Cyber-Physical Analytics: Environmental Sound Classification at the Edge

Recent advances in embedded system technology have created opportunities for alleviating or eliminating common Big Data problems, by providing the resources necessary to perform AI/ML algorithms on-board edge devices. This has led to the emergence of a sub-discipline of Measurements and Signal Intelligence (MASINT) known as Cyber-Physical MASINT, wherein analysts can receive and exploit data directly from cyber-physical devices, and execute algorithms on-board, without the need to transfer to cloud servers. This type of edge analytics can decrease latency, improve security, decrease the amount of data transferred out of the device, and increase the quality of the data being transferred. With this motivation, we approach the task of environmental sound classification, a task which has seen a substantial amount of research in recent years, but which has had very limited implementation at the edge. In this work, we design and deploy an application on a mobile device to perform event detection and sound classification using a novel ensemble of deep neural networks optimized for a mobile environment, capable of classifying six common office sounds with high accuracy and low latency. We provide an accuracy and performance analysis at varying levels of optimization.