Research on Urban Audio Classification Based on Residual Neural Network

Abstract

In recent years, audio classification has been extensively studied, and the classification of urban sounds has great application requirements in criminal investigation and environmental protection. In this paper, a multi-feature hybrid description method is used to classify target city sounds with a multi-layer residual network structure. Firstly, a plurality of feature extraction results were compared with a conventional single feature. Secondly, different network models are studied, and their performance under different characteristics is tested and compared. Finally, comparing Resnet and multi-layer perceptrons, it is found that the Resnet50v2 method under mixed features has a better classification effect on the Ubansound8k data set, reaching 90.7%.