Neural Network Trained on a Single Call Detects Blue Whale Songs Across Ocean Basins

by Simon Mansfield
Sydney, Australia (SPX) ) May 01, 2026
Researchers at UNSW Sydney have trained a neural network to detect blue whale songs in decades-long acoustic recordings using just a single example call, achieving an accuracy of 99.4 percent across real-world datasets spanning multiple ocean basins. The study, led by PhD candidate Ben Jancovich, is published in Scientific Reports.

The work adapts a speech-detection neural network architecture to the problem of marine bioacoustic monitoring. Rather than relying on thousands of labeled whale recordings - the conventional requirement for machine learning models in this domain - the team generated a large synthetic training dataset from one high-quality blue whale call.

Semi-synthetic augmentation techniques including pitch shifting, time stretching, and noise embedding were applied to mimic the natural variation in whale vocalizations and simulate the effects of ocean sound propagation. The model was trained on a standard laptop computer in a matter of hours.

"Machine learning models traditionally need to be trained on thousands of recordings of the very whale song that they're trying to find," Jancovich said. "However, this new model was trained on only one recording of a blue whale call."

The resulting detector matched the performance of models trained on much larger datasets. It was validated against recordings from the Indian Ocean, including a 25-year archive previously too large and labour-intensive to analyse manually. The model was optimised for stereotyped, repetitive vocalizations - such as those produced by blue whales near Madagascar and Antarctica - and the researchers note it is not suited to highly variable calls such as dolphin whistles.

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

The implications extend well beyond blue whales. The approach could be adapted for other species with consistent call structures, including birds and insects, and could enable monitoring of rare or elusive animals that have rarely been recorded. The researchers argue the method opens up vast archives of underused ecological data to systematic analysis.

"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," Jancovich said.

Research Report:One whale song unlocks oceans of data