We propose to use acoustic features of both clicks and whistles to classify odontocete sounds to species. The species studied are bottlenose dolphin (Tursiops truncatus), spinner dolphin (Stenella longirostris), melon-headed whale (Peponocephala electra), and short- and long-beaked common dolphin (Delphinus delphis and D. capensis). An energy-based detector is used for echolocation click detection, and Roch's Silbido algorithm is used for whistle detection. Detected whistles are characterized by maximum and minimum frequencies, duration, slope, spectral maxima, spectral gaps, number and frequency of inflection points, number of "loop" repetitions and other acoustic characteristics. Detected clicks are characterized by cepstral characteristics, as well as by a set of noise-resistant statistics. Clicks that occur within a certain time neighborhood of a whistle have the corresponding feature vectors merged to produce the input to the classification system. Random forest and Gaussian mixture model classifiers are tested on the resulting features and performance is characterized. [Funding from ONR.]

This content is only available via PDF.