Researchers at UNSW Sydney have developed a deep learning model capable of detecting blue whale calls in acoustic recordings with up to 99.4% accuracy, utilizing a single training recording.
The study, led by PhD candidate Ben Jancovich, represents a significant leap in bioacoustics. The tool is open-source and designed to make the analysis of long-term acoustic data—a task currently described as labor-intensive and costly—much more efficient and accessible.
Methodology
The model employs a data augmentation technique that generates thousands of semi-synthetic calls from one original example.
This is achieved by applying pitch shifting, time stretching, and embedding background noise. The neural network is based on an existing system originally designed for human speech detection.
The method relies on the stereotyped nature of blue whale calls; individuals within the same population produce nearly identical sounds. Crucially, training can be completed on a standard laptop in hours, requiring minimal computational resources.
Testing and Accuracy
The detector was tested on real-world recordings and performed comparably to models trained on large datasets.
For one pygmy blue whale population, the model correctly identified 99.4% of calls.
Limitations
The technique is not suitable for species with variable vocalizations, such as dolphins, where each individual has a unique whistle.
Applications
The tool can be applied to long-term acoustic datasets, including a 25-year recording archive from the central Indian Ocean, to study changes in blue whale song.
Similar methods could be used to detect calls of other species that produce consistent, repeatable sounds, including birds and insects.
According to Jancovich, the tool could help ecologists study changes in whale populations over time. Manual analysis of long-term acoustic recordings is labor-intensive and costly, which the new tool aims to make more efficient and accessible.