Unlocking Soundscapes: Harnessing Machine Learning for Sound Classification

The concept of Smart Cities has been implemented in several major areas to improve the lives of citizens. One of the technologies used is sound classification, which monitors and identifies levels of urban noise. Convolutional Neural Networks (CNNs) have proven to be an effective method for sound classification because they can recognize complex patterns in audio data. Various architectures of Convolutional Neural Networks, such as ResNet, DenseNet, and others, exist for this purpose. This research aims to identify the CNN architecture that performs best in urban sound classification. The methodology involves training and testing existing CNN architecture and evaluating them based on several parameters, including the number of parameters, memory usage, convergence speed, and training time. Experimental results show that the performance of the ResNet family tends to degrade when the network depth increases. The best performance was achieved by the Resnet-18 architecture, with accuracy, precision, recall and F1-Score values of 91.5%, 92.2%, 91.2%, 91.6%. On the other hand, the use of complex features from various scales causes the DenseNet and GoogleNet architectures to be unableto achieve the best performance.