Dynamic time warping for automatic classification of killer whale vocalizations

Abstract

A large number of sounds from the captive killer whale population at Marineland of Antibes, France have been classified perceptually into call types. [A. Hodgins-Davis, thesis, Wellesley College (2004)]. The repetition rate of the pulsed component of five or more examples of each call type was calculated to give a set of melodic contours. These were compared pairwise using dynamic time warping to give a dissimilarity or distance matrix. The distances were then transformed into a component space using multidimensional scaling, and the resulting points clustered with a k-means algorithm. In grouping 57 sounds into 9 call types, there was a single discrepancy between the perceptual and automated methods. Preliminary measurements on a second set of 74 sounds from northern resident killer whales have been made. Both the high-frequency and low-frequency contours have been measured and classified perceptually into 7 groups. Errors for the automatic classification of this group are over 10%. [Sound recording supported by grants from WHOI’s Ocean Life Institute and the Royal Society to Patrick J. O. Miller.]