Leveraging hierarchical relationship between sounds for few-shot sound event classification

Abstract

Machine learning research for sound event recognition has generally focused on audio classes with abundant data. In this work, we exploit hierarchical relationships between sound events in a few-shot learning setup to enable classification of a wider set of sound events, given just a few examples at inference. By iteratively calculating prototypes for each level according to a given hierarchy system, our network’s feature space is encouraged to mirror the predefined relationships between sound events. Compared to a non-hierarchical few-shot baseline, our method leads to a significant increase in classification accuracy and significant decrease in mistake severity on unseen classes. Our work also proposes a new audio label taxonomy with descriptive labels that reflect actual acoustic characteristics.