Comparing call-based versus subunit-based methods for categorizing Norwegian killer whale, Orcinus orca, vocalizations

Abstract

Students of animal communication face significant challenges when deciding how to categorise calls into subunits, calls, and call series. Here, we use algorithms designed to parse human speech to test different approaches for categorising calls of killer whales. Killer whale vocalisations have traditionally been categorised by humans into discrete call types. These calls often contain internal spectral shifts, periods of silence, and synchronously produced low and high frequency components, suggesting that they may be composed of subunits. We describe and compare three different approaches for modelling Norwegian killer whale calls. The first method considered the whole call as the basic unit of analysis. Inspired by human speech processing techniques, the second and third methods represented the calls in terms of subunits. Subunits may provide a more parsimonious approach to modelling the vocal stream since (1) there were fewer subunits than call types; (2) nearly 75% of all call types shared at least one subunit. We show that contour traces from stereotyped Norwegian killer whale calls yielded similar automatic classification performance using either whole calls or subunits. We also demonstrate that subunits derived from Norwegian stereotyped calls were detected in some Norwegian variable (non-stereotyped) calls as well as the stereotyped calls of other killer whale populations. Further work is required to test whether killer whales use subunits to generate and categorize their vocal repertoire.