Vocalizations in social interactions among captive dolphins

In order to understand the role of vocalization in dolphin social interactions, we need to know which animals are engaged with which other animals, what activity they are engaged in, and which animals are vocalizing. In this poster we describe our efforts to address the linked problems of identifying individual animals and identifying which animal produced which sounds. Solving these problems is a necessary foundation for understanding the role of vocalizations in dolphin social interactions. We first address how we are developing tools to facilitate the work of human observers who document the locations of dolphins. Secondly, we show the development of computational techniques to automatically extract a particular vocalization type, the whistle, from our huge audio corpus. Finally, we explore computational methods to cluster one whistle type into multiple subtypes, as previous research suggests that vocalizer identity might be linked to whistle variation. We evaluate our whistle detector and present the first known results of attempting to cluster vocalizer by whistle subtype. We conclude by discussing future improvements to our methods, ways to use our new massive dataset of whistles, and how we can use these vocalizations to inform our understanding of dolphin behavior and communication.