Singing maps: Classification of whalesong units using a self-organizing feature mapping algorithm

1 Humpback whales also make a variety of social sounds that are heard most often when the whales are interacting in groups [Thompson et. al 1977]. These sounds appear to be subject to different rules from those influencing songs. Moreover, both genders make social sounds whereas almost all observed singing humpbacks have been male [Payne & Payne 1985]. The role that song plays in the lives of humpback whales is unclear. Traditionally it was believed be purely a cultural phenomena  playing a part in courtship analogous to bird song. However, the low-frequency, repetitive, patterned vocalizations of the humpback whale may also/instead be used for environmental sensing [Frazer et. al 1996]. In this paper we refer to these vocalizations as "song" for historical reasons. A widespread problem in the study of humpback whale song vocaliza-tions involves evaluating the similarity of song elements within a whale's repertoire, between individuals of a social group, and between social groups separated by time and space. Whilst humpback whale songs demonstrate a remarkable amount of regular high level structure, they are composed of a variety of complex and transient elemental phonological units. Reliable classification of song structure requires robust unit classification  a feature which has made this process difficult to automate. This work presents a fully automated technique for performing multiple-resolution unit classification. In this scheme, units are simultaneous assigned membership to a series of increasingly general acoustic classes such that degrees of song structural similarities (and differences) emerge from analysis of units classified at different resolutions.