The strangest thing about some new research on AI whale sounds is where the technology learned to listen. Google researchers took a model trained mainly on birds and other animals on land and tested whether it could help classify underwater calls. Despite having little marine audio in its original training, the model proved useful in new whale-sound tasks. It is an unexpected example of how a method for studying one part of nature may help scientists investigate another.
In a February 2026 research account, Google described using its Perch 2.0 bioacoustics model for marine analysis. The project did not teach a computer to speak whale or decode animals' thoughts. It examined whether patterns learned from one set of animal sounds could provide a useful starting point for recognising another. That apparently modest distinction matters. Wildlife researchers often have far more recordings than human specialists can listen to, and sorting them can be difficult even when the sounds are already known.
Why AI whale sounds are such a difficult problem
Underwater sound can travel through places that are almost impossible for people to inspect visually. Whales may vocalise while submerged, moving or travelling far from a research vessel. Hydrophones can record their acoustic environment for extended periods, gathering material that may include animal calls, ship noise, changing weather and many other sounds. The challenge is not merely recording audio. It is recognising which part of a complicated soundscape belongs to the animal a scientist wants to study.
A human specialist can identify familiar patterns, but checking a large archive by ear is slow and demands concentration. Automatic systems can help flag passages for closer attention. They may also make it easier to compare observations across times and locations. The risk is that noise, unfamiliar calls or changes in recording equipment can produce false matches. A reliable result therefore needs validation against carefully identified examples, not only a confident computer classification.
Listening is particularly valuable for species that are hard to observe directly. A whale may surface briefly, while a recording can capture activity that occurs out of sight. Acoustic evidence can inform questions about presence and movement, but it does not automatically tell researchers the number of animals present or the reason they made a particular sound. The interpretation still depends on the species, recording conditions and scientific context.
What birds have to do with the ocean
Perch 2.0 was developed as a model for biological sounds. Its training included many vocalising species, with a strong emphasis on terrestrial animals. A model of this kind learns useful representations of audio patterns: features that can help distinguish one sound from another. Researchers can then use those representations as the starting point for a more specialised classification problem, even when the new environment differs substantially from the original one.
Google reported that the approach transferred well to marine acoustic work despite the limited underwater training material. The researchers examined how a model trained for fine-grained recognition of land-animal vocalisations could assist tasks involving whales. This is known as transfer learning, though the everyday idea is simpler. A system that has learned how to separate complicated natural sounds may have acquired skills that remain useful beyond the animals it first heard.
That does not mean bird calls and whale songs are the same. The frequencies, habitats, recording conditions and biological uses of sound can be radically different. Successful transfer is noteworthy precisely because those differences exist. Specialists must still check whether the model works on the actual marine recordings they want to analyse. An impressive experiment does not establish identical reliability across oceans and whale populations.
A mysterious noise shows why confirmation matters
There is a real-world precedent for machines helping people investigate strange underwater sounds. The US National Oceanic and Atmospheric Administration reported the identification of a peculiar call nicknamed the biotwang, associated with Bryde's whales in the western North Pacific. The sound had been detected earlier without a confirmed species match. Researchers later brought together visual observations and acoustic evidence to establish the source.
The NOAA account explains how collaboration with Google AI helped search extensive recordings and study when and where the calls occurred. Crucially, the species identification was supported by direct scientific observations, not merely an algorithm assigning a mysterious audio clip to the most convenient animal. That is the difference between an intriguing prediction and a credible biological finding.
The biotwang case also illustrates why such work attracts public interest. There are sounds in the ocean that humans record long before they understand who makes them. A computer can help find repetitions and patterns that would take considerable effort to identify manually. But the resolution of a mystery still involves evidence about the animals themselves. Technology helps organise the clues; researchers must determine what those clues establish.
This is classification, not a whale dictionary
Popular discussion of animal AI can jump quickly from recognising a call to translating communication. Those are separate scientific goals. A model that detects a particular kind of whale sound may tell a researcher when similar calls occur in another recording. It cannot on that basis explain whether a whale is calling to a companion, coordinating movement or responding to a disturbance. The functional meaning of a call may require behavioural observations that the audio alone does not provide.
The research also raises questions about dataset bias. Recording archives are often uneven: some areas and species have been studied extensively while others remain poorly represented. A model trained on easy examples may struggle with recordings affected by ship noise or an unfamiliar acoustic environment. Researchers need to ask what happens when the system encounters a species, location or sound that it has not previously been shown.
A practical system should be able to report uncertainty, allowing human experts to focus on ambiguous material. Misclassifying an unfamiliar call as a familiar one could be particularly damaging if it causes scientists to overlook something rare. A model that is good at identifying known patterns should not be mistaken for a complete map of everything the sea can sound like.
Why this matters beyond whales
The wider benefit is a way to handle ecological information at a scale that people alone struggle to review. LiveAIWire has described AI approaches to identifying orcas and birdsong recognition used by wildlife observers. Those applications have different purposes, but all confront the same practical bottleneck: a recorder or sensor can gather much more evidence than a specialist can evaluate continuously.
Citizen observations and other sources of natural-history data can add context. Our reporting on birdwatchers' observations and animal interactions has explored how human knowledge can help AI-assisted ecological studies. The strongest approach is often a partnership between an automated search tool and people who know the animals and habitats involved. Neither the machine nor the observer has to do every job alone.
For conservation, more efficient listening could help researchers notice where certain calls are becoming more or less frequent and decide where to direct additional fieldwork. Yet the step from calls to population estimates is not automatic. Some individuals vocalise more than others, recording equipment has limits and environmental noise changes. Any ecological conclusion needs an appropriate study design and independent checks.
Recorded sound is also vulnerable to mistakes in the way a camera image is. Boat noise, overlapping calls and changing water conditions can make a recording difficult to interpret. A model that flags a candidate whale sound can narrow the search, but a specialist may still need to consider the location, time, species and quality of the original audio. Repeated matches in similar conditions may strengthen a hypothesis without proving exactly what the animal was doing.
There is a practical benefit in making the search itself repeatable. Researchers can revisit older collections as methods improve, compare recordings from different sites and document why a particular sound was selected for closer attention. That kind of careful archive work is less dramatic than announcing a machine that 'talks whale', but it is much closer to what the evidence supports.
The next mystery may already be in an archive
The appealing feature of transfer learning is that it may reduce the amount of specialised labelled data needed to begin a new investigation. Scientists can try a general sound model, evaluate it against a small set of trusted marine examples and improve it for a particular research question. That is potentially useful when collecting more examples would be expensive or when a species is rarely encountered.
There are obvious dangers in overclaiming. A computer trained on bird sounds has not thereby developed an instinct for whale behaviour. It has learned mathematical patterns that happen to be useful in particular acoustic comparisons. Those patterns may assist science, but they do not remove the requirement to confirm what the sounds mean in the world outside the computer.
The interesting development is that a tool originally intended for the sounds of land animals could help scientists examine the oceans. Much of the evidence may already be waiting in long recordings made years ago. AI cannot guarantee an answer, but it can help researchers find which few seconds deserve a closer listen.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.