Fig. 5. Simulation results.

A) Simulated fMRI responses to natural sounds (NS) and model matched sounds (MM) for a representative voxel in putative primary and non-primary auditory cortical regions (PAC and non-PAC), respectively, and the associated noise corrected normalized squared error (ncNSE). The ncNSE is a measure of the similarity between two response vectors (i.e., the similarity between the response to a set of NS and the corresponding MM sounds). In this simulation, sounds are assumed to belong to two categories (e.g., voices and non voices). The simulations are built so that a voxel’s fMRI response is always driven by the acoustic content of the sounds, while the influence of the sound category differs between cortical regions (PAC vs. non-PAC) and the type of sounds (NS vs. MM). In particular the categorical influence in PAC is small and equal between NS and MM sounds. Instead, in non-PAC NS sounds are characterized by a larger contribution of the sound category. For these simulated responses, concluding that there is no influence of acoustics based on the ncNSE therefore represents a false negative. B) For different levels of SNR, the proportion of false negatives (type-II error; blue curve) associated to an α-level of 0.05 obtained after 5000 simulations, together with the data reliability (green curve) as measured in Norman-Haignere and McDermott (2018). The left and right panel report the proportion of false negatives in the putative PAC and non-PAC scenario, respectively (i.e. in case of an equal vs. different influence of categorical information between NS and MM sounds). The proportion of false negatives associated with a data reliability of 0.4 is highlighted by the red circle.