Google DeepMind adapts Perch 2.0: a bird-song model that recognizes whale calls

Google DeepMind has found an unexpected way to study the ocean: Perch 2.0, a model built to recognize the songs of birds and other land animals, also handles whale vocalizations reliably. This could cut the time needed to develop separate marine models and speed up acoustic monitoring of rare populations.

How the model was tested

Perch 2.0 is a foundational bioacoustic model trained on millions of recordings of birds, amphibians, insects and mammals. It was originally built not for the ocean but for analyzing terrestrial soundscapes. Google DeepMind and Google Research decided to test whether, instead of building a new system from scratch for whales, they could reuse an existing foundation.

The logic is simple: if a foundation model transfers knowledge between different types of signals, scientists will not have to spend as much compute and time on separate development.

For the test, the team took three marine audio datasets with whale vocalizations and other underwater noise. Each five-second fragment was converted into a spectrogram, a visual map of frequencies and sound intensity over time. Perch 2.0 then turned this data into embeddings, compact sets of features that can distinguish, for example, an orca whistle from a humpback whale signal.

The researchers then trained a simple logistic classifier on just a few examples: from four to 32 embeddings per dataset. Even with so few examples, quality was high and kept rising as data was added.

Why the transfer worked

The key idea is transfer learning. A model first learns general acoustic patterns from a huge body of one type of data, then applies that knowledge to a different but related task.

With Perch 2.0 the transfer looks especially surprising: birds sing in air, while whales communicate underwater. But the model seems to pick up more than the medium the sound travels through: it captures subtler patterns such as the shape of a whistle, frequency dynamics, duration and the microstructure of the signal.

"We train this model to find small details in soundscapes."

The researchers offer several explanations. Birds and marine mammals may have evolutionarily similar sound-production mechanisms. Large models trained on diverse data also often perform well outside their original domain. Finally, recognizing bird vocalizations is itself very hard, so the model is forced to notice minute differences. That probably helps it underwater. According to the team, the whistles of some orca populations even fall in a similar spectrogram range to many bird signals.

Why biologists care

For ocean researchers, this result matters as more than a neat demonstration. Bioacoustics keeps uncovering new signal types, and some underwater sounds still have no confident classification at all. If, instead of building a separate model for each species, researchers can take a powerful foundation model and quickly fine-tune a lightweight classifier, the research cycle gets noticeably shorter. This is especially useful for passive acoustic monitoring, where scientists spend months listening to huge volumes of recordings from buoys, hydrophones and autonomous stations.

• Faster launch of new models for specific whale populations
• Lower costs for training from scratch and architecture search
• Works even with very few labeled examples
• More flexible search for rare and still undescribed signal types

It also matters that Perch 2.0 was compared not only with Google's previous whale model but also with other bioacoustic models for birds, animals and coral reefs. In that comparison it ranked either first or second. So this is not a lucky one-off test but a strong result against specialized alternatives. For conservation projects, it is a good sign: a single foundational audio tool may work across several ecosystems.

What it means

The Perch 2.0 story shows that foundation AI models are starting to benefit not only chatbots and content generation but also field science. If transfer learning works between birds and whales, biologists get a chance to track population health faster, spot changes in animal behavior and better protect vulnerable species.