Machine learning models can predict bottlenose dolphin health status from whistle recordings

Abstract

Recently, there have been incredible strides in applying machine and deep learning methods to predict the presence of disease states in humans from recordings of voice. Pattern analyses of dolphin whistles have historically depended on human operators visually and audibly differentiating between sound types. The training and implementation of machine and deep learning strategies for audio analyses reduce human bias and may improve feature detection. Our team at the Navy Marine Mammal Program in San Diego, CA has developed a substantial vocal catalog of whistle recordings for a group of focal dolphins. Additionally, an extensive health history database for each animal is maintained as part of a preventive medicine program. Together, these result in a unique, labeled dataset comprised of tens of thousands of whistles (input) emitted during differing health states (output) that have been leveraged for training machine and deep learning models to classify health status from whistles. We present preliminary results that suggest that health information may be encoded across dolphin whistle characteristics, similar to changes in the human voice. Further, we describe the applied goals for testing and implementing these innovative tools for early predicting changes in dolphin health status from non-invasive recordings.