Frequency-domain multichannel signal enhancement: Minimum variance vs. minimum correlation

In this paper we compare two optimization criteria that can be used in multichannel signal enhancement. The first criterion, Minimization of the output Variance (MV), is related to conventional adaptive beamforming whereas the second one, minimization of the output cross-correlation (or Symmetric Adaptve Decorrelation - SAD), is related to blind source separation techniques. We develop an implementation-independent comparison framework, where the quality of the MV adaption control is identified with the power of the target. We account for the correlation of the source of finite length. Our analysis shows that SAD is attractive in cases where target and interferer have about the same power level at each sensor. We evidence the fact that SAD requires longer signals, not only to maintain a low source correlation coefficient, but also because SAD learning rules suffer from larger estimation errors.