Oscillatory whistles – the ups and downs of identifying species in passive acoustic recordings
Vincent M. Janik – email@example.com
Scottish Oceans Institute
School of Biology
University of St Andrews, UK
Popular version of paper 1pAB6 Oscillatory whistles—The ups and downs of identifying species in passive acoustic recordings
Presented Tuesday afternoon, June 8, 2021
180th ASA Meeting, Acoustics in Focus
Many dolphin species communicate using whistles. Because whistles are produced so frequently and travel well under water, they are the focus of a wide range of passive acoustic studies. A challenge inherent to this type of work is that many acoustic recordings do not have associated visual observations and so species in the recordings must be identified based on the sounds that they make.
Acoustic species identification can be challenging for several reasons. First, the frequency contours of dolphin whistles are variable, and each species produces many different whistle types. Also, whistles often exhibit significant overlap in their characteristics between species. Traditionally, acoustic species classifiers use variables measured from all whistles, regardless of what type they are. An assumption of this approach is that there are underlying features in every whistle that provide information about species identity. In human terms, we can tell a human scream or grunt from those of a chimpanzee because they sound different. But is this the case for dolphin whistles? Can a dolphin tell whether a whistle it hears is produced by another species? If so, is species information carried in all whistles?
To investigate these questions, we analyzed whistles produced by short- and long-beaked common dolphins in the Southern California Bight. Our previous work has shown that the whistles of these closely related species overlap significantly in time and frequency characteristics measured from all whistles, so we hypothesized that species information may be carried in the shape of specific whistle contours rather than by general characteristics of all whistles. We used artificial neural networks to organize whistles into categories, or whistle types. Most of the resulting whistle types were produced by both species (we called these shared whistle types), but each species also had distinctive whistle types that only they produced (we called these species-specific whistle types). Almost half of the species-specific whistles produced by short-beaked common dolphins had oscillations in their contours, while oscillations were very rare for both long-beaked common dolphins and shared whistle types. This clear difference between species in the use of one specific whistle shape suggests that whistle type is important for species identification.
We further tested the role of species-specific whistle types in acoustic species identification by creating three different classifiers for the two species – one using all whistles, one using only whistles from shared whistle types and one using only whistles from species-specific whistle types. The classifier that used whistles from species-specific whistle types performed significantly better than the other two classifiers, demonstrating that species-specific whistle types collectively carry more species information than other whistle types, and the assumption that all whistles carry species information is not correct.
The results of this study show that we should re-evaluate our approach to acoustic species identification. Instead of measuring variables from whistles regardless of type, we should focus on identifying species-specific whistle types and creating classifiers based on those whistles alone. This new focus on species-specific whistle types would pave the way for more accurate tools for identifying species in passive acoustic recordings.