Dividing the acoustic repertoires of animals into biologically relevant categories presents a widespread problem in the study of animal sound communication, essential to any comparison of repertoires between contexts, individuals, populations, or species. Automated procedures allow rapid, repeatable, and objective categorization, but often perform poorly at detecting biologically meaningful sound classes. Arguably this is because many automated methods fail to address the nonlinearities of animal sound perception. We present a new method of categorization that incorporates dynamic time-warping and an adaptive resonance theory (ART) neural network. This method was tested on 104 randomly chosen whistle contours from four captive bottlenose dolphins (Tursiops truncatus), as well as 50 frequency contours extracted from calls of transient killer whales (Orcinus orca). The dolphin data included known biologically meaningful categories in the form of 42 stereotyped whistles produced when each individual was isolated from its group. The automated procedure correctly grouped all but two stereotyped whistles into separate categories, thus performing as well as human observers. The categorization of killer whale calls largely corresponded to visual and aural categorizations by other researchers. These results suggest that this methodology provides a repeatable and objective means of dividing bioacoustic signals into biologically meaningful categories.
We thank the staff of Zoo Duisburg for the opportunity to work with their animals and for their support during the recording of dolphin whistles, especially Roland Edler, Reinhard Frese, Manuel García Hartmann, Friedrich Ostenrath, and Ulf Schönfeld. Recordings for the analysis of killer whale vocalizations were generously supplied by Nancy A. Black, John K. B. Ford, P. Dawn Goley, Dan McSweeney, Paul Spong, and Richard L. Ternullo. The ART2 neural network algorithm was adapted from a program originally written by Aaron Garrett, and Mary Royerr helped with statistical aspects of this paper. Earlier drafts of this manuscript benefited from comments by Karen E. McComb, John K. B. Ford, Michael J. Ritchie, and Peter J. B. Slater. V.M.J was funded by a Royal Society University Research Fellowship, and V.B.D received financial support from a DAAD Doktorandenstipendium aus Mitteln des 3. Hochschulsonderprogramms during part of this study.
A. Categorization of soundpatterns by humans and computers
B. Time and frequency resolution in the auditory perception of birds and mammals
C. Unsupervised learning in artificial neural networks
A. Data sets, acoustic analysis, and contour extraction
B. ARTwarp—Combining dynamic time-warping and adaptive resonance theory
C. Experiment I: Categorization of bottlenose dolphin whistles
D. Experiment II: The appropriate fineness of categorization
E. Visualization of neural network performance
A. Experiment I: Categorization of bottlenose dolphin whistles
B. Experiment II: The appropriate fineness of categorization
C. Visualization of neural network performance
A. Categorization of bottlenose dolphin whistles
B. Choosing the vigilance parameter
C. Applicability to other categorization problems in the study of behavior
Data & Media loading...
Article metrics loading...
Full text loading...