AI is revealing how sperm whales build their calls

A sperm whale’s social calls contain sequences of clicks known as codas. Researchers once grouped them mainly by the number of clicks and the gaps between them. Work by the MIT and Project CETI collaboration has shown how much more information about their structure becomes visible when those features are examined across exchanges between whales.[2]

Timing has more than one dimension

The 2024 Nature Communications study distinguishes rhythm, the relative spacing of clicks, from tempo, the overall duration of a coda. It also identifies rubato, gradual timing adjustments across successive codas, and ornamentation, an extra click added to a familiar pattern. The researchers found that these features combine, and that some variation depends on neighbouring calls in an exchange. Here, context refers to the surrounding codas; it does not establish the subject of a conversation.[1]

A simple musical analogy helps separate the ideas. Tap a short pattern slowly, then repeat it faster: the proportions can stay similar while its total duration changes. Add one final tap and another property changes. The analogy illustrates independent dimensions of a sound pattern; it is not a claim about what a whale intends to communicate.

MIT describes the use of machine learning to expose this richer organisation across thousands of recordings. The useful shift is from assigning each call one coarse label to examining the features that make calls similar or different. That can guide analysis of how a whale responds to a preceding call, or which combinations occur together.[2]

The sound inside the clicks adds another dimension

In 2025, UC Berkeley and CETI reported two recurring spectral patterns that researchers described as a-like and i-like vowels, together with patterns resembling transitions between vowels. Spectral analysis asks how sound energy is distributed across frequencies. Two calls with similar click timing can therefore differ in acoustic quality.[3]

Berkeley’s account describes generative adversarial networks as helping point the linguists towards productive features to examine. The researchers then analysed the details. That is a useful role for machine learning in this setting: proposing an organisation within recordings so that a specific acoustic hypothesis can be tested.[3]

The 2026 Royal Society paper reports several additional patterns, including duration differences between vowel qualities and neighbouring codas influencing the quality of an edge click. Its detailed duration comparison focused on one coda type from four well-recorded whales. The study explicitly compares acoustic organisation with human phonology while leaving the meanings of particular codas unresolved.[4]

A recording becomes more useful when it has a setting

The next scientific opportunity is to connect sounds with independently observed behaviour. CETI’s field programme records audio alongside information such as pressure, movement and orientation, with drone footage adding another view. Synchronising those streams lets researchers relate a vocal exchange to the animals’ activity and social setting.[5]

Consider a hypothetical pattern that appears frequently before a change in group movement. Researchers would need to ask whether the association persists across different whales and encounters, and whether another event explains both the call and the movement. This is a proposed reasoning example, not a result reported in these papers. Finding repeatable associations would turn an acoustic category into a more specific behavioural hypothesis.

For readers who want to see the ideas, CETI’s research page provides short visual explainers, including Vowels 101 and Whale Sound 101. Its combination of recordings, acoustic analysis and observation shows why progress depends on the field data as much as the model.[5]

The immediate achievement is a richer description of a sound system: dimensions that researchers can measure, compare and investigate across exchanges. AI helps make those patterns visible. Each well-defined pattern then gives marine biologists and linguists something concrete to test against life in the water.

Report an error or suggest a correction