Convolutive blind source separation using an iterative least-squares algorithm for non-orthogonal approximate joint diagonalization

摘要: In this paper, we present an approach of recovering signal waveforms speech sources from observed signals in noisy and reverberant environments. The is based on approximate joint diagonalization estimate to provide interference suppression source reduce echoes distortions separated signals. the proposed approach, mixing matrix estimated by minimizing constrained direct least-squares (LS) criterion model. Exclusively under condition where not full rank, it replaced a full-rank matrix. unmixing obtained setting frequency response composite mixing-unmixing filter identity cross-spectral density diagonal matrices are precisely indirect LS These operations fulfilled using alternating algorithm. correlation between interfrequency power ratios used prevent misalignment permutation Finally, compare BSS with number conventional methods environments both artificial actual conditions.