Classification of bird song syllables using singular vectors of the multitaper spectrogram

Classification of song similarities and differences in one bird species is a subtle problem where the actual answer is more or less unknown. In this paper, the singular vectors when decomposing the multitaper spectrogram are proposed to be used as feature vectors for classification. The advantage is especially for signals consisting of several components which have stochastic variations in the amplitudes as well as the time- and frequency locations. The approach is evaluated and compared to other methods for simulated data and bird song syllables recorded from the great reed warbler. The results show that in classification where there are strong similar components in all the signals but where the structure of weaker components are differing between the classes, the singular vectors decomposing the multitaper spectrogram could be useful as features.