Classification of diverse call types using cluster analysis techniques

We investigated the problem of categorizing the repertoire of a group of highly varied vocalizations. A set of Weddell seal Leptonychotes weddelli in-air calls recorded near Davis, Antarctica, were examined. The repertoire size was estimated by first subjectively assigning each call (based on auditory and spectrographic patterns) to one of a large number of provisional call types. A set of frequency, duration and waveform measurements were made on every call. For each provisional call type, the mean value of each of these measures was calculated. These mean values were used to perform Cluster Analyses of the provisional call types. Beginning with calls clustered closest together, the most similar provisional call types were successively amalgamated until further joining would link two clearly dissimilar vocalizations (rising versus falling frequency sweeps in this case). The Weddell seal repertoire contained 12 call types ranging from long sinusoidal upsweeps to growls. This procedure provides a method of estimating the minimum repertoire size of a sample of calls. Dichotomous Sorting using Principal Components Analysis can also be used to categorize calls but will probably be of greater value when applied to finding subdivisions within a single call type.