Abstract
The time-frequency contours of bottlenose dolphin (Tursiops truncatus) whistles are commonly extracted as features for the classification and comparison of these signals. This research presents a new method to extract and track the whistles’ fundamental frequency contours. This method takes a spectrogram as an input and uses a Bayesian inference approach with a random walk model for the frequency transitions. The output of the method is the maximum a posteriori probability estimate of the frequency contour. This technique is similar to the Kalman filter except that the states are probability densities, and it uses a probabilistic state transition model. For each time segment, the method combines the prior estimate of the probability distribution with the current observation to obtain the current estimate of the probability distribution. Experimental evaluation indicates that this method greatly improves the reliability of contour extraction from noisy signals when compared to the ad hoc contour extractor proposed by Buck and Tyack [J. Acoust. Soc. Am. 94 (1993)]. This technique has previously applied to target tracking in radar [Bethel and Rahikka, IEEE Trans. Aerosp. 23(6) (1987)]. [Work supported by NSF Ocean Sciences.]