Baleen whales live at extended timescales. Study of how these animals use their vocalisations for communication requires massive data sampling over long periods. The volume of data precludes traditional hands-on analysis techniques, at least during the first stages of data reduction. This paper describes a system for automating the sampling and analysis of baleen whale calls. Blue Balaenoptera musculus and fin B. physalus whale calls are very stereotypical. Blue whale 'A' and 'B' calls have fundamental frequencies of approximately 17 Hz, narrow bandwidth and well-defined harmonic structure, and typical duration of 15-25 sec. Fin whale 'pulses' have fundamental frequencies of approximately 17 Hz, but are broadband in nature and short (~ 1 sec) duration. The homogeneous call structure lends itself to automated detection. Stable acoustical differences in call structure lend themselves to automated species identification. We have benchmarked a series of bioacoustical call identification algorithms against a set of blue and fin whale calls while systematically manipulating the signal to noise ratio. The results demonstrated a typical tradeoff of speed versus accuracy. The best algorithm was inserted into the underwater sound recording system and its signal-detection theoretic performance was quantified. Results will be discussed with respect to technological and ecological aspects of baleen whale bioacoustics (Prolect CS-1082 of the Strategic Environmental Research and Development Program).
Abstract: