A novel initialization approach for solving permutation ambiguity of frequency domain blind source separation

摘要

Due to the importance of speech signals in the recent years, many techniques have been proposed to solve the permutation ambiguity of blind source separation in frequency domain. Our goal in this paper is to present a new method based on initialization and navigation resources to solve the permutation ambiguity in blind source separation in the frequency domain. In this work, first the separation matrices are obtained from each frequency subband, into a diagonal matrix. After sorting the eigenvalues, the matrix will initialize to the next frequency subband as an initial value. Simulation results show that our separation method has better separation quality and has higher speed of convergence.