MINTFormer: A spatially aware channel equalizer

Reverberation is a process that distorts a wanted signal and impairs perceived speech quality. In the context of multichannel dereverberation, channel-based methods and beamforming are two common approaches. Channel-based methods such as the multiple input/output inverse theorem (MINT) can provide perfect dereverberation provided the exact acoustic impulse responses (AIRs) are known. However, they have been shown to be very sensitive to AIR estimation errors for which several modifications have consequently been proposed. Conversely, beamformers are significantly more robust but provide comparatively modest dereverberation. While the two approaches are conventionally considered independent, both can be formulated as a filter-and-sum operation with differing filter design criteria. We propose a unified framework, termed MINT-Forming, that exploits this similarity and introduces a mixing parameter to control the tradeoff between the potential performance of MINT and the robustness of beamforming. Empirical results show that the mixing parameter is a monotonic function of channel estimation error, whereby a MINT solution is preferred when channel estimation error is low.