Developed in partnership with the SETI Institute and marine biologists, the project aims to build the first systematic tool for analysing dolphin communication using a Large Language Model (LLM) architecture, similar to that used in human text-based AI systems.

Training data

DolphinGemma was trained on the Dolphin Communication Project’s extensive archive of dolphin vocalisations, alongside environmental and behavioural data gathered from bottlenose dolphin populations. The model attempts to map vocal signals to observed behaviour, with researchers hoping to identify patterns that may point to structured communication or symbolic reference.

The approach mirrors the one used in decoding human languages—vector representations of sounds analysed to detect patterns, possible meanings and conversational structures. Google claims that DolphinGemma can now cluster certain sounds with observed social contexts, such as group formation or play behaviour.

Caveats

While the project has drawn public interest—and headlines touting an AI “translator” for dolphins—researchers involved have urged caution. No formal, peer-reviewed studies have yet confirmed that the model accurately interprets dolphin communication, and key questions remain about whether dolphin vocalisations contain anything resembling human language structure.

According to the team, DolphinGemma is intended as a research tool, helping marine scientists uncover structure and statistical patterns in dolphin vocalisations, not to draw conclusions about meaning or syntax.

There is also concern in the broader research community about oversimplifying animal cognition and overstating the capabilities of current AI models in interpreting non-human communication systems.

Open access for scientific use

DolphinGemma is being shared with a few research teams for testing and validation. Its open-source framework is intended to promote further collaboration in the fields of marine biology, bioacoustics and artificial intelligence. While decoding dolphin speech remains speculative, the goal is to accelerate cross-disciplinary collaboration in the emerging field of computational bioacoustics.