Unsupervised classification of graded animal vocalisations using fuzzy clustering

Graded vocal repertoires represent a challenge for bioacoustics. We present an unsupervised classification procedure designed to take gradation into account, based on Mel frequency cepstral coefficients and fuzzy clustering. Cepstral coefficients are well defined for tonal, broadband and pulsed sounds. They compress information about the distribution of energy across frequencies into a limited number of variables. The Mel scale mimics the perception of pitch by mammalian ears. Fuzzy clustering is a soft classification approach. Instead of assigning samples to a single category, it describes their position relative to overlapping clusters and can therefore identify stereotyped and graded vocalisations within a repertoire. We evaluated the performance of this procedure on a set of long-finned pilot whale (Globicephala melas) calls. Fuzzy clustering was much less time-consuming than manual classification (days vs. months), but identified a smaller number of categories (three to six fuzzy clusters compared to 11 human-defined call types). Some fuzzy clusters were similar to sets of human-defined call types, but some call types were spread over several fuzzy clusters. Fuzzy clusters provide new quantitative insight about the gradation of vocal repertoires. We discuss the results and the need to investigate the functions of call gradation in future research.