Birdsong is a rare example of learned vocal communication. Like speech, song consists of acoustic units (syllables) that are learned from adults and organized into temporal sequences (syntax) during development. Tutors' and pupils' syllables and syntax are often the same, suggesting that both song features are learned. Syntax is highly species-specific and consistent across individuals and populations of the same species, unlike syllables. Here, we hypothesized that song syntax can be accurately predicted by computational models based on the relationship between syllable acoustics and order that is consistently observed in the songs of conspecific individuals. We tested this hypothesis using techniques inspired by natural language processing to generate and test a species rules model for each of two species of estrildid finches whose songs are composed of stereotyped syllable sequences that differ in syntax. We modeled species rules as the significant correlations between a syllable's acoustic features and sequence position in a song, measured by analyzing song acoustics in many individuals of the same species. We used only a bird's syllable types and its species rules to predict syllable sequence in the birds' adult songs, without using tutors' songs to train or test the model. We then quantified the accuracy of the species rules model in predicting a bird's syllable sequence compared to the accuracy of the tutor's song in predicting a bird's syllable sequence. Results showed that species rules models predicted birds' actual sequences as well as did tutors' sequences, in both species and across different colonies. Results support the hypothesis that species-specific rules based on syllable acoustics and order can explain the species-specific syntax of birdsong. The modeling approach developed here has general utility for detecting, predicting and comparing sequential structure in complex audio signals.
Edwards et al. (Tue,) studied this question.