A fixed acoustic voice score increased mean response error in 20 iEEG patients and decreased it in ten fMRI adults

We tested whether one acoustic voice score predicts neural responses beyond known voice and nonvoice class means. Separate public studies supplied signed intracranial EEG responses and fMRI response estimates. Four patients supplied development. Fixed tests used 20 other patients and ten fMRI adults. Acoustic fitting, neural fitting and test sounds were separate. Mean gains against intercept-only error were -2.079 and +9.983 points. Their 95% person t intervals were -2.845 to -1.313 and +9.047 to +10.918 points. All 20 patient gains were negative. All ten fMRI gains were positive. Gains over shuffled acoustic axes were positive in both studies. The fixed shared rule failed, p=1. An exact Gaussian noise control used the same electrical acoustic design. It gave a 7.446-point gain over shuffled axes with no neural sound information. The fMRI noise control gave 0.029 points. Different prediction variances can thus favor one axis without useful neural prediction. Extra sound splits, scan checks and site-count checks kept the signs of the mean actual gains. A sound-resampling interval included zero for the electrical advantage over shuffled axes. These results concern fixed models and reused cohorts. They do not establish a causal modality difference, absent graded voice coding, or a shared positive neural effect. The package gives code, permitted electrical inputs, source versions and terms, plans, all outcomes, figures, a noise reference and repeat steps.

human auditory cortex · intracranial EEG · fMRI · voice · prediction · negative result · acoustic features · prediction variance · reproducibility