A BLIND SOURCE SEPARATION ALGORITHM USING WEIGHTED TIME DELAYS

Bert-Uwe Kohler and Reinhold Orglmeister
E-mail: koehler@tubife1.ee.tu-berlin.de

A new algorithm for the blind separation of mixed non- white signals is proposed. After removing the second order cross-correlation between the signals for zero time delays (sphering), the signals are linearly transformed in order to reveal common components that are still existent within the signals. The linear transform is accomplished by ltering the sphered signals using a non-recursive lter which is equivalent to the summa- tion of weighted time delayed signal values. This leads to a non-zero cross-correlation that is used to eventu- ally separate the signals. In this paper it is shown how the lter coeÆcients are determined and how this prin- ciple is applied to blind source separation.