A multi-resolution hidden markov model for optimal detection, tracking, separation, and classification of marine mammal vocalizations

Abstract

We employ the multi-resolution hidden Markov model (MRHMM) to develop an improved algorithm for modeling marine mammal wandering tone vocalizations (whistles). A vocalization is modeled by a series of time segments in which the signal has a constant frequency rate (chirps). Rather than using chirps of uniform length, the segments are allowed to be of variable size, thus adapting to both short rapid changes in frequency rate as well as long segments of constant rate. The method supports the goals of finding the single best segmentation or the average joint probability density function of the data over all possible segmentations, weighted by the a priori probability of each segmentation. The probability density function (PDF) projection theorem is used to allow likelihood comparisons in the raw data domain. Simulated data and recorded marine mammal vocalizations are used to demonstrate the technique.