Comparative Evaluation of CNN, VGG16, Conformer, and YAMNet Models for Skateboard Riding Sound Detection

In this paper, we present a comparative evaluation of four models, namely CNN, VGG16, Conformer, and YAMNet, as core components of a skateboard riding sound detection system. This system is designed to help reduce monitoring costs in public areas where unauthorized skateboarding may cause pedestrian obstruction or noise problems. We focus on the trade-off between model complexity, measured by the number of parameters, and detection performance, as well as the robustness of each model in realistic noisy environments. Experimental results show that the VGG16-based model achieves the highest performance, with area under the curve (AUC) values of 0.987 and 0.948 for clean and noisy test data, respectively. Notably, the CNN model, despite having fewer parameters and lacking pre-training, demonstrates strong performance, outperforming YAMNet and Conformer under noisy conditions, with AUC values of 0.970 and 0.904 for clean and noisy test data, respectively.