Urban Sound Classification Using Features with Multiple Windows Fusion and a Comparative Analysis of CNN, DNN, and LSTM

This study addresses the task of environmental sound classification (ESC) by introducing features composed of multiple window channels as input, to compare the performance of convolutional neural networks (CNN), deep neural networks (DNN), and long short-term memory (LSTM) models. The primary contribution of this research lies in the rowwise concatenation of three distinct window channels, each generating a 64-dimensional Mel-frequency. Given the consistent distribution of these features and the presence of frequency band differences, the mix-up feature augmentation technique is employed to further enhance performance. Three mainstream classification models are utilized to validate and compare the proposed feature extraction approach and its effectiveness. The CNN model achieves an accuracy of 99.28%, representing stateof-the-art performance on the UrbanSound8K dataset.