Exponential Moving based Features for Acoustic Scene Classification

Abstract

Acoustic scene classification (ASC) aims to classify a sound recording, which is recorded from a particular environment into a predefined category that describes the environment. In order to better model event behaviors within acoustic scenes, we investigate the use of exponential moving (EM)-based statistical representations for the ASC task. To this end, we design a convolutional neural network (CNN) based approach using statistical EM-based representations of log mel-band energies. We evaluate our proposed method on the publicly available performance dataset, DCASE 2022 Low-Complexity Acoustic Scene Classification Challenge dataset. Results show that the EM-based statistical representations achieve higher classification accuracy and better log losses compared to just using the log Mel feature.