Imagine watching a demonstration described as “whale translation”. A recording goes in, a new sequence of clicks comes out, and the result sounds convincing. Before deciding what has been achieved, ask what the demonstration claims to preserve: the acoustic character of a sound, the identity of its source, or the meaning of a message?
Those are different research questions. Recent work on sperm whales offers a particularly interesting way to explore them. It combines careful observation, analysis of sound structure and generative AI. Reading these contributions separately makes it possible to appreciate the progress without supplying an imaginary conversation.
A click sequence has a structure to investigate
Sperm whales communicate through short click sequences called codas. In a Nature Communications paper published on 7 May 2024, Pratyusha Sharma and colleagues used a dataset of 8,719 codas from the Eastern Caribbean 1 clan. The underlying recordings came from The Dominica Sperm Whale Project, collected between 2005 and 2018.
The analysis identified combinations of rhythm and tempo, gradual changes in coda duration called rubato, and occasional extra terminal clicks called ornamentation. Importantly, “context” in this paper means neighbouring vocalisations, rather than an activity such as diving. The authors explicitly describe a study of structure, without characterising call meanings or conducting playback experiments.
That distinction changes how to read the result. A structured exchange gives researchers patterns to explain. It does not, on its own, assign an English sentence to each pattern.
The vowel comparison concerns sound organisation
A Proceedings of the Royal Society B paper published on 15 April 2026, by Gašper Beguš and colleagues, examined vowel-like qualities in codas. The researchers initially retained 1,144 clearly recorded codas produced by the tagged whales; some analyses applied further exclusions.
The authors report interactions between coda types based on click count and timing and acoustic categories labelled a and i. They also identify duration patterns and similarities between clicks at the edges of adjacent codas. These are proposed parallels with the organisation of human speech sounds. The paper expressly leaves the information conveyed by the sequences aside. Its vowel labels are analytical categories, rather than translations of human words.
What a generative whale model is tested to do
Orr Paradise and colleagues’ NeurIPS 2025 paper, WhAM, describes a Whale Acoustics Model adapted from VampNet, a model pretrained on musical audio, using approximately 10,000 coda recordings. WhAM turns audio prompts into synthetic sounds with the acoustic style of sperm whale codas. The authors explicitly distinguish this acoustic conversion from translating meaning.
Evaluation included a measure of similarity between audio collections and a study involving five specialists in marine biology or underwater acoustics. Experts compared natural and synthetic recordings and identified shortcomings in generated clicks and background noise. The paper also notes that echolocation sequences had entered datasets intended for communication codas. Its results concern sound generation and acoustic classification; they do not establish that whales understand a generated message.
Keep the achievement and its interpretation visible
My practical reading rule is to write down the demonstrated task before interpreting a result. “Generated a synthetic recording”, “classified a sound category” and “tested a proposed meaning” describe different achievements. A report can be impressive on one task while leaving another entirely open.
This matters when comparing demonstrations. A recording that sounds convincing to a human specialist answers a question about human assessment. Evidence that an animal responds in a particular way would address a different question. Even then, the proposed explanation would need to be compared with alternatives. Keep track of who or what supplied the judgement.
A hypothetical pattern before a dive
Consider an invented future study in which researchers notice that a particular coda pattern often precedes a group dive. It might be tempting to give the pattern the English gloss “We are diving now”. That would be a proposed interpretation of the observation, with several questions still to answer.
Perhaps the pattern is associated with an individual who commonly vocalises at that point. Perhaps it forms part of a longer exchange, and the relevant information depends on what came before it. Perhaps a feature of how those occasions were recorded helps explain the apparent relationship. These are hypothetical alternatives, not findings about the codas in the cited papers.
A stronger investigation would ask what happens on comparable occasions when the pattern is absent, whether the same relationship appears in new recordings, and which alternative explanations survive those comparisons. The prediction should be stated clearly enough that a disappointing result can count against it. The interpretation would then have something concrete to answer to.
Ask questions that match the claim
For readers, creators and business owners encountering an AI research story, a short evidence note can keep the discussion specific:
- Input: Which recordings were used, and how were they selected and labelled?
- Task: Was the system detecting, classifying, generating or testing an interpretation?
- Evaluation: What did the success measure count, and who assessed the result?
- Boundary: Which larger question remains unanswered by that test?
Use the answers when writing a caption, briefing or article. Name the result precisely, then identify any interpretation you are adding. If you illustrate a possible meaning, label it as hypothetical beside the example. A reader should be able to tell which part comes from the study and which part is your explanation.
The MaryChuks guide to reading a research paper offers a fuller route through the question, methods, result and limitations. For this topic, the useful habit is to follow each claim back to its particular task and evidence.
Leave room for discovery
There is plenty to be curious about in a communication system whose patterns are becoming more measurable. A careful account can explain why a sound distinction is interesting, what a model can manipulate and what a proposed test would need to establish. It can also acknowledge uncertainty without treating it as failure.
For anyone sharing this research, the next useful question is specific: what observation would change the interpretation? Keeping that question visible gives future findings a clear place to contribute, instead of asking a striking demonstration to settle everything at once.
Sources
- Sharma and colleagues, Contextual and combinatorial structure in sperm whale vocalisations, Nature Communications, 7 May 2024.
- Paradise and colleagues, WhAM: Towards A Translative Model of Sperm Whale Vocalization, NeurIPS 2025. Author manuscript posted 1 December 2025.
- Beguš and colleagues, The phonology of sperm whale coda vowels, Proceedings of the Royal Society B, 15 April 2026.
Discover more from Marychuks.com AI, Psychology, Business & CreativeVerse
Subscribe to get the latest posts sent to your email.
