Research/Earth
Humpback song's extra structure vanishes once its repeats are removed
A pre-registered held-out test on the data behind a 2025 Science paper. Humpback song has structure a second-order Markov chain cannot reproduce, but once its repeated phrase cells are collapsed the test finds nothing beyond that chain, though it still finds English words in data of the same size.
Nov 19, 2026 · 7 min read

In 2025 a paper in Science reported that humpback whale song, cut into chunks where the next sound becomes hard to predict, follows Zipf's law the way words in human language do. Over five pre-registered tests on the authors' own data we have now followed that structure to its source. Humpback song does contain sequential structure that a second-order Markov chain cannot reproduce on unseen song. But it sits in the song's repeats. When every repeated phrase cell is collapsed to one copy, a held-out prediction test finds nothing beyond a second-order chain (p = 1.000), while the same test finds English words in 5 of 5 samples of the same size. On these data, the structure behind the 'language-like' chunks is phrase repetition plus short-range rules.
p = 1.000
Collapsed whale song against 1,000 second-order surrogates
8 of 8
Years in which looking further back predicts unseen song worse
5 of 5
Size-matched English samples detected by the same test
0 of 3
Songs with random repeats added that fooled the test
The question
Arnon et al. (2025, Science 387:649) found Zipfian chunk frequencies in eight years of humpback song from New Caledonia. Earlier in this project we reproduced their number to the last digit and showed that a second-order Markov chain, which remembers only the last two song elements and has no words, produces song that is just as Zipfian (RESULTS-WHALE.md). A stronger instrument, a held-out prediction test, then found that whale song does have structure such a chain lacks (p = 0.005; RESULTS-HOLD.md). A third test found that the song repeats its cells far more than the chain does: collapsing repeats removes 63% of its elements against 50% for the chain, and 35 of its 80 most frequent 'words' are repetition cells (RESULTS-RUNS.md). So the decisive question was: once the repeats are gone, is any structure beyond a second-order chain left?
What we did
We registered the test, its decision rules, seven predictions with priors, and the code that grades it before it ran (HOLDC-PREREG.md). Every tandem repeat in each year's song, meaning a cell of 1 to 8 elements sung twice or more in a row, was collapsed to one copy. The held-out test then asked, for each year, whether models that look 3 to 5 elements back predict unseen recordings better than a model that looks 2 back, compared with 1,000 second-order chains generated under the same collapse rule. Three controls were fixed in advance. Real English (child-directed speech written as phonemes) was cut into samples as large as the collapsed whale years, to check that the test can still see real units at this size. Second-order surrogates served as negative controls. And synthetic song made of second-order chains with random repeats added asked whether repetition alone can fool the test.
What we found
Nothing beyond second order in the collapsed song. The whale's mean held-out gain was -0.1454 bits per element, against -0.0955 for the surrogates; none of the 1,000 surrogates scored lower, so p = 1.000. In every one of the 8 years, looking further back made predictions of unseen song worse. In 4 of the 8 years a first-order model predicted best of all.
Held-out gain beyond second order, collapsed data
bits per element
The test had the power to see real units. English samples with as many phonemes as the collapsed whale years were detected in 5 of 5 replicates. None of the 5 English negative controls was detected, nor the whale negative control, so the registered control gate passed.
Collapsed whale song: held-out cross-entropy by context length
Random repetition does not fool the test. We had expected (prior 0.80) that song with random repeats added would be detected before collapsing. It was detected in 0 of 3 data sets. By the reading rule we fixed in advance, this means random repetition alone does not explain the earlier HOLD result. So the extra structure is carried by the repeats, but not by repetition as such. It lies in how the whale repeats. One candidate is run length: a chain that remembers two elements cannot count to three, so a whale that sings a cell a set number of times would look like structure beyond second order. That is a hypothesis, not yet tested.
Five registered tests on the whale data
| test | question | whale result |
|---|---|---|
| WHALE | Zipf law beyond a second-order chain? | no, p = 0.970 |
| HOLD | beyond second order on unseen song? | yes, p = 0.005 |
| RUNS T2 | more repetition than a second-order chain? | yes, 0.6290 vs 0.4982 removed |
| RUNS T3 | top chunks are repetition cells? | 35 of 80 vs 26.89 expected |
| HOLDC | beyond second order once repeats are collapsed? | no, p = 1.000, with power |
What it means
The Zipf law of humpback song is computed on chunks that are largely the same short cell sung over and over. A second-order chain matches the law, and once the repeats are removed a calibrated test that still finds English words at the same size finds no longer-range ordering in the song. The structure behind the whale's 'language-like' chunks is phrase repetition, which biologists have described since Payne and McVay (1971), plus short-range rules between elements. That does not mean humpback song lacks meaning or units; it means these statistics give no evidence of word-like ordering beyond repetition. The working manuscript (paper/DRAFT.md) now covers all ten tests on whale and finch song, and recommends that any claim of language-like structure in animal song be checked again after collapsing repeats.


