Abstract:Temporal correlation matrices of the convolutive-mixture signals possess the struture of tri-factor multiplication and so can be jointly diagonalized blockwise. Though cutting to pieces and then reprogramming of the 3-D matrix composed of a group of correlation matrices, an alternating and iterative approach is proposed to achieve the least-squares solutions and then to estimate the channel mixture matirx for realizing convolutive blind separation in time domain. Compared with traditional joint block-diagonalization algorithms, simulation results show that the proposed one has better and more stable separation performance irrespective of initial parameters. The dissimilarity index and Bark sepectral distortion are improved by 4.35 dB and 0.22 respectively.