2026-08-12
Applying an infant-speech segmentation method to Bengalese finch song, the detected subsequences follow a Zipfian power law with slope 1.05, matching human language.
Human language has two statistical properties that are rare in nature. It is built from statistically coherent subsequences, words, whose internal sound transitions are more predictable than the transitions between them. And the frequency of those words follows a Zipfian distribution, a power law where the most frequent unit appears about twice as often as the second, three times as often as the third, and so on.
These properties are not accidents. A large body of work shows that segmenting speech into coherent chunks and having their frequencies fall on a Zipfian distribution both make language easier for infants and adults to learn. That supports an appealing hypothesis: the two properties exist because they make language learnable, and cultural transmission between generations favors whatever transmits faithfully. Last year the same authors found both properties in humpback whale song, issuing a strong prediction that any complex sequential signaling system that is culturally transmitted should look this way. This paper takes the test to birds.
Prior birdsong work had only measured the distribution of individual syllables (it is skewed, sometimes Zipfian). But the learning pressures on single syllables differ from those on larger sequences, so no one had shown whether birdsong, like speech and whale song, contains statistically coherent subsequences whose frequency is Zipfian.
The core move is to apply a segmentation pipeline built for human data and whale song to Bengalese finch song. Bengalese finches are domesticated munias with complex, culturally transmitted song. The segmentation cue is the transitional probability between adjacent syllables: within a coherent unit, syllable-to-syllable transitions are tighter, so within-unit transitional probabilities run higher than between-unit ones. Infants use this exact dip to find word boundaries (the classic Saffran 1996 finding).
For each recording the authors estimate the transitional probability between every pair of consecutive syllables, then place a cut whenever one transition drops to less than half the previous one (a ratio below 0.5, the same threshold used for whale song; 0.3 and 0.7 are also tested). Using ratios rather than raw probabilities catches relative dips: 0.4 is not low on its own, but a fall from 0.9 to 0.4 is meaningful.
The corpus comes from Takahasi and Okanoya: six males, five recordings each, from the earliest analyzable song around day 60 to crystallized song around day 120, with roughly ten song bouts per recording. Syllables are classified by eye, each assigned a letter, with 93.3 percent inter-coder agreement. This developmental data, which whale song lacks, lets the authors ask whether the structure is adult-only or runs through development.
To rule out the Zipfian fit being an artifact of the syllable distribution, three baselines are each run 1000 times: shuffling syllables (preserving syllable frequency, destroying sequence), preserving bigram statistics while disrupting longer sequences, and rotating the cut points (preserving order and chunk length but removing their alignment with transitional-probability dips).
Bengalese finch song does break into statistically coherent subsequences whose frequency is Zipfian.
| Threshold | Power-law fit R² | Cuts per bout |
| 0.3 | 0.75 | 5.01 |
| 0.5 | 0.89 | 10.72 |
| 0.7 | 0.90 | 16.75 |
At the 0.5 threshold the mean slope is 1.05, close to reported values for human language. All three baselines fall far behind: shuffled syllables z=17.05, bigram-preserved z=6.77, rotated cuts z=15.69, all P<0.00001. The fit is driven by genuine sequential structure above the bigram level, not by syllable frequency or the cutting scheme.
The developmental data carries two sharper findings. Statistical coherence increases with development: the gap between within-subsequence and between-subsequence transitional probabilities widens (β=0.0005, P=0.033), and the detected subsequences grow more similar to each bird's adult tutor song, measured by normalized Levenshtein similarity (β=0.0021, P=0.0044). Yet the power-law fit R² does not change across development (P=0.56); only the slope grows, making the distribution more skewed (P=0.0056). Even when juvenile song is still a noisy version of the adult target, the Zipfian shape is already there.
The detected subsequences also obey Zipf's law of brevity: more frequent subsequences are shorter (Poisson mixed-effects model, β=−0.04, P<0.00001), a signature of efficient coding.
Humans, humpback whales, and Bengalese finches sit hundreds of millions of evolutionary years apart, yet all three build signaling systems from coherent, power-law-distributed parts. The claim is that this structure emerges because it makes the signal learnable, and cultural transmission selects for learnability: convergent evolution.
For people working on language models, there are two hooks. First, the Zipfian distribution is the shape of natural-language training data itself. This paper offers an answer to why it has that shape: it is not an arbitrary property of language but a near-inevitable product of learnability pressure acting under cultural transmission, which explains why the same shape recurs across corpora and within parts of speech and semantic categories. Second, the thesis that structure emerges from transmission and learnability is the same thesis that the multi-agent emergent communication line of ML research keeps drawing on, and Simon Kirby's work is a primary source for it. This paper extends the evidence from humans to birds.
Bengalese song is a mating display, and the detected subsequences carry no semantic meaning. That shows the core statistical properties of language can arise in a system with transmission and learning pressure but without meaning.
The sample is small: six birds, each learning from a single tutor, though in semi-natural conditions Bengalese finches can learn from several. Domesticated Bengalese finches sing more variably than their wild munia ancestors, an effect of domestication.
The 0.5 cutting threshold is borrowed directly from the whale-song analysis; there is no random baseline for real birdsong to calibrate it, which the authors acknowledge, though the matching results at 0.3 and 0.7 are a robustness backstop.
More fundamentally, whether the detected subsequences are the functional units the birds themselves process is unverifiable. Bengalese finches are sensitive to transitional probabilities, but that does not mean they segment their own song this way. And the evidence is correlational: it confirms the prediction that culturally transmitted systems are Zipfian, but cannot establish that every Zipfian distribution comes from culture, a directionality the authors flag explicitly.