Estimating the Repertoire Size in Birds using Unsupervised Clustering techniques

Abstract

Birds produce multiple types of vocalizations that, together, constitute a vocal repertoire. For some species, the repertoire size is of importance because it informs us about their brain capacity, territory size or social behaviour. Estimating the repertoire size is challenging because it requires large amounts of data which can be difficult to obtain and analyse. From birds vocalizations recordings, songs are extracted and segmented as sequences of syllables before being clustered. Segmenting songs in such a way can be done either by simple enumeration, where one counts unique vocalization types until there are no new types detected, or by specific algorithms permitting reproducible studies. In this paper, we present a specific automatic method to compute a syllable distance measure that allows an unsupervised classification of bird song syllables. The results obtained from the segmenting of the bird songs are evaluated using the Silhouette metric score.