Active listening and learning for orca sound detection

Abstract

Southern Resident Killer Whales are in danger of extinction, however, the current approaches to detect and study their habitat are very time and labor consuming. To streamline this process we have built an active learning system leveraging the vast streams of passive acoustic data collected by hydrophones, the advances in machine learning for audio analysis, and the expertise of trained bioacousticians. The system visualizes predictions of a machine learning algorithm on raw data and allows input from the user to verify and correct them, thus supplementing the existing training dataset and also engaging the user in the annotation process. Researchers can focus on labeling only observations for which the algorithm has high uncertainty, while most of the data get labeled automatically. We show that the active learning system improves the performance of a Convolutional Neural Network algorithm on a two-category classification problem (presence/absence of a call) even with very few extra annotations: the f1-score increased from .83 to .84 with 50 new annotations (∼3% increase of the labeled dataset). We have designed the system flexible to incorporate other algorithms and facilitate result comparison within the community. It is also open source and easy to deploy on diverse computing infrastructures.