Context-dependent syllable transitions in birdsong

Like phonology in human speech, syllable sequences in birdsong follow probabilistic rules. Here we analyze the Bengalese finch songs, which consist of variable sequences of 7-15 stereotypical syllables. We show that the sequences can be statistically modeled with partially observable Markov models (POMMs), which are Markov models in the state space, with each state associated with a single syllable. The syllable sequences are in general non-Markovian, therefore the POMMs must have multiple hidden states associated with the same syllables. This state multiplicity encodes context-dependent syllable transitions. Since the numbers of states associated with individual syllables in POMMs are not known a priori, they must be inferred from the observed syllable sequences. We show that this inference process can be casted into a form of hypothesis test of whether a model overgeneralizes. Using this method, we analyze how auditory feedback impact the context-dependencies in the Bengalese finch songs. Deafening significantly reduces but does not eliminate context-dependencies, suggesting that auditory feedback is an important but not the only brain mechanism for context-dependent syllable transitions.