Explainable Acoustic Scene Classification with Audio Tags and Feature Ablation
Darshit Amit Pandya, Heiner Stuckenschmidt · 2025
Computational Auditory Scene Analysis (CASA), also known as Machine Listening, has gained increasing attention, requiring a deeper understanding of the methods and models used in this field. Acoustic Scene Classification (ASC) is a key task within CASA, providing contextual information essential in several domains. This study investigates ASC's decision-making process by integrating feature ablation with fine-grained audio tags (AT). We demonstrate that such integration can give us a better, finer explanation of a model's decision-making process. This approach provides deeper insights into the learned representations of ASC models, improving interpretability and guiding the development of more robust and generalizable audio classification systems.