soundClass: An automatic sound classification tool for biodiversity monitoring using machine learning

Abstract

Passive acoustic monitoring, a non-invasive technique, is increasingly used to study animal populations and habitats at much larger spatial and temporal scales than standard methods. However, easy to apply tools for reliable detection and classification of signals of interest among hundreds or even thousands of hours of recording are still lacking. We introduce the r package soundClass, a tool to train convolutional neural networks, and employ them to classify sound events in recordings. soundClass provides a sound event classification pipeline, from annotating recordings to automating trained networks usage in real-life situations. We illustrate the package functionality on bat echolocation calls, bird songs and whale echolocation clicks, showing that the package can be used to train networks for several types of sound events, taxonomic groups and environments; and exemplify its application. This tool facilitates the creation and usage of trained networks and was developed with a strong focus on graphical user interfaces to be used by non-specialist scientists in statistics and programming.