Strange “biotwang” ID’d as Bryde’s whale call

Multi-species machine-learning model can unlock insights into new whale species.

In 2014, researchers monitoring acoustic recordings from the Mariana Archipelago picked up an unusual whale vocalization with both low- and high-frequency components. It seemed to be a whale call, but it sounded more mechanical than biological and has since been dubbed a "biotwang."

Now a separate team of scientists has developed a machine-learning model to scan a dataset of recordings of whale vocalizations from various species to help identify the source of such calls. Combining that analysis with visual observations allowed the team to identify the source of the biotwang: a species of baleen whales called Bryde's (pronounced "broodus") whales. This should help researchers track populations of these whales as they migrate to different parts of the world, according to a recent paper published in the journal Frontiers in Marine Science.

Marine biologists often rely on a powerful tool called passive acoustic monitoring for long-term data collection of the ocean's acoustic environment, including whale vocalizations. Bryde's whale calls tend to be regionally specific, per the authors. For instance, calls in the eastern North Pacific are pretty well documented, with frequencies typically falling below 100 Hz, augmented by harmonic frequencies as high as 400 Hz. Far less is known about the sounds made by Bryde's whales in the western and central North Pacific, since for many years there were only three known recordings of those vocalizations—including a call dubbed "Be8" (starting at 45 Hz with multiple harmonics) and mother-calf calls.

That changed with the detection of the biotwang in 2014. It's quite a distinctive, complex call that typically lasts about 3.5 seconds, with five stages, starting at around 30 Hz and ending with a metallic sound that can reach as high as 8,000 Hz. "It's a real weird call," co-author Ann Allen, a scientist at NOAA Fisheries, told Ars. "Anybody who wasn't familiar with whales would think it was some sort of artificial sound, made by a naval ship." The 2014 team was familiar with whale vocalizations and originally attributed the strange sound to baleen whales. But that particular survey was autonomous, and without accompanying visual observations, the scientists could not definitively confirm their hypothesis.

That's where Allen and her co-authors came in. Allen's research involves conducting visual and acoustic surveys of US territories and islands with the goal of estimating the health and status of the whale populations in those regions. Her team was surveying the Mariana Archipelago in 2018, incorporating visual observervations.

Although Bryde's whales can be challenging to identify, those observers spotted 10 different Bryde's whales over the course of the survey. They also collected acoustic recordings of the biotwang for nine of those 10 instances. "Once is happenstance, twice is coincidence," said Allen. "Nine times, it's definitely a Bryde's whale."

A call and response?

The exact purpose of the biotwang sound among whales is still unknown. "With baleen whales, vocalizations can serve several different purposes," said Allen, citing breeding calls, feeding calls, and contact calls as some examples. Allen said the biotwang might be a contact call, comparing it to a "Marco Polo" call-and-response among Bryde's whales. But it will take more research to make a conclusive determination on that score.

Key to this research was the development of a machine-learning model to sift through the gigantic dataset of what is currently over 200,000 hours of recordings. Allen confessed that she was initially overwhelmed by the magnitude of the dataset until a friend who worked for Google put her in touch with co-author and Google data scientist Lauren Harrell. Initially the model was designed to recognize the songs of humpback whales. But the biotwang was an entirely new whale sound, so Harrell and Allen had to figure out how to integrate it into what has become an open source multi-species model.