High-Level Sound Classification in the ESOUNDMAPS Project

Abstract

ESOUNDMAPS is an ongoing research program that aims in developing a wireless audio sensor network (WASN) and deploying it at the surrounding environmental area of the Technological Institute of Piraeus in Attica, Greece. The proposed WASN will be used for the environmental monitoring of the area and aims to be used in the future for assessing the impact of human generated noise to the wildlife. Collected sound samples will be forwarded by the sensor nodes to a central server where they will be automatically evaluated with respect to their identity; environmental sound maps will be generated based on the evaluation of these sound samples. High-level sound classification is defined herein as the act of classifying a sound sample to three broad categories namely anthropogenic, biophysical (other than human) and geophysical sounds. In this paper we present an integrated platform that includes sound sample denoising using wavelets, feature extraction from sound samples and Gaussian mixture modeling of these features, and a powerful two-layer neural network classifier for the automated high-level classification of incoming sound samples. Classification results, obtained using digital sound samples, exhibit outstanding classification accuracy (sometimes reaching or exceeding 98% correct vs. incorrect estimates), thus demonstrating the feasibility of the proposed approach in realistic environments.