Abstract
The time-frequency contours of bottlenose dolphin (Tursiops truncatus) whistles are commonly extracted features for the classification and comparison of these signals. We have previously developed a method to extract the whistles’ fundamental frequency contours [J. Acoust. Soc. Am. 108, 2635 (2000)] by sequentially tracking the contour. These contours are often extracted from recorded rather than real-time signals, which allows us to read ahead or access the ‘‘future’’ signal. This talk extends our previous work to propose a technique to estimate the fundamental frequency at each time from the ‘‘future’’ values of the recorded signal in addition to the past and present signals. This technique is similar to the Kalman smoother except that the states are probability densities, and it uses a probabilistic state transition model with a Bayesian projection to produce the maximum a posteriori probability estimate of the frequency contour based on all recorded data, not just the present and past signal. Experimental evaluation indicates that this method greatly improves the reliability of contour extraction from noisy signals, with considerable resistance to strong harmonics, some dropouts and echolocation clicks, when compared to our previous technique. A huge improvement is achieved from the technique proposed by Buck and Tyack [J. Acoust. Soc. Am. 94, 2497–2506 (1993)].