An experimental study of the robustness of multichannel inverse filtering systems to near-common zeros

Reverberation can significantly reduce the intelligibility and degrade naturalness of speech signals in hands-free communications. One method to achieve dereverberation is to perform identification and inverse filtering of multichannel acoustic systems. In this approach, we will show that the ‘energy’ of the components of a multichannel inverse filtering system play a key role in its robustness to channel noise and system identification errors, where the ‘energy’ of an MA filter is defined as the sum of the squared coefficients of it. The energy of the components of inverse filtering systems obtained from multiple-input/output inverse theorem (MINT) is usually high, resulting in non-robust inverse filtering systems. In this paper, the characteristics of acoustic channels which give rise to high energy in the components of inverse filtering systems are investigated. Specifically, the features of channel zeros that cause the components of inverse filtering systems to be of high energy are shown.