Acoustic Events Processing with Deep Neural Network

Abstract

Safety is one of the society requirement, what we need for cheerful live. The principal purpose is to recognize potentially dangerous acoustic events (gun shooting and glass breaking). This document compares a Neural Network (NN) based on the detection system and a hidden Markov model based on the acoustic event detector. For both methods, the same database was used. The database consisted of shots, glass breaks and background noise. Proposed deep neural network processes an acoustic signal through two hidden layers. The whole process may divide into three parts. Training, testing and evaluation part. As the main resulting parameter accuracy has been chosen. This computation process uses a confusion matrix for reliable detection. Accuracy is compared with previous research in this area, as well.