Using Multiresolution Time Series Motifs to Classify Urban Sounds

Abstract

The automatic classification of urban sounds is important for environmental monitoring. In this work we employ SAX-based Multiresolution Motif Discovery to generate features for Urban Sound Classification. Our approach consists in the discovery of relevant frequent motifs in the audio signals and use the frequency of discovered motifs as characterizing attributes. We explore and evaluate different configurations of motif discovery for defining attributes. In the automatic classification step we use a decision tree based algorithm, random forests and SVM. Results obtained are compared with the ones using Mel-Frequency Cepstral Coefficients (MFCC) as features. MFCCs are commonly used in environmental sound analysis, as well as in other sound classification tasks. Experiments were performed on the Urban Sound dataset, which is publicly available. Our results indicate that we can separate difficult pairs of classes (where MFCC fails) using the motif approach for feature construction.