Combined frequency-domain dereverberation and noise reduction technique for multi-microphone speech enhancement
Simon Doclo, Marc Moonen · 2001
In this paper a frequency-domain technique is described for estimating the acoustic transfer functions, when reverberated speech signals are corrupted by spatially coloured noise. This technique is an extension of the frequencydomain procedure of [1], which is only optimal in the case of spatially white noise. Using the estimated acoustic transfer functions, dereverberation can be performed with a matched filtering approach. Also a GSVD-based noise reduction technique is discussed, which computes an optimal (MMSE) estimate of the speech components in each frequency bin. It is shown that the optimal estimate of the clean dereverberated speech signal is obtained by dereverberating the optimal estimate of these speech components. Since the same decomposition is required for dereverberation and noise reduction, both procedures can easily be combined.