A sparse neural code for acoustic targets enables speech-like vocal flexibility in parrots

Abstract

English speech flexibly recombines 44 phonemes into >20,000 words and limitless phrases. Here we identify a novel neural code supporting combinatorial vocal control in budgerigar parrots. Deep learning based acoustic analyses show that both English imitated and natural warble songs are composed of subsyllabic acoustic elements that are flexibly reused across acoustic contexts, analogous to phoneme reuse in words. Such reuse poses a scaling problem for songbird-inspired timing-based codes, which require distinct neural sequences for distinct syllables even when syllables share acoustic components. We identified neurons in the parrot frontal cortical nucleus MO that exhibit sparse bursts locked to reused acoustic elements independent of syllable identity, timing, or surrounding acoustic context. These results provide the first evidence for a sparse, acoustically-grounded vocal motor code that can be flexibly reused to assemble a large and combinatorial vocal repertoire.