DolphinGemma: Google wants to communicate with dolphins

DolphinGemma is a kind of large language model for dolphins. It is used by a dolphin organization via Pixel smartphones.

listen Print view

(Image: Google)

3 min. read

April 14 is National Dolphin Day in the USA and Google has announced that it is working on an AI model that will enable people to communicate with the animals. DolphinGemma was developed in collaboration with Georgia Tech and the Wild Dolphin Project (WDP). The model learns structures from the sounds of dolphins and can generate them.

The WDP provides the knowledge about dolphins. This includes social behavior, for example, but also initial findings about sounds. For example, dolphin mothers and calves come together through whistles that only they use, writes Google. There is a type of screeching during arguments and a clicking hum during courtship and pursuit by sharks. Google has now used the data from the WDP to train an AI model.

A SoundStream tokenizer is used to represent the sounds. The model architecture is optimized for complex sequences, writes Google. With 400 million parameters, DolphinGemma can run on the Pixel smartphones of WDP employees. Basically, the model works in the familiar way of a large language model, except that animal sounds are used instead of human language. The AI learns patterns and structures and calculates probabilities for the output.

Videos by heise

"By identifying recurring sound patterns, clusters and reliable sequences, the model can help researchers uncover hidden structures and potential meanings in the natural communication of dolphins – a task that previously required immense human effort." It is already possible, for example, to point out objects to dolphins that the animals like to play with.

The collaboration has also resulted in another system that is used to communicate with dolphins. CHAT stands for Cetacean Hearing Augmentation Telemetry. It is an underwater computer with which a common but simple vocabulary is to be developed. It also generates synthetic whistles, but they sound different from those of dolphins. The whistles are then associated with objects, i.e. the dolphins are classically conditioned. The researchers now also hope that the dolphins will imitate the sounds at some point in order to demand the associated objects –, such as seaweed. Pixel devices from Google are also used for CHAT.

DolphinGemma is to be made available as an open model this summer. Although the model has been trained on the sounds of dolphins in the Atlantic, Google expects that it can also be used with other dolphin species – by means of fine-tuning, for example. The blog post also states that the aim is to narrow the gap between human and dolphin communication. Of course, the overarching goal is cross-species communication, not just with dolphins.

(emw)

Don't miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.