Sparse coding of harmonic vocalization in monkey auditory cortex

One of the potential foundations for the perception of harmony is the neurons in the primary auditory cortex (A1) that specifically respond to harmonically related frequencies, whose underlying organizing principles are yet unclear. We hypothesize that such harmony-related responses result from adapting to natural harmonic sounds by using sparse coding, a computational model that originally related natural image statistics to the neural properties of the primary visual cortex (V1). Our previous work has shown that a set of harmonic responses found in A1 of marmoset monkeys can emerge by sparse coding of highly harmonic sounds; however, the work was only preliminary because the specific sounds we used were unfamiliar to monkeys. In the present study, we recorded voices of marmosets, and showed that the same model applied to the conspecific vocalizations can reproduce the harmony-related responses. This result more directly supports the hypothesis that A1 adapts to natural sounds, in particular voices comprising harmonic overtones, under the principle of sparse coding. ► We model A1 neurons using sparse coding of behaviourally imporant sounds. ► We record marmoset voices and use them as inputs to sparse coding model. ► Sparse coding of monkey voices explains a rich repertoire of A1 tuning curves. ► Harmonically related distant modulations of A1 tuning curves are reproduced. ► A1 and V1 may share the sparse coding principle, but input statistics are different.