Urban sound classification on the edge: exploring the accuracy-efficiency trade-off

Sound source classification is a valuable addition to noise monitoring, providing ‘further insights into local soundscapes. For privacy preservation, this classification often must be conducted on the edge, i.e., in real time on noise sensors. This puts constraints on the size and complexity of the classification models that can be used. Furthermore, there is a trade-off between accuracy and efficiency, which needs to be balanced on battery or solar powered sensors. However, little is known about this trade-off under consideration of constraints imposed by such sensors. In this paper, we explore the scope of sound classification models that can run efficiently on low-cost sound sensors. Specifically, we investigate the Pareto frontiers between model accuracy and computational complexity, providing insights into the trade-off necessary for deploying such models on very constrained hardware. Building on these findings, we train new classification models optimized for edge devices. The models are trained on publicly available audio samples and a new Dutch Urban Sounds dataset specifically collected to enhance the accuracy of sound source classification in urban environments. The models and implementation are open source, enabling researchers and practitioners to adopt, adapt, and build upon our work.