A General-Purpose Sound Taxonomy for the Classification of Heterogeneous Sound Collections

Abstract The task of sound classification is fundamental for decoding information embedded in the audio domain. Taxonomies are integral to this process by defining the way in which the audio domain is conceptualized. Existing sound taxonomies often feature complex hierarchies and can be very specific to a particular use case. Thus, navigating and describing audio content using such taxonomies becomes a difficult task, not only for computers but also for non-expert humans. In this paper, we introduce a user-centric methodology that integrates both theoretical and data-driven approaches to formulate a general-purpose and user-friendly taxonomy for sound classification, the Broad Sound Taxonomy (BST). The taxonomy comprises a two-level hierarchical structure with 5 broad top-level categories and 23 second-level categories capturing finer semantic distinctions. To inform the taxonomy's design process, we leverage the context of the Freesound website, a popular example of a large and heterogeneous sound collection with a substantial user base. We consider the needs of different user profiles including sound designers, musicians, and researchers, and we conduct a user-based evaluation to assess the usefulness of the taxonomy and users' understanding of its categories. Results suggest that the taxonomy is appropriate for classifying sounds of diverse nature and can be used effectively by participants with different levels of expertise. The proposed taxonomy has several applications in the fields of multimedia retrieval, machine listening, audio analysis and, more generally, computational audio.