In the murky depths of the South China Sea, an artificial intelligence model originally trained to recognize objects in images has learned to hear. Researchers led by Zhuo Xiao of Guangxi Minzu University have repurposed a foundation AI model called Segment Anything Model (SAM) to detect Bryde's whale calls hidden within seismic data—and it's working with remarkable precision. The discovery, published in Seismological Research Letters, offers a new tool for understanding one of the ocean's most elusive creatures without ever having to chase them down.
Whale population monitoring has long been a challenge for marine biologists. Traditional methods require researchers to physically encounter these animals or deploy expensive equipment, making large-scale observation difficult and intrusive. But whale calls carry a signature: when converted into spectrograms—visual representations of sound frequencies over time—they appear as clear, repeating patterns. This insight was the key that unlocked a fresh approach. "That turns call detection into an image segmentation task, which SAM excels at," Xiao explained. The AI model, which was trained on massive generic image datasets but never specifically on whale calls, could identify these acoustic patterns with 96% accuracy when analyzed seismic data from a single station on Xieyang Island in the Beibu Gulf, a shallow-water feeding ground for Bryde's whales.
What makes this achievement particularly striking is that the AI caught calls that human analysts had missed. When Xiao's team examined seismic recordings from January 26 and July 11, 2021, the model identified vocalizations the trained eye had overlooked. The researchers also discovered something unexpected in the data: seasonal variation in how the whales call. The time interval between acoustic pulses—known as the interpulse interval—was shorter and more intense in winter, suggesting the whales were coordinating more actively with one another. In summer, the calls were more spread out, indicating more solitary vocalization. This window into whale behavior would have been invisible using older detection methods.
The implications ripple far beyond a single research station. Accurate detection of baleen whale calls lets scientists monitor population sizes, track seasonal movements, and detect behavioral changes—all without the noise and disruption of traditional observation. Passive acoustic monitoring via seismometers and hydrophones is noninvasive and scalable, meaning the technique could eventually be deployed across ocean basins. To prove the method's versatility, Xiao and colleagues tested their AI on fin whale recordings from Ireland and blue whale recordings from Canada, and the results held strong. "They generalize remarkably well to new domains without major workflow changes," Xiao noted.
The technique isn't perfect. False positives and missed calls still occur, and disentangling whale calls from ocean noise remains challenging. But Xiao's team is already planning the next phase: combining seismic data with acoustic sensors, ocean-bottom seismometers, and fiber-optic distributed sensing arrays to further sharpen detection. They're also working to fine-tune foundation models specifically for cetacean calls. As our relationship with the ocean deepens and whale populations face mounting pressures from shipping, climate change, and fishing, new ways to listen to them—and understand what they're telling us—could prove invaluable.