Considering non-stationarity for blind signal separation

We investigate the exploitation of non-stationarity for signal separation. A second-order decorrelation method is used to separate synthetic independent autoregressive signals that are made up of stationary blocks that have been convolutively mixed. We compare results obtained by not taking into account the non-stationarity with those that do. Under certain conditions, exploiting non-stationarity results in more robust separation. We present simulation results that vindicate this fact. In addition, we apply the decorrelation method to real microphone signals, to see how exploiting non-stationarity affects separation quality.