A model can learn what sound tends to come next without establishing what that sound means. That distinction is the starting point for reading about DolphinGemma.
What the researchers describe
Google’s April 2025 announcement describes DolphinGemma as an audio-in, audio-out model trained using the Wild Dolphin Project’s Atlantic spotted dolphin recordings. It learns structure in sequences and predicts subsequent sounds.
That is a useful research direction. It gives researchers a way to investigate regularities in recordings. It does not supply a verified dictionary of dolphin meanings.
Three questions to keep separate
Can we describe the sound? An acoustic description concerns features such as frequency and timing. A visual pattern can be striking without telling us what caused it or who produced it.
Can we predict another sound? Prediction concerns regularities in a sequence. A successful prediction should be judged against a stated task and evaluation, not an unspecified claim of “accuracy.” Ask what was predicted, what the comparison was, and what examples the model had already encountered.
Can we establish meaning? This asks for evidence about the relationship between a signal, its use, and the response it elicits. It is a different claim from finding a pattern in the recording. Treating the questions as interchangeable makes both the achievement and the uncertainty harder to understand.
How to read the next headline
Start with the source. Look for a clearly defined task, an explanation of the data, and a description of how a result was evaluated. If a story supplies a percentage without defining the outcome being measured, the number cannot do the explanatory work being asked of it.
Then separate the reported result from the proposed application. “This could help researchers investigate communication” and “this translates communication” require very different evidence. Ambitious possibilities can be discussed without presenting them as completed experiments.
My editorial position is that this distinction makes the research more interesting. We can value better tools for asking questions without pretending those questions have already been answered.
Continue with the research archive or the dolphin sound guide.