EVD-based multi-channel dereverberation of a moving speaker using different RETF estimation methods
Ina Kodrasi, Simon Doclo · 2017
The multi-channel Wiener filter (MWF) for dereverberation relies on estimating the late reverberant power spectral density (PSD) and the relative early transfer functions (RETFs) of the target signal from a reference microphone to all microphones. State-of-the-art multichannel late reverberant PSD estimators also require an estimate of the RETFs, which may be difficult to estimate accurately, particularly in highly reverberant and noisy scenarios. Recently we proposed a more advantageous late reverberant PSD estimator based on an eigenvalue decomposition (EVD) which does not require knowledge of the RETFs, thereby avoiding the propagation of any RETF estimation errors into the PSD estimate. However, the performance of the proposed EVD-based estimator was analyzed by using it in an MWF with simulated RETF estimation errors for a fixed speaker position in noiseless scenarios. In this paper the EVD-based estimator is combined with several practical RETF estimation methods, i.e., the covariance whitening, covariance subtraction, and least-squares methods. The performance of the MWF using the EVD-based estimator and the considered RETF estimation methods is then investigated for a fixed and a moving speaker in different noiseless and noisy scenarios. Experimental results show that while combining the EVD-based estimator with any of the considered RETF estimation methods yields a high performance, in noiseless scenarios the covariance whitening and subtraction methods result in the best performance, whereas in noisy scenarios the least-squares method results in the best performance.