Second-Order Characteristic Window Approach to Blind Speech Signal Separation

In this paper, a second-order characteristic window approach to blind signal separation based on short-time stationarity of speech signals is proposed. This approach uses a new whitening algorithm via joint difference correlation matrix to remove the effect of colored noise for one thing, then uses characteristic window that the length is approximately equal to the fundamental period to divide the whitening observation data into series. In every different data frame, time-lag covariance matrices are computed and jointly approximately diagonalized to estimate the rotation parameters which could make the sources obtained. This approach avoids colored noises and separates speech signals only by means of second-order information. Its effectiveness is shown via computer simulation.