Analysing differences among animal songs quantitatively by means of the Levenshtein distance measure

Abstract

The Levenshtein or string edit distance is an objective measure of the difference between two strings of elements. Levenshtein distance analysis has previously been applied to humpback whale songs, where it provided a quantitative measure of song change from year to year. This analysis is extended and a first approach to a robust statistical test is developed. The statistical test addresses the central issue whether two groups of songs (either from different individuals, different groups or different years) belong to the same population of songs or are significantly different. This is accomplished through derivation of the Kohonen median song sequence, which has the smallest possible summed Levenshtein distance to all songs of the group. By a simple t-test or nonparametric equivalent it is tested whether the median distance to the Kohonen median song sequence of a second group is significantly larger, which indicates that the groups are different. The test is expanded to handle multiple comparisons among several groups of songs.