Young Bengalese finches learn an adult’s song the same way – in chunks. Researchers have now found that the song itself is built from “words.”

Simon Kirby of the University of Edinburgh, Kazuo Okanoya of Teikyo University, and Inbal Arnon of the Hebrew University of Jerusalem led the work. They borrowed a method first used on human infants and ran it on the songs of six young male finches. In an interview with Earth.com, Arnon – also a Leverhulme Trust visiting professor at Edinburgh – said the chunks behave like words in two ways.

Hear one sound in a chunk and you can guess the next, and a few chunks come up constantly while most stay rare. “This is really exciting because both properties arise in human language because they help learning,” she said. Only the males sing, and each young bird learns its song from a single adult tutor. Each male’s song is built from about eight distinct syllables, and each bird strings them in its own order.

The team counted how often every syllable in a recording followed every other – the same cue infants track in speech. When the next syllable was less than half as easy to guess as the one before, they cut the song at that spot. With each syllable represented by a letter, one adult’s song began: cc, bbc, aaaaaa. The bird then sang one longer chunk, bccddeefaaa, over and over for the rest of the song.

In earlier experiments, chicks raised with several tutors spliced together sections from different adults’ songs. They made the joins where the next syllable was hardest to guess. Across languages, the most common word tends to appear about twice as often as the runner-up and three times as often as the third. In English, that word is “the.” Linguists call the pattern Zipf’s law.

In the 30 finch recordings, the team found the same curve. Its steepness was 1.05, almost exactly the same as in human languages. The curve might have come from which syllables each bird used instead of their order. The team tested that: they shuffled the syllables inside every song and made 1,000 scrambled versions. In every single one, the scrambled songs followed the curve less closely than the real songs. The team ran two harsher checks. In one, neighboring syllable pairs stayed together; in the other, they slid all the cuts a random distance along the song. In both tests, the real songs still fit the curve best. The chunks also followed a second rule of human speech: the most frequent ones were the shortest.