CAN AI DECODE WHALE AND DOLPHIN SPEECH?

For decades, researchers recorded hours of clicks, whistles and codas from whales and dolphins, then ran up against a hard limit: there was simply too much audio for any team to label by ear. Machine learning has changed the scale of what is searchable, clustering millions of sounds and flagging recurring structure that humans then test against behavior. AI is the discovery instrument, not the interpreter.

Thank you for reading this post, don't forget to subscribe!

Project CETI: finding structure in sperm whale codas

Sperm whales communicate in codas, short, patterned bursts of clicks. In a 2024 study, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory working with Project CETI applied machine learning to 8,719 codas recorded from Eastern Caribbean whale families by the Dominica Sperm Whale Project. The analysis identified a combinatorial coding system built from features the team named rhythm, tempo, rubato and ornamentation, with at least 143 distinguishable coda combinations frequently occurring. Crucially, the authors framed this as evidence of structure, a “phonetic alphabet,” not a decoded language: the AI revealed how the sounds are organized, while the question of what they mean remains open for biologists to test. Source: Sharma et al., Nature Communications, 2024. Read the paper.

“AI widens the funnel of what is worth investigating; people still make the call.”

DolphinGemma: a model that runs in the water

In 2025, Google introduced DolphinGemma, a roughly 400-million-parameter model built on the Gemma family and developed with Georgia Tech and the Wild Dolphin Project. It uses Google’s SoundStream tokenizer to turn dolphin vocalizations into sequences, then learns to predict the next sound, much as a language model predicts the next word, helping researchers spot recurring patterns and structure. Compact enough to run on a Pixel phone, it is designed to assist analysis in the field rather than from a distant data center. The model is trained on the Wild Dolphin Project’s decades of labeled audio of Atlantic spotted dolphins; its job is to accelerate pattern-finding, leaving interpretation to the researchers who know the animals. Source: Google / DeepMind, 2025. Project page.

Earth Species Project: open foundation models for bioacoustics

The Earth Species Project, a nonprofit whose stated premise is that more than 8 million species share the planet while humans understand the language of one, released NatureLM-audio, described as the first audio-language foundation model built for bioacoustics. Prompted in plain English alongside an audio clip, it can classify species and exhibit emergent abilities such as counting individuals in a recording. On the BEANS-Zero benchmark, NatureLM-audio reported state-of-the-art zero-shot performance on 7 of 9 datasets, and the model weights, code and benchmark were released openly. As a general-purpose tool, it is explicitly meant to give biologists a faster way to triage and query recordings, not to assign meaning on its own. Source: Earth Species Project, ICLR 2025. Announcement.

What’s next, and what to watch for

The near-term trajectory is more data and tighter feedback loops: field-deployable models like DolphinGemma can flag patterns during an expedition, letting researchers test hypotheses against behavior in something closer to real time. The honest caveats matter too. None of these systems has demonstrated two-way conversation or a verified “translation,” and confirming that a sound pattern carries a specific meaning requires controlled behavioral evidence, not model output alone. The risk to guard against is over-reading a cluster as a “word.” The promise is real, but it is a discovery accelerator, and the burden of proof still sits with the scientists.

The through-line: decision support, not replacement

Every project here follows the same division of labor. The model handles what humans cannot: clustering millions of clicks, surfacing combinatorial structure, predicting the next sound, querying archives by natural language. The biologist handles what the model cannot: pairing a flagged pattern with observed context, designing the experiment that tests for meaning, and deciding whether a finding holds. AI widens the funnel of what is worth investigating; people still make the call. That is the durable shape of this field, decision support, with the decision staying human.

Methodology & sources

• 8,719 codas, combinatorial structure and the four named features (rhythm, tempo, rubato, ornamentation) — Sharma et al., Nature Communications (2024)
• MIT CSAIL / Project CETI framing of the “phonetic alphabet” — MIT News (2024)
• DolphinGemma ~400M parameters, SoundStream tokenizer, runs on Pixel, built with the Wild Dolphin Project — Google DeepMind (2025)
• NatureLM-audio, zero-shot state of the art on 7 of 9 BEANS-Zero datasets, open release — Earth Species Project (2024/2025)
• Wild Dolphin Project founded 1985, longest-running underwater dolphin study, ~40 years of data — Wild Dolphin Project