Limitations of Fir Multi-microphone Speech Dereverberation in the Low-delay Case

In this paper multi-microphone dereverberation is considered under the constraint that no or little additional delay should be introduced by the FIR deconvolution filters. This is crucial for a number of applications such as hearing aids etc. Assuming that the acoustic impulse responses (AIRs) are known – e.g. by estimation, we determine the maximum degree of attainable dereverberation. Even though the AIRs are in general non-minimum phase, complete dereverberation can be accomplished in principle, using causal FIR filters of the same order as the AIRs, yielding no or only a little additional delay. We show that complete dereverberation with no or little delay will, however, reduce the SNR. For a given SNR gain and low delay, therefore, the achievable dereverberation is limited. We employ a time domain FIR multichannel Wiener filter with a delay constraint to find the MSE-sense optimal deconvolution filters. Dereverberation performance and SNR gain are demonstrated for typical AIRs with reverberation times of T60 ≈ 500ms and N = 4000 taps which have been measured in a conference room. Furthermore, we propose a new method utilizing a shaped desired total response, which is capable of selectively eliminating late reverberation while maintaining the SNR.