Differentiating marine mammal clicks using time-series properties

Automatic detection of marine animal vocalizations is increasingly used to analyse the extensive acoustic datasets collected from autonomous passive acoustic recorders, resulting in a constant effort to improve detector accuracy and develop new and more efficient detection methods. Differentiating between the clicks produced by odontocete species can be especially problematic due to overlapping time and frequency characteristics. Classifiers for clicks are often based on short-time Fourier transforms or Wigner-Ville transforms which are computationally expensive. We propose a computationally efficient method of detecting and differentiating between clicks based on a two stage classifier. First, an initial classification is obtained using three features derived from the time series of the click. Second, the cepstrum is used to determine the inter-click interval. The method was tested on recordings which included clicks produced by small dolphins, killer, pilot and sperm whales, as well as at least three species of beaked whales. This new approach increases the efficiency of analysis and provides reliable species classification.