Wavelet Scattering Transform for Multiclass Support Vector Machines in Audio Devices Classification System

This paper presents an acoustic device classification system using multiclass support vector machines (MCSVM) on wavelet features. Using the MCSVM classifier that includes different binary SVMs, device-wise classification accuracy is achieved with shorter computational time than a convolutional neural network (CNN). Other types of kernels for the MCSVM classifier are compared with the k-nearest neighbors (kNN) algorithm, multiclass Naive Bayes, and Decision Tree. The experiment results demonstrate that the classification accuracy of MCSVM is better than Naive Bayes, kNN, and Decision Tree when tested on Detection and Classification of Acoustic Scenes and Events for 2020 (DCASE2020) dataset. The device-wise classification accuracy for the proposed MCSVM classifier exhibits approximately 15.6% better than the baseline (via CNN) results in DCASE2020-Task1A. Hence, it has good potential in robotic and drone systems for acoustic device detection.