Applications of semi-supervised learning on soundscape Ecology

Abstract

The study of ecosystems and species from sound recordings collected over time enabled by Soundscape Ecology is a promising strategy for autonomously monitoring natural environments at scale, due to its low implementation cost and ease of collecting large amounts of data. However, the high cost of labeling the recordings is a challenge for expanding the practical application of these techniques. The analysis of collected data is often possible only on a small subset of the data available, i.e. those instances for which labels are assigned by specialists, for instance indicating at which moments in the recordings a particular species can be heard. In this context, Semi-Supervised Learning has been effectively applied to similar problems, such as Speech Recognition, in which data collection is again easier than data labeling. Our goal is to adapt and develop techniques for Self and Semi-Supervised Learning for application in Soundscape Ecology data so that it becomes possible to use the large amounts of unlabeled data available to improve the effectiveness of learned models.(AU)