Abstract
Neural networks (NNs) have been widely applied in speech processing tasks, and, in particular, those employing microphone arrays. Nevertheless, most of the existing NN architectures can only deal with fixed and position-specific microphone arrays. In this paper, we present an NN architecture that can cope with microphone arrays on which no prior knowledge is presumed, and demonstrate its applicability on the speech dereverberation problem. To this end, our approach harnesses recent advances in the Deep Sets framework to design an architecture that enhances the reverberant log-spectrum. We provide a setup for training and testing such a network. Our experiments, using REVERB challenge datasets, show that the proposed position-agnostic setup performs comparably with the position-aware framework and sometimes slightly better, even with fewer microphones. In addition, it substantially improves performance over a single microphone architecture.
Abstract (translated)
URL
https://arxiv.org/abs/2010.11875
