Abstract
The organization of sound collections is a common task with applications in sound browsing, analysis and retrieval. The growing availability of online audio databases has brought attention to the need for additional structure and improved characterization within heterogeneous sound collections. Our research addresses this by creating hierarchical taxonomies for broad audio content, prioritizing user-driven decisions to aid user navigation and live music creation. We employ deep learning techniques for automated audio classification, integrating audio and text-based metadata through a cross-modal approach, making use of existing data from collaborative collections. Our next step involves further characterizing the audio samples based on their category. To improve our taxonomy, datasets, models and descriptors, we create a constant feedback mechanism integrated into our methodology by gathering user opinions and manually annotated data. We primarily use the Freesound platform as a real-world case study, and we measure user satisfaction in environments such as live coding, where precise search control and real-time retrieval are crucial.
Bio
Panagiota Anastasopoulou is a PhD researcher at the Universitat Pompeu Fabra (UPF) within the Music Technology Group (MTG) and contributes as a member of the Freesound project. Her research is centered on the tasks of characterizing and retrieving audio from large and diverse sound collections. With a strong foundation in musicology, she brings a humanistic perspective to the technological advancements emerging from her research by considering their practical applications for users and their impact on artistic expression. Beyond her academic pursuits, Panagiota is an artist who continues to create music and actively engages with artistic communities. She is a practitioner of live coding and currently participates in TOPLAP Athens and TOPLAP Barcelona communities.