Abstract: Because the performance of the blind source separation method based on generalized eigendecomposition for congruent pencils in the time domain is affected by the filter or delay, a blind source separation method based on generalized eigendecomposition for congruent pencils in the wavelet domain is proposed. Using the orthogonal property of wavelet, the non-gaussianity of the signals is increased and the difficulty in separating signals is decreased. Since the bi-orthogonal wavelet has the linear phase property and good approximation ability to signals, the generalized eigendecomposition for congruent pencils of wavelet coefficients is carried out, and the steady separation matrix is obtained. The algorithm not only retains the virtue of the time domain algorithm, but also can use two random numbers as the filter coefficients, and the time delay method used by it can separate the mixtures when the number of the sources is more than three. The simulation results of four voice signals verify the effectiveness of this method.

Key words: blind source separation, generalized eigendecomposition, wavelet transform, congruent pencils