91ɫ��Ƭ

Turning one call into thousands

Traditionally, researchers comb through recordings, marking each call by hand. Automated methods have been developed, but the large, labelled training datasets they require simply don’t exist for rare or hard-to-record species.

To solve the problem, Jancovich and the research team built an automated detector system – based on an existing system originally trained to detect human speech – capable of scanning vast audio archives for whale calls.

Deep learning was a powerful tool for this task, though tricky to implement.

“You need loads of training data and you need loads of compute power,” Jancovich says. “That’s a major limitation in some applications.”

So, instead of collecting thousands of training recordings, the researchers used a single example of a blue whale song to generate a training dataset comprised of thousands of semi-synthetic songs.

The method worked by copying the original call and applying modifications such as pitch shifting, time stretching and embedding different types of background noise.

“These modifications are representative of natural variations in the animals’ vocal behaviour,” Jancovich says, “as well as what happens to sound as it propagates through the ocean.”

When tested on real-world recordings, the detector trained on this data performed at a level comparable to detectors trained on far larger datasets.

For one pygmy blue whale population, it correctly detected 99.4% of calls.

“The surprising outcome is that a relatively simple data augmentation process enables really good performance from that one single training example,” Jancovich says.

“You would think you’d need more data, more variation – but because these animals produce sounds that are so similar to one another, it works.”

The result is a large, realistic dataset that can be used to train a detector.

Why blue whales?

The approach relies on the consistency of the target animal’s vocalisations.

Blue whales, the largest animals on Earth, are difficult to study. They are endangered, widely dispersed and spend most of their lives underwater. But they also produce highly stereotyped calls – meaning individuals within the same population make almost identical sounds.

“For example, all the blue whales that live around Madagascar sing the same song, and all the ones near Antarctica sing a different song,” Jancovich says.

This predictability makes it possible to model realistic variation from a single call, but it also means the new method is limited to animals that produce stereotyped calls.

“It wouldn’t work for something like dolphins, where every individual has its own unique whistle.”

A lighter footprint

Deep learning is a technique within the field of machine learning that uses deep neural networks, which are the foundation that AI models like ChatGPT are built on.

“Training large, deep neural networks can consume considerable amounts of electricity,” says Jancovich. “So one of our aims was to do something that’s very compute-efficient.”

The result is a model that can be trained on a standard laptop in a matter of hours, rather than weeks.

“It’s a smaller model – and we’re fine-tuning something that’s already been trained, so it doesn’t need as much data, as much training time or compute power.”

Unlocking decades of data

Around the world, vast archives of underwater recordings have been collected through passive acoustic monitoring – hydrophones left in the ocean to record continuously.

But without efficient detection tools, much of that data remains underused.

“The raw data is quite hard to exploit because researchers don’t always have access to high-performance automated detection methods, or the necessary training data and compute resources,” Jancovich says.

The next step for the research team is to apply the detector to a 25-year dataset from the central Indian Ocean, tracking long-term changes in blue whale song.

It may also open new windows into animal behaviour. Whale songs, for example, are not just signals.

“They help us study things like animal culture – the way animals learn songs from each other across generations,” Jancovich says.

And the same technique could be applied to detecting and monitoring other species – from birds to insects – that produce consistent, repeatable calls. Ecologists already have microphones placed not just in the oceans, but in forests and other remote environments.

“If accurate detectors can be trained from a single good recording, this can help us study rare and elusive species that have seldom been heard by humans,” says Jancovich.