Killer whales are apex predators whose vocal behaviours offer insights into their ecology. At the Prince Edward Islands, killer whales’ vocalisations align with seasonal prey availability and social activities, providing a unique research opportunity. This project aims to improve the understanding of killer whale behaviour and its response to environmental changes by analysing an extensive dataset of recordings between 2021 and 2023. The dataset spans 4,200 hours, with 70 hours manually annotated to include over 15,000 annotations. However, the dataset’s complexity, including multi-species recordings featuring both killer whales and humpbacks, required innovative solutions to distinguish between species based on call frequencies.
THE PROBLEM
Manually analysing large acoustic datasets is time-intensive and prone to human errors. For the 4,200-hour dataset from the Prince Edward Islands, annotators had to distinguish killer whale social calls from other sounds, such as noise or humpback whale vocalisations, some of which share overlapping frequencies. These challenges increase the risk of errors, with a small percentage of inaccurate annotations. Without efficient and accurate tools, extracting meaningful ecological insights about killer whales’ social behaviours and environmental adaptations becomes a significant bottleneck.
THE SOLUTION
This project developed an automated detection system to identify five distinct killer whale vocalisation types: upsweeps, downsweeps, tones, squeaks, and clicks. The model achieved impressive accuracy, with less than 5% of predictions being incorrect or misidentified. Notably, many false positives (calls identified but annotated as noise) were actual social calls, while most false negatives (annotated calls misclassified as noise) were genuinely noise segments. Automation has streamlined the analysis process, saving significant time and allowing researchers to focus on ecological interpretation. These results improve our understanding of killer whale vocalisation patterns, benefiting long-term conservation efforts and offering methods to handle complex, multi-species datasets.
THE PROJECT’S TEAM
Danielle Hausler
ML Researcher
Dr Fannie W. Shabangu
Position: PhD, Marine Biologist at the Department of Agriculture
Our people are important to us much as the whales!
WHO WE ARE?
Our people are important to us much as the whales!
Our volunteers are motivated and collaborative and share a passion for solving problems and the sea. We seek individuals inspired by the prospect of developing new technology and driven by the will to make on environmental change. Deep Voice includes five teams: Research, Bio-annotations,Development, Product, and Media.
