NMF-Based Spectral Analysis for Acoustic Event Classification Tasks

Abstract

In this paper, we propose a new front-end for Acoustic Event Classification tasks (AEC). First, we study the spectral contents of different acoustic events by applying Non-Negative Matrix Factorization (NMF) on their spectral magnitude and compare them with the structure of speech spectra. Second, from the findings of this study, we propose a new parameterization for AEC, which is an extension of the conventional Mel Frequency Cepstrum Coefficients (MFCC) and is based on the high pass filtering of acoustic event spectra. Also, the influence of different frequency scales on the classification rate of the whole system is studied. The evaluation of the proposed features for AEC shows that relative error reductions about 12% at segment level and about 11% at target event level with respect to the conventional MFCC are achieved.