A Study of Sound Event Detection Techniques for Home Activity Monitoring

Abstract

This thesis is dedicated to an experimental study of home acoustic activity monitoring within sound event detection systems. The principal goal is to develop an efficient system for activity classification using a large set of audio activities within DCASE 2018 datasets. We have built a monitoring system by extracting features (Log Mel-Band energies) from time frames of each audio signal. Then, we have trained the extracted features using a deep neural network, namely Convolutional Neural Network (CNNs). Eventually, our study shows that the combination of Log Mel-band Energy features and CNN learning algorithm helps getting a good performance that allows the system to show a strong generalization ability. Key words: Acoustic Activity Monitoring, Sound Event Detection, Feature Extraction, Machine Learning.