Addressing challenges in automated whistle extraction

Abstract

Automated extraction of dolphin whistles from field recordings presents a challenging problem, as typically these recordings are noisy, contain multiple overlapping whistles and interfering signals. This results in many spurious detections, partial extraction and fragmentation of the whistle contours, which could cause problems for applications such as whistle-based localization. This paper presents the probability hypothesis density (PHD) filter, a non-traditional whistle tracking method based on multi-target Bayesian framework. Two implementations of the PHD filter, one based on Gaussian mixtures and one based on particle methods, are adapted for this task and tested on a large real-world dataset consisting of six dolphin species. The proposed filters successfully track simultaneous whistles, and compare favourably to standard methods. Moreover, it is shown how the proposed methods can be used to enhance the localization efforts with linear towed arrays.