I spent three days last week at a marine acoustics workshop in Townsville. Twenty-something researchers, a few of them PhD students, a few of them long-time hydrophone operators who’ve been listening to the Great Barrier Reef since the 1990s. The whole event was, in one sense, a reckoning. The amount of underwater audio being recorded across Australian reefs and coastal waters has grown roughly a hundredfold in the past decade. The number of trained ears available to listen to that audio has not.

This is where AI comes in, and where the conversation gets both genuinely exciting and a bit complicated.

What we’re actually recording

A reef makes a remarkable amount of sound. Snapping shrimp create the constant crackling backdrop you hear in any tropical underwater recording. Damselfish make pops and chirps to defend territory. Parrotfish scrape coral with audible rasps. Larger predators make their own signatures. And during coral spawning events, the soundscape changes in characteristic ways that can be measured.

Healthy reefs sound different from degraded ones. This is established science, dating back to work by AIMS and collaborators in the 2010s. Degraded reefs are quieter, with fewer species and less acoustic variety. A 2017 study showed that the sound of a healthy reef can actually attract settling fish larvae to a degraded area — which has interesting restoration implications.

Knowing this and being able to act on it at scale are different things. Until recently, “analysing” a reef recording meant a human researcher listening to hours of audio with headphones, manually tagging species and behaviours. A single 24-hour recording could take a week of person-time to fully annotate. Multiply by hundreds of recording sites and continuous monitoring, and the bottleneck becomes obvious.

The machine learning shift

In the past five years, marine bioacoustics has gone through what computer vision went through in the early 2010s. Models trained on labelled audio datasets — fish calls, whale calls, shrimp clicks — can now identify species and behaviours in raw hydrophone recordings with accuracy that often exceeds first-pass human annotation. The models aren’t perfect. They miss things. They invent things. But they’re fast, they don’t get tired, and they don’t bring their human preconceptions to the data.

A research group at JCU has been running a long-term acoustic monitoring project across multiple GBR sites with a custom-trained model that flags interesting acoustic events for human review. The human researcher’s job has changed: instead of listening to everything, they review the model’s flagged segments. Productivity is up by an order of magnitude. They’re catching events — late-night spawning, rare visitor species — that would never have surfaced in traditional sampling.

I worked briefly last year with Team400 on a related question — how to take research-grade audio models built for one site and adapt them quickly to another site with different ambient noise, different species composition. The challenge is what’s called “domain transfer” in the AI world. A model trained on Heron Island data does not just work, off the shelf, in the Whitsundays. The shrimp soundscape is different. The fish community is different. The boat noise is different. Adapting requires both fresh labelled data and someone who understands what acoustic features matter biologically. That second piece — the biology — is what keeps the work honest.

What we’re finding

A few things have shown up in the analysis that surprised us.

Diel patterns are shifting. The daily sound cycle on the reef — quiet at midday, loud at dawn and dusk, different at night — has been measured for decades. On warming reefs we’re seeing the dusk chorus starting earlier and the midday quiet period lengthening. Whether this reflects fish stress, behavioural adjustment to higher water temperatures, or species composition changes is still being worked out. See the BBC’s coverage of related research from the Caribbean for context.

Reef restoration sites are starting to sound healthier. A few coral restoration projects on the southern GBR are now four to six years old, with hard coral cover returning. Their acoustic signatures, compared against control sites, are measurably different — closer to healthy-reef profiles. This is not a one-to-one indicator of ecosystem recovery, but it’s a useful complementary metric to visual surveys.

Boat noise penetration is worse than we thought. Vessels are noisy underwater, especially small fast boats. Long-term recordings from sites that were considered “remote” show meaningful boat-noise contamination during daylight hours, particularly on calm days. This affects everything — fish behaviour, larval settlement, predator-prey acoustic detection. We’re going to need to think more carefully about how marine park zoning interacts with the underwater noise environment, not just the visible-surface activity.

The limits I keep coming back to

A model can tell you a parrotfish is in the recording. It cannot tell you why. It cannot tell you whether the parrotfish is feeding, spawning, fleeing, or simply passing through. It cannot tell you, on its own, whether the patterns it detects are biologically meaningful or just statistical artefacts of how the model was trained. That work — interpretation — is still a job for a researcher with a doctoral education and field experience.

The risk in this field, as in any field where AI suddenly makes data analysis fast and cheap, is that we mistake speed for insight. We end up with mountains of beautifully annotated audio and very little time spent in the water actually understanding what fish are doing and why. I worry about this. The early-career researchers I spoke to in Townsville worry about this. The model is a tool. The understanding is the work.

What I’m doing next

My team is starting a small project this winter to deploy three hydrophones at a single mid-shelf site for six months continuous recording. We’ll be running the audio through a community-trained model and also having two researchers (myself included) listen to randomly sampled segments to check the model’s accuracy. We expect to find disagreements. We expect to learn from them. We expect to publish the disagreements, not just the successes, because that’s how the field moves forward.

There’s an excellent Wikipedia overview of marine bioacoustics if you want background. If you want to actually contribute, several Australian universities are now running citizen-science acoustic projects where amateur listeners can help annotate clips and improve the training data. It’s slow, satisfying work. The reef is talking. We’re just learning, finally, to hear at scale.

Dr Sarah Winters