WIRED reported on November 5, 2026, that Denise Herzing described work with Georgia Tech and Google DeepMind to find patterns in dolphin sounds, alongside limited attempts at two-way sound interaction. Google describes DolphinGemma as predicting likely next sounds, while its separate CHAT effort uses synthetic whistles linked to objects. These approaches study communication; neither amounts to a demonstrated translation of dolphin meaning.
What DolphinGemma does with dolphin sounds
Google announced DolphinGemma on April 14, 2025, as an audio-in, audio-out model trained on the Wild Dolphin Project’s recordings of wild Atlantic spotted dolphins. It uses SoundStream to represent sounds and predicts likely continuations of a sound sequence. Google gives its size as about 400 million parameters and says it is sized to run on Pixel phones used in the field. These specifications describe the announced model, not a measure of its accuracy.
The Wild Dolphin Project has studied wild Atlantic spotted dolphins in the Bahamas since 1985. Its records pair underwater audio and video with individual dolphins, their life histories and observed behavior. That context gives researchers ways to study sounds alongside what dolphins are doing; a predicted next sound still does not, by itself, identify what a vocalization means.
DolphinGemma and CHAT take different approaches
Google describes DolphinGemma as analyzing sequences of natural dolphin sounds. CHAT, or Cetacean Hearing Augmentation Telemetry, is a separate underwater effort that uses synthetic whistles associated with objects to try to establish a simpler shared vocabulary.
| System | Signal and task described | Research aim |
| DolphinGemma | Natural dolphin sounds; finds patterns and predicts likely next sounds | Analyze sound sequences |
| CHAT | Synthetic whistles associated with objects | Try a simpler shared vocabulary |
Google says DolphinGemma may help CHAT identify potential mimics. That is a possible support role between the systems, not evidence that either one translates natural dolphin communication.
Why copying a sound does not prove understanding
In the experiment WIRED described, an underwater wearable computer played sounds associated with objects such as a scarf and sargassum seaweed. Two researchers modeled the sound-and-object association and observed the dolphins’ responses. The account describes attempts at interaction, not a completed conversation.
A dolphin repeating a signal could be mimicking it without understanding what it refers to or what function it serves. Herzing identified that distinction as especially difficult to assess in the wild.
A predicted next sound describes a pattern in a sequence; it does not, on its own, translate a dolphin’s meaning.
Google says adapting DolphinGemma to other cetacean species may require fine-tuning, so its stated training scope remains the wild Atlantic spotted dolphins in the project’s Bahamas dataset.
