Researchers have built an AI algorithm that automatically detects sounds and classifies what they are – and when they ran the model, it categorised a sound no one has identified before. The findings are published in PLOS Computational Biology.
Coral reefs are surprisingly noisy places, full of clicks, chirrups and pops. It’s so loud that it can be hard to pick out and identify individual sounds made by reef animals.
“Imagine a crowded room filled with people all talking at once, and you’re trying to pick out words from multiple people,” says the study’s author Daniel Duane, an oceanographer at the Naval Undersea Warfare Center Division Newport.
Many scientists use AI to interpret data – such as recordings of sounds heard in underwater environments – more quickly. For example, machine learning helped marine scientists finally work out that the metallic, spaceship-like noise (dubbed biotwang) they’d heard echoing through the Mariana Trench was the call of a Bryde’s whale.
Often, researchers use what are known as supervised AI algorithms. These are trained on datasets that humans have already categorised before being let loose on the data. “For example, you give the model a thousand examples of a blue whale calls, and it learns to identify other whale blue whale calls from the examples you gave it,” says Duane.
Healthy reefs that are full of diverse animals tend to be noisy, so listening to underwater soundscapes can provide researchers with lots of information about how coral habitats are doing. “It has been shown that degraded reef communities produce a lot less sound,” he says.
But what if you want to eavesdrop on a noisy coral reef full of many unknown and overlapping animal sounds? Several models are unable to identify specific sounds. “For example, you might have a minute-long audio sample with lots of overlapping fish sounds, and the algorithm can identify that sample as being ‘biologically active,’” Duane says. “This can be useful for comparing the health of different reefs, but it doesn’t tell you much about the specific animals that are producing the sounds.”
To try to tease the mish mash of reef sounds apart, Duane and colleagues turned to unsupervised AI. “The algorithm goes through a large dataset and clusters signals based on their similarity,” he explains.
The model ploughed through a year’s worth of data recorded in four different sites in Hawaii. This resulted in “more than seven million detections which were automatically sorted into 29 clusters,” he says.
The scientists listened to some of the clips in each ‘cluster’ at random to identify them. They were hoping it might manage to classify one or two specific sounds but it found nine. These included “two call types from damselfish, feeding sounds from parrotfish, calls from holocentrids (squirrelfishes and soldierfishes), an unidentified fish sound, three separate humpback whale song units, and ship noise,” he says. “The remaining clusters mostly contained unidentified sounds (random “bumps and knocks”).”
The unknown fish sounded very strange. Its calls were “almost like a very deep purr,” he says. Duane thinks this species’ nocturnal activity might be why their noises haven’t been identified before now. “They seem to only vocalise at night, when visual surveys with cameras or divers are not practical,” he says.
- What's the loudest bird? Meet the strange bird that produces a deafening sound
- Why do howler monkeys scream? The secrets behind the loudest monkey in the world's deafening howl
The scientists cross-checked the damselfish vocalisations against videos captured by FishEye Collaborative using a 360° camera and found the model was correct for at least 82 per cent of the calls.
It also uncovered trends within the sounds. “For example, the holocentrids vocalised at sunrise and sunset, with big increases during the full moon,” says Duane. “The unknown fish also displayed a lunar pattern, with calling occurring during nighttime hours when the moon was not in the sky.”
Now, they want to keep working on the model so it can pick up an expanded range of sound types and see how it performs when analysing other reef environments.
They hope that the ability to deconstruct species sounds within the chatter of a reef will help conservationists better understand what’s going on in the ecosystem so they can find ways of protecting it. “Different groups respond to environmental stressors in different ways,” he explains. “For example, a bleached reef will see a decrease in population for most fish but an increase in parrotfish activity.”
More wildlife stories from around the world
- There's a dolphin in Australia that chases fish until they vomit – then eats it
- This diver was exploring the Great Barrier Reef, when he noticed something strange
- Scientists record world's longest underwater timelapse. What is reveals is astonishing
- Underwater robot stalks barracuda for 10 minutes – and discovers this
Top image: Courtship and territoriality sounds of dascyllus albisella. Credit: FishEye Collaborative, under Creative Commons CC BY-NC 3.0.
