Variable syllable context depth in Bengalese finch songs: A Bayesian sequence model

Birdsong is an important model for vocal learning and sequential motor behavior. Similarly to human language, songs, notably those of Bengalese finches and canaries, exhibit higher-order sequence structure, meaning that the statistics of one syllable may depend on a number of previous syllables. However, this number (the context depth) varies in a manner that has challenged previous formal approaches. Here we used a hierarchical non-parametric Bayesian sequence model (based on Teh, 2006; Elteto et al., 2022) that seamlessly combines predictive information from shorter and longer contexts of previous syllables, weighing them proportionally to their predictive power. We fit our model to songs of 8 different Bengalese finches, each with > 300 song bouts (Veit et al., 2021). The model inferred the context depth, showing that it varied substantially, with some syllables depending just on one deterministic predecessor, but others depending on >10 previous syllables. Underlying this variability was syllables forming alternating and repeating chunks, i.e. strings of fixed subsequences. When fitted at the chunk-level, our model revealed different chunk-motifs that characterize how bouts typically start, unfold, and end. The model was also able to predict the flexibility with which birds can learn to switch between syllable transitions based on external cues.