Gunshot detection neural network implemented on a low-cost microcontroller

Abstract

Nowadays, criminal activity is on the rise, and it usually involves some type of firearm. There are automated shot detection systems but in the end, they still require human intervention to decide if it is an actual gunshot. Distinguishing between two similar sounds, such as the detonation of a firearm or a firecracker, is not always possible with the naked ear. There are multiple publications on artificial intelligence to identify gunshots; however, they use convolutional neural networks, which, despite being highly effective, require a system with extensive computational resources. This document presents a fully connected neural network implemented on a microcontroller that can identify up to 90% of firearm detonations. This document will be of interest to students or researchers interested in the design of neural networks for sound recognition on embedded systems.