Beamformer design for nonstationary signals by means of interfrequency correlations

The wide-sense stationary assumption has been frequently employed in array processing, since it results in uncorrelated frequency bins and consequently major simplifications arise. However, unlike stationary signals, significant interfrequency correlations are observable in nonstationary signals like speech. Here, we drop the stationarity assumption and will show that taking interfrequency correlations into account leads to higher noise reduction. We develop a framework to design nonstationary beamformers similar to the stationary Frost beamformers. Based on the noise model, it can be used to design both fixed and adaptive beamformers. The nonstationary beamformer derived here is a set of time-varying filters and hence can be seen as a set of time-frequency or wavelet transform filters.