Language-like statistical structure arises in learned signaling: evidence from birdsong.

Language has two statistical properties that are strikingly rare in other species. First, it is made up of statistically coherent units (e.g., words) consisting of sequences of sounds, whose internal structure is relatively predictable. Second, the frequency distribution of these units follows a characteristic shape known as a Zipfian distribution where the most frequent unit appears approximately twice as often as the second most frequent one, three times as often as the third, and so on (, ). While Zipfian distributions are found across the natural world, with different explanations applicable in different domains (), recent work has shown that this distribution serves a learning function in human language: Having statistically coherent subsequences () and having their frequency follow a Zipfian distribution facilitate language learning in humans (–). For example, infants, children, and adults are better able to discover words in continuous speech when their frequency distribution follows this particular power law distribution (, , ). The fact that both statistical properties aid learning makes them good candidates for arising through cultural transmission: An extensive literature shows properties that aid learning may be present in language precisely because they help its faithful transmission over generations of learners (–). There is experimental evidence that cultural transmission can lead to the emergence of statistically coherent subsequences that follow a Zipfian distribution in humans, even in the absence of meaning or communication (). Both statistical properties were thought to be uniquely human; however, the same statistical structure has been recently found in humpback whale () song, a culturally transmitted complex signaling system where individual sound elements are concatenated to form longer sequences (). Using tools inspired by how human babies segment speech, the study found that whale song has language-like statistical structure (): The song contains statistically coherent subsequences whose distribution is Zipfian. If the similarity between these two very different species is related to culture, then we can make the strong prediction that similar structure will be found wherever complex sequential signaling is culturally transmitted. Passerine birds make an ideal test case for this prediction: Songs are made up of individual syllables used to form longer sequences. The song is complex and learned (, ), and we have data on song development (currently lacking for whales). Previous work on birdsong has examined the distribution of the individual syllables, finding that it is highly skewed (–), with some studies reporting a good fit to a Zipfian distribution (–). However, these findings only pertain to the individual sound elements, whose learning and transmission may be affected by different learnability pressures than those affecting larger sequences. Existing findings cannot tell us whether birdsong, such as human language and whale song, contains statistically coherent subsequences whose distribution is Zipfian, as would be expected if these statistical properties arise to facilitate faithful transmission. We test the prediction that similar structure will be found in culturally transmitted birdsong when applying the same infant-inspired tools used to analyze human data and humpback whale song. The method uses dips in transitional probability between the basic acoustic building blocks to segment the song into subsequences. This is the same cue that is used by infants to segment speech (): Because words are statistically coherent, transitional probabilities between syllables are higher (on average) within words than between words (). Using developmental data, we can additionally ask whether this structure is also present in juvenile song or only in the adult target. We focus here on Bengalese finches ( var. ) whose song has been extensively studied (). Bengalese finches are a domesticated variant of a wild songbird, the white-rumped munia () (). Bengalese song is learned, culturally transmitted, and has complex structure (). Only males sing (), and songs differ between birds, with each bird’s song typically consisting of around eight distinct syllables that are combined to form longer sequences (). Bengalese finches are sensitive to transitional probabilities (), the cue used by our segmentation method. In a seminatural environment, Bengalese chicks can learn from multiple tutors (). They splice sections of song from different adult males when the transition between syllables in the tutor’s song is less predictable (). This indicates that the song has statistical structure and that this information is relied on by chicks during learning: Bengalese finches do not learn songs as a stereotyped, fixed sequence but learn them as smaller chunks based on the transitional probabilities between syllables.