algorithm by l₀ optimization using nonhomogeneous linear mean square estimation (NLMSE) quasi-KLT basis. First, we propose to construct NLMSE quasi-KLT bases from the single-channel mixture by diagonalizing the approximate autocorrelation matrices of speech sources, which are constructed by a linear combination of generated templates. Second, based on the NLMSE quasi-KLT bases, single-channel speech separation is performed by l₀ optimization. Finally, simulation results demonstrating the effective performance of single-channel separation of two speech sources using the NLMSE quasi-KLT bases are presented.