Automatic Detection and Classification of Beluga Whale Vocalizations

In this work, an algorithm has been proposed for real time detection and classification of beluga whale calls. The detection algorithm is based on an adaptive activity detector that exploits a priori knowledge of the longest/ shortest beluga whale sound unit. Optimum parameter values of the proposed detector are obtained by simulation to maximize the difference between Detection Probability and False Alarm Probability. A set of features that allow successful classification by means of a Naive Bayes classification algorithm has been put forward as well. Three classification categories related to observed beluga behaviours were selected. In a different perspective, the proposed features can be employed to obtain clues of how beluga sounds are produced and the degree of control that the specie has over its internal organs. As an example, the presence of nonlinearities such as subharmonics and frequency jumps have been detected and related to some extracted features. This technique can serve as a complement to more rigorous studies based on video information and ultrasonic sensors for whale monitoring.