A common task for researchers of animal vocalisations is statistically comparing repertoires, or sets of vocalisations. We evaluated five methods of comparing repertoires of ‘codas’, short repeated patterns of clicks, recorded from sperm whale (Physeter macrocephalus) groups. Three of the methods involved classification of codas – human observer classification, k-means cluster analysis using Calinski and Harabasz’s (1974) criterion to determine k, and a divisive k-means clustering procedure using Duda and Hart’s (1973) criterion to determine k. Two other methods used multivariate distances to calculate similarity measures between coda repertoires. When used on a sample coda dataset, observer classification failed to produce consistent results. Calinski and Harabasz’s criterion did not provide a clear signal for determining the number of coda classes (k). Divisive clustering using Duda and Hart’s criterion performed satisfactorily and, encouragingly, gave similar results to the multivariate similarity measures when used on our data. However, the relative performance of the k-means techniques is likely data dependent, so one method is not likely to perform best in all circumstances. Thus results should be checked to ensure they extract logical clusters. Using these techniques concurrently with multivariate measures allows the drawing of relatively robust conclusions about repertoire similarity while minimising uncertainties due to questionable validity of classifications.
cluster analysis, classification, vocal repertoire, sperm whale, codas