Abstract: State of the art sound event classification relies in neural networks to learn the associations between class labels and audio recordings within a dataset. These datasets typically define an ontology to create a structure that relates these sound classes with more abstract super classes. Hence, the ontology serves as a source of domain knowledge representation of sounds. However, the ontology information is rarely considered, and specially under explored to model neural network architectures. We propose two ontology-based neural network architectures for sound event classification. We defined a framework to design simple network architectures that preserve an ontological structure. The networks are trained and evaluated using two of the most common sound event classification datasets. Results show an improvement in classification performance demonstrating the benefits of including the ontological information.
Keywords: sound event classification, neural networks, ontology
TL;DR: We present ontology-based neural network architectures for sound event classification.