Tuning in to the sounds of Canada’s wildlife

A new AI tool that picks out bird and amphibian sounds in audio recordings could improve how ecologists monitor and study Canada’s wildlife.

“HawkEars is a software package that analyzes audio recordings to identify bird and amphibian species, and it is trained on species that occur in Canada,” says Jan Huus, a retired software developer and avid bird watcher who created the tool.

After reading about her research, he connected with ecologist Elly Knight, an adjunct professor in the Department of Biological Sciences, and the two have been collaborating ever since, with support from the Alberta Biodiversity Monitoring Institute.

“These acoustic cues have so much information in them because it’s essentially the currency the birds are communicating in,” says Knight, co-director of the Boreal Avian Modeling Center with Biodiversity Pathways.

For a long time, studying animal populations has been labor-intensive. Researchers would have to trek out into the wild to take photographs and videos, and capture audio recordings of animals. But as Knight explains, “The field of ecology is going through a paradigm shift from human-based approaches of measuring and monitoring wildlife to remote-sensing approaches.”

One such approach, known as passive acoustic monitoring, is becoming the standard method for studying animals that communicate through sound. With this approach, researchers can set up remote-controlled recording devices, allowing them to gather audio across a particular season or even over a full year.

There’s just one issue: this method captures massive amounts of information—”far more than we’ve ever dealt with in the past,” says Knight—that then needs to be processed and sorted through in some way. Previously, researchers would do this by sitting down to listen to the audio recordings while looking at the sound represented on a computer screen, drawing on their expertise to annotate the data and classify it.

“It’s really only efficient to do that for about 1% of what we collect,” says Knight. “AI classifiers like the one Jan created open up doors into what we can do with the data and allow us to tap into that huge amount of information.”

There are a few existing global classifiers—BirdNet and Perch being the most common—but they’re trained on thousands of species from around the world. HawkEars focuses on a smaller set of regional species, so it’s more accurate. The paper introducing the tool is published in the journal Ecological Informatics.

As Huus explains, it’s a bit of a balancing act. “If you train a classifier on too small a group of species, it won’t perform as well, but if you train it on tens of thousands, it also doesn’t perform well. There’s a sweet spot in there somewhere.” For Huus and Knight, the regional scope of HawkEars seemed like an optimal middle ground.

HawkEars’ accuracy “enables us to get a lot more out of the data for some of the finer-resolution questions about ecology and wildlife monitoring,” says Knight.

Some applications for HawkEars include monitoring species at risk, as well as testing for changes in the timing and arrival of bird species in certain regions, which may serve as signs of climate change or habitat quality, says Knight. Huus notes it can also be used to analyze the impact of human sounds on bird environments, measuring the effect of being too close to a highway or industrial site, for example.

Huus is now refining the tool for wider use by separating the Canadian bird classifier from the toolkit of code, allowing others to “take the HawkEars software, take their data and train their own classifier.” A lab in Pittsburgh has already trained a classifier using Central American frog recordings and the HawkEars code.

“We’re excited about what HawkEars can provide in terms of conservation in Canadian landscapes, to facilitate research and monitoring,” says Knight.