PATTERN RECOGNITION OF ENVIRONMENTAL SOUNDS USING TIME-FREQUENCY DISTRIBUTIONS

Abstract

Environmental sounds are signals recorded in areas of environmental or ecological interest, that convey information as to the status, the inhabitation and the use / human activities of the area. An increasing research interest in this field has recently produced an increasing number of databases that contain environmental sounds. The need for automatic event classification within the recordings of such databases has accordingly grown in importance. In this paper, a novel method
for the automatic recognition of environmental sounds is presented. The signals tested are echolocation calls produced by different species of bats. In the proposed method, each signal is processed to yield a time-frequency distribution, as the basis for the feature extraction. Time-frequency distributions are then compressed by extracting appropriate features. The feature vectors formed are introduced to an Artificial Neural Network classifier, in order to obtain classification decisions for each sound / event. Experimental results obtained from the classification of the bats’ echolocation calls verify that the proposed method is capable to discriminate the aforementioned family of environmental sounds. The potential of the proposed method to perform well for other classes of environmental sounds is based on its generic, signal independent nature.