SurfPerch: AI Listens to Coral Reefs to Track Their Health

Coral reefs cover less than one percent of the ocean floor yet support roughly a quarter of all marine species. They also shield coastlines, sustain fisheries, and anchor tourism economies across the tropics. But reefs are under siege from warming waters, acidification, and overfishing, and scientists have long struggled to measure their health at scale because divers can only survey small areas, in daylight, in clear water.

SurfPerch, an AI model created by Google Research and Google DeepMind, offers a different approach: listening. The model automatically processes thousands of hours of underwater audio recordings from hydrophones anchored on reefs, classifying the croaks, crackles, and growls of marine life into a bioacoustic picture of how alive a reef really is. It lets scientists, in the words of the research team, hear reef health from the inside.

Why Sound Reveals What Divers Cannot

A healthy reef is a noisy place. Fish grunt and croak, snapping shrimp produce crackles loud enough to be heard from the surface, and the collective hum of activity rises and falls with the biological richness of the ecosystem. A degraded reef, by contrast, falls quiet long before it looks dead to the eye.

Sound also goes where surveyors cannot. It works at night when many reef species are most active. It penetrates deep and murky waters where visibility ends. And because hydrophones record continuously for months, they capture changes over time that a snapshot dive survey would miss entirely. The catch has always been analysis. A single sensor can generate thousands of hours of audio, and human experts who can identify sounds by ear are rare. That backlog is what SurfPerch was built to eliminate.

A Dataset Built by Citizen Scientists

SurfPerch was trained on ReefSet, a dataset of reef audio annotations drawn from Calling in Our Corals, a project created by Google Arts & Culture in collaboration with marine scientists. The project invited people around the world to listen to short clips of reef recordings online and label the sounds they heard. That crowdsourced effort produced a library of new fish sounds that researchers used to build and refine the model.

The result is a model that can distinguish biological reef noises such as croaks, crackles, and growls from specific groups like damselfish, dolphins, and groupers, and from non biological noise like waves and boat engines. Once deployed, SurfPerch can scan archived audio and flag every instance of a sound type, giving scientists a quantitative read on which species are present and how active the ecosystem is.

An Unexpected Lesson from Bird Song

One of the most surprising findings in the SurfPerch work came from transfer learning, the technique of applying knowledge from one domain to another. The team found that adding bird song to the model's training mix actually improved its performance on reef sounds. Despite sounding very different, there were enough common acoustic patterns between bird song and fish sounds for the model to learn from one and get better at the other.

Fine tuning proved powerful too. When researchers fed the model just a handful of examples of a newly discovered fish sound, its ability to detect that sound across thousands of hours of audio improved by 25 percent. That means field scientists who record a new sound once can teach the model to find every other occurrence in the archive, turning isolated observations into ecosystem wide data.

What the First Trial Found

The first research trial combining Calling in Our Corals data with SurfPerch produced findings across three oceans. In the Philippines, the acoustic analysis revealed measurable differences between protected and unprotected reefs, giving conservation managers evidence that marine protected areas change more than fish counts. In Indonesia, the model helped characterize restoration outcomes, showing how replanted or rehabilitated reefs sound compared with untouched ones. And on the Great Barrier Reef, the acoustic data exposed relationships within the fish community that traditional surveys had not captured.

For restoration practitioners, this matters because success is notoriously hard to prove. A transplanted coral fragment may survive, but whether the reef around it is recovering as a living ecosystem is a deeper question. Soundscapes offer a continuous, sensitive indicator of that recovery.

From One Model to a Family of Ears

SurfPerch did not remain a one off. In 2025 Google DeepMind released Perch 2.0, a bioacoustics foundation model trained primarily on birds and other terrestrial animals, and found that it also performed well on underwater tasks, including classifying whale vocalizations and reef sounds, despite never hearing underwater audio during training. A follow up paper presented at a NeurIPS 2025 workshop showed that Perch 2.0 could create useful custom classifiers for marine sounds from as few as four labeled examples per class.

The broader vision is agile modeling: any researcher with a laptop, a small set of labeled recordings, and access to archives like the NOAA passive acoustic collection can build a custom detector in hours instead of months. The tools and tutorials are public, which means small conservation teams in island nations, often the communities most dependent on reef health, get the same acoustic intelligence as large institutions.

Why This Matters Now

The world has lost a significant share of its coral reefs, and mass bleaching events are becoming more frequent as oceans warm. Conservation resources are limited, and they work best when they are aimed at reefs that can still be saved and interventions that demonstrably work. Acoustic monitoring with AI provides exactly that evidence base: continuous, cheap once deployed, comparable across sites, and indifferent to darkness or murky water.

A hydrophone on a buoy. A model trained on citizen science and bird song. A dashboard telling a park manager in the Philippines whether the reef she protects is getting louder, and therefore livelier, year over year. This is what modern conservation monitoring looks like when the ocean itself gets to speak.