The Humpback Whale has a variable and complex song containing a large variety of different syllables, with a typical duration of 10 minutes. It is closely connected to communication with other specimens and reproductive success. However, the quantiﬁcation of such variable song is challenging mainly due to noise which accompanies the underwater sounds and the resemblance between different sounds. In this project, we present automated methods for clustering syllables in Humpback Whale song. The method provides a tool for pairwise comparison of syllables with the aim of grouping them in terms of their similarity. This allows to analyze the repertoire size within an individual. Our method starts with time-frequency analysis of the syllables, each syllable was represented by Spectrograms with 3 different window sizes and Ambiguity Function. Due to the large dimensionality of the data, dimensionality reduction was performed using SVD decomposition. Similarity matrices were calculated using 3 different distance metrics. The matrices rows were used as feature sets for low-dimensional signal representation. K-Means and Hierarchal Clustering algorithms were performed based on the feature sets, and the optimal number of clusters was evaluated using Davies-Bouldin and Silhouette validation indices. The proposed algorithm is evaluated using a synthetic data base and by means of comparison to a manual recorded song investigation done by a specialist. The results show that the metric based on Frobenius Norm gives the best clustering out of the tested metrics. In addition, the Ambiguity Function clustering quality is roughly the same as the Spectrogram clustering quality but is more computationally efficient.
Yaron Duek RAFAEL