Multi-Channel Non-Negative Matrix Factorization for Overlapped Acoustic Event Detection

Abstract

In this paper, we propose two multi-channel extensions of non-negative matrix factorization (NMF) for acoustic event detection. The first method performs decision fusion on the activation matrices produced from independent single-channel sparse- NMF solutions. The second method is a novel extension of single-channel NMF, incorporating in its objective function a multi-channel reconstruction error and a multi-channel class sparsity term on the activation matrices produced. This class sparsity constraint is used to guarantee that the NMF solutions at a given time will contain only a few classes activated across all channels. This indirectly forces the channels to seek solutions on which they agree, thus increasing robustness. We evaluate the proposed methods on a multi-channel database of overlapping acoustic events and various background noises collected inside a smart office space. Both proposed methods outperform the single-channel baseline, with the second approach achieving a 15.4 % relative error reduction in terms of F -score.