Classification of marine mammal vocalizations using machine learning

Abstract

Marine mammal vocalizations are a significant component of the underwater soundscape in many parts of the world. The ability to detect and classify these vocalizations has implications in marine mammal research and tracking, acoustic noise measurement and modeling, and surface and submerged ship navigation. Advances in machine learning and deep neural networks have enabled classification technologies that can match and even exceed human performance in many areas. In this talk, we will explore a supervised, multi-label classifier using a Convolutional Neural Network to discriminate different types of marine mammal vocalizations across species using spectrograms derived from single-channel audio recordings from the Watkins Marine Mammal Sound Database of the Woods Hole Oceanographic Institute. The Watkins database contains over seven decades of recordings of more than 60 professionally identified marine mammal species and is publicly available for scientific research. We will discuss in details the rationale for our data labeling and pre-processing, model training and evaluation, and data and model visualization in order to understand and improve classification performance and robustness.