Automatic recognition of urban soundscenes

Abstract

In this paper we propose a novel architecture for environmental sound classification. In the first section we introduce the reader to the current work in this research field. Subsequently, we explore the usage of Mel frequency cepstral coefficients (MFCCs) and MPEG7 audio features in combination with a classification method based on Gaussian mixture models (GMMs). We provide details concerning the feature extraction process as well as the recognition stage of the proposed methodology. The performance of this implementation is evaluated by setting up experimental tests in six different categories of environmental sounds (aircraft, motorcycle, car, crowd, thunder, train). The proposed method is fast because it does not require high computational resources covering therefore the needs of a real time application. © 2008 Springer-Verlag Berlin Heidelberg.