Density based subspace clustering of urban audio recordings

Abstract

Sound pollution is a common concern in urban environments. Many cities are therefore equipped with networks of microphones, which are merely used to identify loud areas and less loud areas. Additional information, other than the sound levels, is not derived. It is unknown which sound events are responsible for the recorded sound levels. This thesis aims to cluster sound events in an unlabelled audio recording, to distinguish which sound sources cause the recorded sound levels. In order to cluster sound events, audio features are required. As it is unknown which (possibly large number of) sound events to expect, many audio features are extracted from the data. This results in high-dimensional data which introduces several challenges for clustering. To overcome these challenges, it is assumed that relevant features are dependent on the sound event. Each sound event cluster is searched in a different subset of the full-dimensional space. This method, called subspace clustering, is evaluated. In order to evaluate a subspace clustering algorithm, evaluation techniques are used. As evaluation on an unlabelled dataset is difficult, a labelled audio dataset is used, namely the ESC-10 dataset. For unlabelled datasets, an internal metric is proposed. Still, evaluation of a subspace clustering algorithm on this labelled dataset is not straightforward. Because (subspace) clusters are not assigned to any class, a label search is derived. This label search assigns the best subspace clusters to a label. This enables a comparison between a subspace clustering algorithm and a full-dimensional base model, i.e. k-means clustering. The subspace clustering algorithm, SUBCLU, indicates considerable performance improvements when compared to k-means on the labelled ESC-10 dataset. The overall accuracy of k-means is 33% while SUBCLU achieves an accuracy of 47% on this labelled dataset. Although there are some limitations of the implemented methodology, this research demonstrates that subspace clustering is beneficial for clustering (unknown) sound events in audio recordings.