Segmentation of harmonic syllables in noisy recordings of bird vocalisations

The study of birdsong has implications in a number of biological and conservational applications. However, the analysis of bird vocalisations in the natural habitat is still largely a laborious task. One of the bottle necks is the segmentation of bird vocalisations into individual syllables. Simple segmentation in time domain proves difficult because of overlapping signals over different frequency bands. The common approach is to convert audio recordings into a spectrogram and apply image processing techniques to pick out the signal of interest. We examine several methods that have been proposed recently to do just this and find that they are inadequate to deal with harmonic vocalisations. We propose a method that segments syllables by looking for the fundamental frequency first then works its way up in the frequency axis to find other harmonics if they exist. We evaluate our method against another popular method and find that the proposed method can segment correctly more than 70% the number of syllables, more than twice that of the method we are comparing to.