Abstract  Modern theories of learned vocal behaviours, such as human speech and singing in songbirds1 , posit that acoustic communication signals are reproduced from memory, using auditory feedback2 . The nature of these memories, however, is unclear. Here we propose and test a model for how complex song structure can emerge from sparse sequence information acquired during tutoring. In this conceptual model, a population of combination-sensitive (phrase-pair) detectors is shaped by early exposure to song and serves as the minimal representation of the template necessary for generating complete song. As predicted by the model, birds that were tutored with only pairs of normally adjacent song phrases were able to assemble full songs in which phrases were placed in the correct order; birds that were tutored with reverse-ordered phrase pairs sang songs with reversed phrase order. Birds that were tutored with all song phrases, but presented singly, failed to produce normal, full songs. These findings provide the first evidence for a minimal requirement of sequence information in the acoustic model that can give rise to correct song structure.

Also see the News & Views article for this study by Daniel Margoliash.