Abstract:
A nonparametric statistical approach has been developed for representing variation within a class of vocalizations. This statistical representation permits computationally efficient a posteriori estimation, based on training in a supervised learning context, of the probability density associated with the class. This approach avoids the assumption of parametric forms for the probability density; parametric forms may not be suitable in cases where the probability density has a complex, multimodal character. Two applications of the statistical representation are described. First, the a posteriori densities associated with several classes of transients can be computationally ``evaluated'' on a new sample to provide a basis for classification of marine mammal vocalizations as well as a measure of confidence in that classification [J. Acoust. Soc. Am. 96, 3312 (1994); 98, 2969 (1995)]. Second, the statistical representation provides for quantitative characterization of variability using information-theoretic measures. Examples of both applications are presented. [Work supported by ONR and NRL.]