Automatic Detection for Acoustic Monitering of Wild Animals

Abstract

We achieve, to our knowledge, state of the art results for detecting elephant calls from acoustic data on two different tasks of interest to researchers: 1) A segmentation task where we predict for each time step input whether there is an elephant call present and 2) A trigger word detection task where we predict the completion of an elephant call. The latter is of particular interest because of its potential to allow researchers to run in real time computationally efficient models on low-cost devices. We used data collected by passive acoustic monitoring (PAM) systems in the African jungle provided generously by the Cornell Lab of Ornithology.