On the performance output cumulant matching based source separation methods in noise

Abstract

We study the performance of blind source separation methods based on output cumulant matching proposed by Cardoso and others. The output cumulant matching method has the advantage of exploiting invariance to yield performance and convergence properties that are insensitive to channel parameters in the low noise case. In moderate SNR however, the channel parameters do affect the performance via a modified set of parameters involving the noise covariance. A performance study is presented for this case for these algorithms. It is shown that enhancements to overall signal recovery is possible by combining noise suppression techniques with the cumulant based approaches.