Robust multichannel equalization for blind speech dereverberation

Abstract

Acoustic reverberation arises from the reflection of sound waves within an enclosed space. It is generally desirable in music reproduction but can be detrimental to speech-related applications. For the human listener, while the early reflections help to improve speech intelligibility, the late reflections have been shown to impair perceived speech quality. For speech processing technologies such as automatic speech recognizers, reverberation reduces accuracy and performance. Dereverberation is therefore an important research topic with interest driven by increasing availability of communication devices and consumer demand. One approach to dereverberation computes a set of equalizing filters that are used to perform the dereverberation processing, given multichannel inputs and estimates of the acoustic impulse responses (AIRs) between the source signal and microphones. However, estimation errors are inevitable in practice and therefore robust channel equalizers are required. This thesis aims to develop such robust algorithms in a manner that is desirable specifically for speech dereverberation. The framework of channel shortening is used, having been previously shown to give promising results. Subband approaches are also investigated to reduce the computational complexity and achieve finer control of dereverberation in separate frequency bands. A second approach to dereverberation steers the look direction of beamformers towards the source. Reverberant sounds from other directions are treated as noise and accordingly suppressed. The motivation behind beamformer design and channel equalization is similar and in this work, a unified framework termed MINTFormer is proposed. The aim is to combine the robustness of beamformers with the potentially perfect dereverberation ability that can be achieved by channel equalization approaches.