Classification of Urban Sound using Custom 1D and 2D CNN Architectures

In this work, we will study and analyze the performance of Convolutional Neural Networks (CNN) based techniques for sound classification which may help in case of urban surveillance applications.Recently, many researchers have used different feature extraction techniques along with different machine learning (ML) techniques to classify urban sound.Deep learning (DL) models like VGG-16, AlexNet, GoogleLeNet have also been explored in this direction effectively.In this research work, we proposed various custom designed 1D and 2D CNN models along with different extracted features like; Mel Frequency Cepstral Coefficients (MFCC), Chroma short-time Fourier transform (Chroma stft), Mel-spectrogram, and Spectral Contrast for urban sound classification.Detailed ablation studies were carried out with respect to different filter sizes, number of convolutional layers, and various activation functions to obtain state-of-the-art performances on the standard urban sound dataset.The experimental results show better performances with our custom CNN models compared to the classification models proposed in the existing recent literature works.