Munim Technologies

Research/Earth

Sperm whale coda vowels hold steady through whole bouts

A pre-registered test on the public labels behind the sperm whale 'vowel' papers. Neighbouring codas switch vowel about a fifth as often as chance predicts, and the same whale's vowel mix changes between tag deployments. The vowel behaves like a state of the moment, not a choice made coda by coda.

Nov 21, 2026 · 6 min read

In 2025 and 2026 two papers reported that sperm whale codas come in two vowel-like qualities, a and i, and that these pattern like human vowels. We asked a question neither paper modelled: in time, does the vowel behave like a choice made for each coda, or like a state that holds for a while? In a pre-registered test on the authors' own public hand labels, neighbouring codas of one whale switched vowel 75 times where shuffling predicts 352. That is about a fifth of the expected switches (Q = 0.2129, p = 0.0001). The switches that do happen cluster at long silences. The same whale, making the same coda type, also has a different vowel mix on different tag deployments. This does not show that the vowels are artefacts. It does show that the labels move like a slow state, from recording geometry, depth or a vocal register held through a bout, and that tests treating codas as independent overstate the evidence.

75

Within-bout vowel switches observed

352.23

Switches expected if each coda's vowel were independent

p = 0.0001

Permutation test, 9,999 shuffles

6 of 7

Registered predictions that came true

The question

Begus, Sprouse, Leban, Silva and Gero (2025, Open Mind) described two spectral qualities of codas, one with a single peak below 10 kHz and one with two. Begus and colleagues (2026, Proc. R. Soc. B) then reported vowel-like phonology: durations that differ by vowel, an uneven spread across coda types and coarticulation-like effects at coda edges. Critics have suggested that the peaks could reflect how the sound reaches the tag. If each coda's vowel is chosen on its own, two codas from one whale a few seconds apart should differ about as often as that whale's vowel mix predicts. If the vowel is a state, of the whale or of the recording, neighbours will nearly always agree and changes will gather at long silences.

What we did

We joined the 1,375 hand-labelled codas the authors published on OSF to recording times in the public Dominica Sperm Whale Project tables (Sharma et al. 2024). That gave 1,134 labelled, timed codas from 13 whales, 19 tag deployments and 94 bouts, with a bout ending at any silence over 60 seconds. We counted switches between neighbouring codas within bouts. We compared them with the exact number expected if vowels were shuffled within each whale, deployment and coda type, so that differences in vowel mix between whales, tags or coda types cannot explain the result. The design, the decision rules, seven predictions with priors and an independent grader were all published before the job ran (VOWEL-PREREG.md). We calibrated the test on the real timeline. In worlds where every coda's vowel is independent it rejected 11 of 200 times (size 0.055). In worlds with one vowel per bout and a share of codas flipped at random it detected 200 of 200 at a 20% flip rate.

What we found

Vowels persist strongly. Only 75 of 1,040 within-bout neighbours differ in vowel. The shuffled worlds never came close: their 9,999 counts ranged from 306 to 397. By the rule fixed in advance (p at most 0.05 and at most half the expected switches) the verdict is VOWELS_CLUSTER_STRONGLY. Persistence is strongest at the shortest gaps and fades as the silence grows.

Vowel switching by silence before the coda

Q = observed switches / expected switches; 1 means switching as often as the vowel mix predicts. Pairs per bin: 926, 114, 19, 37. Exploratory side reading. Source: RESULTS-VOWEL.md

It holds for every whale where it can be measured. Each of the seven whales with more than five expected switches switched less than expected. One of them, FRUIT, was close to chance. Five whales never used both vowels within one deployment and coda type.

Vowel switching within bouts, by whale

Whales with more than 5 expected switches. Within-bout pairs: 335, 25, 107, 271, 126, 48, 28; all 1,040. Per-whale values are exploratory. Source: RESULTS-VOWEL.md

The same whale sounds different on different tags. For the four whales with at least two usable deployments, the vowel mix of a given coda type differs between deployments far beyond chance (G = 65.899 against a null mean of 8.934, p = 0.0001). Only 20 of the 96 whale, deployment and coda-type groups contain both vowels at all.

Calibration on the real timeline

worldrejectedrate
independent vowel per coda11 of 2000.055
one vowel per bout, 5% of codas flipped200 of 2001.000
one vowel per bout, 10% flipped200 of 2001.000
one vowel per bout, 20% flipped200 of 2001.000
one vowel per bout, 30% flipped185 of 2000.925
Each world uses the real codas, bouts and strata with 999 permutations. The observed Q of 0.2129 is closest to the 5% world (mean Q 0.2663, exploratory). From RESULTS-VOWEL.md.

What it means

The data look like one vowel per bout, with roughly one coda in twenty showing the other. That fits a slow state better than a fresh choice for each coda. Our test cannot say which state. A whale could hold a vowel through a bout the way a speaker holds a register, and the coarticulation-like effects reported in 2026 fit that reading. Tag position, orientation and depth would produce the same pattern, and the change between deployments of one whale points that way. Either way, analyses that count codas as independent draws of a vowel treat roughly a thousand correlated labels as a thousand pieces of evidence. Claims about how vowels pair with coda types, durations or whales need tests that respect bouts and deployments. The sharpest next check would set the deployment effect against tag placement and depth records, which are not in the public files.

More research