Deep Representation Learning for Orca Call Type Classification

Abstract

Marine mammals produce a wide variety of vocalizations. There is a growing need for robust automatic classification methods especially in noisy underwater environments in order to access large amounts of bioacoustic signals and to replace tedious and error prone human perceptual classification. In case of the northern resident killer whale (Orcinus orca), echolocation clicks, whistles, and pulsed calls make up its vocal repertoire. Pulsed calls are the most intensively studied type of vocalization. In this study we propose a hybrid call type classification approach outperforming our previous work on supervised call type classification consisting of two components: (1) deep representation learning of killer whale sounds by investigating various autoencoder architectures and data corpora and (2) subsequent supervised training of a ResNet18 call type classifier on a much smaller dataset by using the pre-trained representations. The best semi-supervised trained classification model achieved a test accuracy of 96% and a mean test accuracy of 94% outperforming our previous work by 7% points.