Classifying Urban Sounds using Time Series Motifs

Abstract

The aim of this work is to describe a SAX-based Multiresolution Mo- tif Discovery approach to generate features for Urban Sound Classification. The idea is to discover relevant frequent motifs in the audio signals and use the dis- covered motifs and their frequency as characterizing attributes. We also describe different configurations of motif discovery for defining attributes. For classifi- cation we use a decision tree based algorithm, random forests and SVM. We compare the results obtained with the results using Mel-Frequency Cepstral Co- efficients (MFCC) for feature generation. MFCCs are commonly used in envi- ronmental sound analysis. Experiments were performed on the publicly available Urban Sound dataset. The results obtained suggest that the motif approach is able to identify discriminating features especially in the cases where MFCC failed.