Convolutive blind signal separation via polynomial matrix generalised eigenvalue decomposition

摘要

An extension of the generalised eigenvalue decomposition (GEVD) to polynomial matrices, that is, a polynomial GEVD is proposed. A method for its application to convolutive blind signal separation is then introduced. The author shows that the source signals can be estimated using two related, but different, `target' polynomial matrices. These polynomial matrices are parahermitian matrices, corresponding to two different signal time intervals, which capture the non-stationarity of the sources. The validity of our method in separating the sources from their convolutive mixtures is demonstrated with computer simulations.