Dr. Dolittle Project: A Framework for Classification and Understanding of Animal Vocalizations

Granting Agency: National Science Foundation
Award number: 0326395
Investigators:

• Michael T. Johnson
• Michael Darre
• Anne Savage
• Peter Scheifele
• Elizabeth von Muggenthaler

Abstract
The fundamental goal of this research is to develop a broadly useable framework for pattern analysis and classification of animal vocalizations, by integrating successful models and ideas from the field of speech processing and recognition into bioacoustics. Current approaches to automated methods for analyzing and classifying animal vocalizations are still significantly behind the capabilities that exist in the field of human speech processing, and cross-fertilization with other fields offers a tremendous opportunity for making advancements in this area. The PI and his team plan to employ an innovative cross-disciplinary strategy to develop a framework based on robust feature analysis and selection and Hidden Markov Model classification, with additional expertise drawn from such fields as psychology, biology, linguistics, machine learning, and signal processing, for the purpose of making significant advancements to the current state-of-the-art in bioacoustics algorithms and animal communications research. The algorithms developed through this process will be applied to a wide range of important tasks, including automatic vocalization classification and labeling, individual identification, call type classification, behavioral-vocalization correlations, stress analysis, language acquisition, and seismic infrasonic communication. Species being targeted for study include domestic and agricultural animals, marine mammals, and several endangered species. The primary contribution of this work will be the development of well-founded models and methods for improving our understanding of animal communication and behavior. Experimental tasks of significance to a wide variety of research problems within this field are also planned. In addition, many aspects of the HMM classification framework, including hierarchical feature analysis and selection, variable frame sizing, and learnable model topologies, offer the potential for advances in machine learning as well as speech and signal processing.
Broader Impacts: The development of an easily adaptable framework for applying speech-processing techniques to animal vocalizations will allow other researchers to create systems that pertain to the specific species they are studying. This will speed the development effort of these systems and allow those not familiar with speech and signal processing techniques to incorporate advanced methods in their research. A primary underlying motivation for such research is the preservation of endangered species and the improvement of careand habitats for animals in captivity, and this project has the potential for significant, long-term impact toward these goals.