Wavelet convolution: an urban sound source recognition method

As urban environmental noise pollution continues to worsen, the demand for high-precision real-time monitoring systems based on acoustic recognition methods has become increasingly critical. Wavelet convolution has demonstrated strong performance in two-dimensional natural image modalities, particularly by enlarging receptive fields and enhancing the representation of low-frequency features. However, the feasibility of wavelet convolution in non-natural graphical modalities, such as spectrograms, remains to be systematically investigated. In this study, an urban sound classification system based on wavelet convolution is proposed. The experiments were carried out on a dataset that included five categories of urban environmental sounds. The wavelet convolution model achieved an accuracy and F1-score of 0.991 on the validation set, outperforming conventional baseline models. Moreover, compared with the baseline, the model achieved reductions of 64.2% and 64.1% in total parameter count and inference time, respectively. These results demonstrate that the proposed wavelet convolution-based sound classification system combines high classification performance with lightweight efficiency, making it highly suitable for real-time monitoring applications.