Sensitivity analysis of the multi-frame MVDR filter for single-microphone speech enhancement

Recently, a multi-frame minimum variance distortionless response (MFMVDR) filter for single-microphone noise reduction has been proposed, which exploits speech correlation across consecutive time frames. It has been shown that the MFMVDR filter achieves impressive results when the speech interframe correlation vector can be accurately estimated. In this paper, we analyze the influence of estimation errors for all required parameters, i.e., the speech interframe correlation vector and the undesired correlation matrix, on the performance of the MFMVDR filter. We compare the performance difference between oracle estimators and practically feasible blind estimators. Experimental results show that even small estimation errors substantially degrade the speech quality, where the most critical parameter is the speech interframe correlation vector.