抄録
Blind signal separation based on statistic independence has been studied extensively. Particularly difficult is seperation of convolutive mixture. This can be solved approximately by a method with an FIR filter which is a cascade connection of many delay elements. It is, however, more desirable if the signal can be deconvoluted with an IIR filter (having feedback loop) of a modest order. This paper proposes such a demixer by projecting the conventional learning law based on the FIR approximation toward the parameter space of IIR filters. Numerical simulation is carried out to evaluate the proposed method.
