Classification of Environmental Sounds Using Convolutional Neural Network with Bispectral Analysis

Abstract

To realize a useful acoustic environmental recognition system, we propose a new method that classifies sound signals using a slice bispectrogram, which is a third-order version of a spectrogram. The classified sound was used as input data for a convolutional neural network. We conducted a fundamental classification experiment using UrbanSound8k, which was an open dataset consisting of 10 classes of environmental sounds. Our proposed method gave high accuracy and stability. Furthermore, a relationship between the accuracy and non-Gaussianity of sound signals was confirmed.