Automatic detection of regularly repeating vocalizations

Abstract

Many animal species produce repetitive sounds at regular intervals. This regularity can be used for automatic recognition of the sounds, providing improved detection at a given signal-to-noise ratio. Here, the detection of sperm whale sounds is examined. Sperm whales produce highly repetitive ‘‘regular clicks’’ at periods of about 0.2–2 s, and faster click trains in certain behavioral contexts. The following detection procedure was tested: a spectrogram was computed; values within a certain frequency band were summed; time windowing was applied; each windowed segment was autocorrelated; and the maximum of the autocorrelation within a certain periodicity range was chosen. This procedure was tested on sets of recordings containing sperm whale sounds and interfering sounds, both low-frequency recordings from autonomous hydrophones and high-frequency ones from towed hydrophone arrays. An optimization procedure iteratively varies detection parameters (spectrogram frame length and frequency range, window length, periodicity range, etc.). Performance of various sets of parameters was measured by setting a standard level of allowable missed calls, and the resulting optimium parameters are described. Performance is also compared to that of a neural network trained using the data sets. The method is also demonstrated for sounds of blue whales, minke whales, and seismic airguns. [Funding from ONR.]