The DolphinGemma project is the most significant development in cetacean bioacoustics in a decade. Here's what it actually does.
Released by Google in collaboration with the Georgia Tech Research Institute and the Wild Dolphin Project, DolphinGemma is a transformer-based model fine-tuned on bottlenose dolphin vocalizations. It classifies whistle contours, identifies call types, and models temporal patterns in vocal sequences with a degree of precision that was previously impossible without significant compute investment.
What it doesn't do: translate. The jump from "we can classify sounds with high accuracy" to "we understand what dolphins are communicating" is enormous and largely unsolved. DolphinGemma is a measurement instrument, not a decoder. The newsletter covers this distinction carefully because the coverage almost never does.
As a documentary filmmaker working on a film about DolphinGemma's development, I have access to the researchers, the methodology, and the data that doesn't make the press releases. That context informs every issue.
Research archive →