Ensemble of Handcrafted Features of Environment Sound Classification Using a Deep Convolutional Neural Network to Enhance Accuracy and Reduce Computational Complexity

Abstract

Environmental sound classification (ESC) is an area of active research in signal and image processing that has made significant strides over the past several years. The goal of ESC is to classify environmental sounds by extracting and analyzing handcrafted and deep features from various acoustic events. The task is complex because environmental sounds are typically unstructured, nonstationary, and overlapping. Multiple deep learning (DL) approaches have successfully tackled the ESC problem and outperformed conventional classifiers like k-nearest neighbors (kNN) or support vector machine (SVM). However, most DL approaches have high computational costs, making them unsuitable for use in embedded systems applications. In this dissertation, we propose four models that require low computational costs and achieve high accuracy in classifying environmental sounds. Model 1 uses kNN to analyze and extract multiple temporal and spectral handcrafted features. Model 2 extracts deep features from different spectrograms using a proposed deep convolutional neural network (DCNN) with six convolutional layers and four max-pooling layers, totaling 150k parameters. Models 3 and 4 combine handcrafted and deep features to improve classification accuracy. We tested the proposed models on a public dataset called Urbansound8k and achieved a classification accuracy of 95.3%.