Google details DolphinGemma, a new 400M-parameter LLM to decode dolphin communication by analyzing the vocalizations of wild Atlantic spotted dolphins

DolphinGemma, a large language model developed by Google, is helping scientists study how dolphins communicate — and hopefully find out what they're saying, too.

The Keyword 2025-04-14

Context & Ripple Effects

DolphinGemma extends Google’s model work into a narrowly defined scientific use case: interpreting recordings from wild Atlantic spotted dolphins rather than serving general-purpose text tasks. It sits alongside the company’s broader Gemma effort, which later expanded into translation-focused open models and models aimed at reasoning and agentic work.

The significance is not the parameter count alone, but the attempt to apply LLM-style pattern modeling to a nonhuman communication dataset, where scientific validation matters as much as model output.

First-order effects

Scientists working with wild Atlantic spotted dolphin vocalizations gain a 400M-parameter model designed to help organize and analyze those recordings.
Google establishes DolphinGemma as a specialized LLM application, shifting the immediate focus from generic language generation to research interpretation of animal signals.

Second-order effects

The project raises the value of high-quality, behavior-linked acoustic datasets: model results will be useful only insofar as researchers can validate vocal patterns against observed contexts.
It gives adjacent bioacoustics research a concrete example of task-specific AI, while making evaluation methodology—not merely model scale—the likely point of differentiation.

Third-order effects

If similar systems can be rigorously validated, AI development could increasingly split between broad foundation models and smaller domain models built around scarce, expert-curated scientific data.
This points toward a wider shift in which progress claims for specialized AI are judged by reproducible scientific interpretation rather than benchmark performance alone.

The trend: DolphinGemma is one instance of foundation-model techniques being adapted into compact, domain-specific systems for scientific analysis.

Discussion

@timkellogg.me Tim Kellogg on bluesky
GemmaDolphin: an LLM for dolphins — the weirdest cool thing — they fine-tuned Gemma. No text, instead they're pairing audio & video — sounds to dolphin movements — they've been able to identify key structures, like dolphin names and “battle cries” — blog.google/technolog…
@gusthema Gus on bluesky
🐬 + 💎 = 🤯 — I Knew Gemma models were powerful for human communication, but never guessed they could help with interspecies communication when I joined the project! — DolphinGemma is kind of like auto-complete for 🐬 audio! — Mind blown! — blog.google/technology/a...
@googledeepmind @googledeepmind on x
Built using insights from Gemma, our state-of-the-art open models, DolphinGemma has been trained using @DolphinProject's acoustic database of wild Atlantic spotted dolphins. It can process complex sequences of dolphin sounds and identify patterns to predict likely subsequent [ima…
@googledeepmind @googledeepmind on x
Meet DolphinGemma, an AI helping us dive deeper into the world of dolphin communication. 🐬 [video]
@sundarpichai Sundar Pichai on x
Introducing DolphinGemma, an LLM fine-tuned on many years of dolphin sound data 🐬 to help advance scientific discovery. We collaborated with @dolphinproject to train a model that learns vocal patterns to predict what sound they might make next. It's small enough (~400M params)
@osanseviero Omar Sanseviero on x
Introducing... DolphinGemma! 🐬 🐬Next dolphin token predictor model (audio-to-audio) 🤏Just 400M parameters - can run directly in phones! 🏝️Fin-tuning possible for different species 🌊To be released for the next dolphin season https://blog.google/...

Chronicles