Conditioning for marine bioacoustic signal detection and classification

Abstract

Marine acoustic signals are characterized by certain types of noise and interference. Conditioning methods applied to spectrograms can be used to reduce or even remove these sounds, making bioacoustic signals more evident and simplifying the tasks of detection and classification. One family of methods is for making a long-term estimate of noise at each frequency and subtracting this estimate from the spectrogram; this has the beneficial effects of whitening the noise spectrum and removing relatively stationary noise sources such as vessel sound, but has the detrimental effect that relative spectrum levels - important in echolocation click classification - are altered. Another method estimates the spectrum in narrow bands at each time step and subtracts this estimate from the corresponding spectrogram frame; this method is useful for tonal sound detection and classification in that it removes short-duration clicks from snapping shrimp and echolocating animals. Other methods for removing other, more rare types of noise are presented as well. Examples and performance characterization of these methods are presented. [Funding from ONR and N45.]