On contour-based classification of dolphin whistles by type

Classification of cetacean vocalizations may help marine biologists study their behavioral context in different environments yet automatic classification of vocalizations for their information content has not been adequately addressed in the literature. Since classifier performance has a strong dependence on the extent to which features cluster, we, in this paper, explore the effect of two feature sets on two classifiers and assess their performance and computational complexity. We choose two feature sets that are exemplary of very different methods: The first set consists of Tempo-Frequency Parameters (TFPs) that are hand-picked to describe the spectral whistle contours. The second feature set embodies spectral information measured with the Fourier Descriptors (FD) commonly used in image processing for contour representation. The computed feature vectors are fed into the K-nearest neighbor (KNN) and Support Vector Machine (SVM) classification algorithms. The KNN in its basic form is a simple classifier that works well if feature clusters have clear margins and SVM uses a data dependent margin chosen for optimal performance. We argue that KNN serves to accentuate the effect of the feature sets and the SVM acts as the scientific process control. Experimental results show best results with the combination of the TFP feature extractor and the SVM classifier, suggesting a future research direction of developing non-linear kernels for SVM.