Passive Acoustic Monitoring for the Detection and Identification of Marine Mammals

This project is a multi-pronged study to advance the state of the field in three areas. The development of automated auditory scene analysis for delphinid tonal calls will permit subsequent work by this investigator or others to exploit the use of whistles for classification and localization. Our approach is to dynamically build hypothesis graphs using phase-frequency representations of the signal. In parallel to this effort, two modeling techniques are being pursued to improve existing passive acoustic monitoring capabilities based on echolocation clicks of odontocetes. The first of these examines the use of a universal background model as proposed by Reynolds et al. (2000) for human speaker verification tasks. Reynolds’ problem, which is similar in nature to ours, is how can one reject observations from a speaker (or dolphin species) for which there is no data to create a model. We adapt his idea of a universal background model by training a generalized odontocete model using the data of a number of species. This model is not specific to any one species. Using Bayesian learning, training data from a specific species adapts the parameters of the generalized model, thus serving as a foil against vocalizations that sound similar to one of our species. The second approach for echolocation clicks exploits recent machine learning work on submanifold learning (Dasgupta and Freund, 2007; Dasgupta and Freund, 2008; Freund et al., 2007). In order to detect and classify odontocetes, features, or poignant characteristics of their signals, must be extracted from the audio signal. As the underlying process of sound generation cannot be measured directly, nor is it well understood, classification techniques must attempt to infer information about the producer of the signal (e.g. species) through a typically higher-order set of features. Submanifold learners focus on learning a subspace of the high-order feature space that can be more conducive to providing robust classification.