A graph search algorithm for delphinid whistle contours.

Abstract

A graph search algorithm is presented to extract tonal contours from audio signals in a fully automated manner. The algorithm is capable of tracking multiple contours simultaneously. Search is done in the spectral domain, with contours being treated as paths through a phase‐magnitude space. A graph search formalism is combined with dynamic programming techniques to track candidate paths that may overlap or cross. Disambiguation of path crossings is accomplished by examining a combination of slope and the derivative of phase. The algorithm is general in nature and applicable to tonal calls from many animals with little modification. We target this work specifically toward delphinid whistles and demonstrate its efficacy on calls from bottlenose (Tursiops truncatus), short‐ and long‐beaked common (Delphinus delphis and D. capensis), and Risso’s (Grampus griseus) dolphins recorded in the Southern California Bight. Characterizing the complete contour will permit the development of call comparison techniques that take into account finer grained detail than the typically used contour statistics (e.g., number of inflection points, minimum, and maximum frequencies). [This work is sponsored by the Office of Naval Research.]