Dolphins may soon have an unexpected interlocutor: artificial intelligence. Google is developing a model capable of deciphering their sound exchanges, opening up unprecedented perspectives on their communication.
This initiative builds on decades of research conducted by the Wild Dolphin Project (WDP). By combining field observations and advanced technologies, scientists hope to unlock the secrets of the clicks, whistles and other sounds emitted by these cetaceans.
DolphinGemma: a cetacean translator
DolphinGemma is based on an architecture inspired by human language models, but adapted to the acoustic specificities of dolphins. The system analyzes sound sequences to identify recurring patterns, with a precision impossible to achieve manually.
Thanks to SoundStream technology, vocalizations are converted into data usable by AI. The model, optimized to work on smartphones, allows real-time analysis during expeditions.
Initial tests suggest that DolphinGemma could predict certain sound sequences. This breakthrough would facilitate the identification of potentially meaningful structures, without proving the existence of a language per se.
CHAT: towards interspecies dialogue
In parallel, the CHAT system explores a more interactive approach. It associates synthetic whistles with concrete objects, such as seaweed or toys, to establish a common vocabulary with dolphins.
Researchers use Pixel smartphones to capture and process sounds underwater. This lightweight solution avoids the need for bulky equipment while offering sufficient performance for experiments in the wild.
The goal is not to translate the complexity of natural exchanges, but to create simple bridges. If dolphins reproduce the learned whistles, it would mark a step towards voluntary communication between species.
How does the model learn?
DolphinGemma trains on thousands of hours of recordings provided by the WDP. This data, annotated with behavioral context, allows the system to link certain sounds to specific situations.
The model uses a method similar to that of voice assistants, but adapted to dolphin frequencies and modulations. Each sound is broken down into analyzable segments, then compared to other occurrences.
This approach reveals similarities imperceptible to the human ear. For example, tiny variations could correspond to different "dialects" among dolphin groups.
What are the limitations of this technology?
There is still no evidence that dolphins use a structured language. The identified sounds could simply reflect basic emotions or intentions, without elaborate syntax.
Underwater noise (waves, boats) also complicates the recording of vocalizations. Algorithms must therefore filter out these interferences to isolate relevant sounds.
Finally, adapting the model to other species will require adjustments. The Atlantic spotted dolphins studied by the WDP do not have the same vocalizations as bottlenose dolphins or orcas.
Article author: Cédric DEPOND