Munim Technologies

Research/Earth

Whale song's Zipf law is matched by a simple Markov chain

In a pre-registered re-test on the authors' own data, a second-order Markov chain with no words in it made humpback song more Zipfian than the real song (R^2 0.939 against 0.930). A power check calibrated on real English says this kind of test cannot yet tell whether whale song has words.

Nov 18, 2026 · 6 min read

In 2025 a paper in Science reported that humpback whale song, cut into chunks the way infants are thought to find words, follows Zipf's law as closely as human language (mean R^2 0.93). We re-ran that analysis on the authors' own data and code, reproduced their number to the last digit, and then asked a question the paper's controls do not: is the pattern more than simple rules about which sound follows which? It is not, at least not detectably. A second-order Markov chain, which only remembers the previous two sound elements and contains no word-like units, produced song that was slightly more Zipfian than the real song. A power check we calibrated on real English shows why this is not the end of the story: at the amount of song available per year, the same test also misses real English words.

0.930

Zipf R^2 of real whale song (reproduced exactly)

0.939

Mean Zipf R^2 of song from a second-order chain

p = 0.970

Real song against that chain (one-sided)

1 of 8

Years where the power check says words are detectable

The question

The whale paper (Arnon et al. 2025, Science 387:649) and a 2026 finch paper (Kirby et al., Science Advances) both cut song wherever the probability of the next element dips, then count the chunks. Chunk frequencies fall off with rank the way word frequencies do, and the papers show that this beats two controls: shuffling the elements, and rotating them. Both controls destroy the song's transition structure. Neither asks whether that transition structure alone, with no units at all, is enough to produce the law. That is the test we registered in PREREG-AMENDMENT-3.md and PREREG-AMENDMENT-5.md before opening the data.

What we did

We read the authors' released data for eight years of song (2010-2017, 55 recordings, 45,749 sound elements) and passed an exact reproduction gate: their Zipf R^2 of 0.9302888665216734 came out as 0.9302888665216728, with the same 3,386 chunk types, and their most frequent chunk in 8 of 8 years. We then built 1,000 surrogate song sets per null model: the paper's shuffle and rotation, within-bout resampling of first and second order, and pooled Markov chains of first and second order fitted to each year's song. Every null was cut and scored exactly like the real song.

What we found

Zipf R^2: real whale song against song made by each null model (theta 0.5)

Null values are means over 1,000 surrogate data sets. The real song beats the paper's two controls (z 23.8 and 13.9) and both first-order nulls, but not either second-order null. Source: RESULTS-WHALE.md

By the rule we registered, this is explained by second order. Because an earlier calibration showed the test cannot always separate word-like from wordless song at whale size, it is reported as not distinguishable. The label is not robust to the cutting threshold: at the two loosest thresholds the real song does beat the pooled second-order chain, though never the second-order resampling.

Robustness across cutting thresholds

thresholdreal R^2second-order chain meanplabel
0.250.8320.7950.010beyond second order
0.30.8580.8390.030beyond second order
0.5 (registered)0.9300.9390.970explained by second order
0.70.9510.9590.990explained by second order
0.750.9530.9631.000explained by second order
200 null data sets per row except the registered threshold (1,000). Only the registered threshold is graded.

Can the test see words at all?

A negative result only means something if the test could have come out positive. So we ran the identical pipeline on real child-directed English, cut into phonemes, at five data sizes (17 replicates). Below about 2,700 phonemes per recording, English is also "explained by second order". Above it, English words are detected every time. A cheap held-out check, which we call D1, predicted this in advance: replicates where a second-order model generalises better than a first-order one in most recordings were detected 7 of 7 times, the others 1 of 10.

Real English: share of replicates where words beat the second-order chain

words detected (theta 0.5)D1 majority (power check passes)
00.250.50.751200040006000800010000phonemes per recordingshare of replicates
Replicates per size: 5, 5, 4, 2, 1. All 7 registered predictions came true (Brier score 0.120). Source: RESULTS-CDS-SIZE.md

In the whale data, D1 favours order 2 in only 1 of 8 years (2015, the largest, with 14,429 elements). By element count alone the whale years look large enough, but by this power check they sit where English words are missed too. A likely reason, not yet tested, is that whale song is very repetitive within a year, so it carries less independent information than its element count suggests.

What it means

The controls the whale and finch papers rely on are passed alike by English, by finch song, by whale song and by synthetic song with no words. A Markov chain fitted to the whale's own transitions reproduces its Zipf law in full at the registered threshold. The fair reading is narrow: the published Zipf evidence does not separate word-like units from second-order element rules, and at this data size no test of this kind can yet. It is not evidence that whale song lacks words. Two things would decide it: more song per year, or a test that predicts held-out chunks instead of refitting the null on the same data. We suggest reporting D1 next to any animal-song Zipf result; it needs no null models and takes seconds. Our next test is an animal positive control: canary song, whose phrase syntax is known to reach three or more phrases back.

More research