At an 80% target, limits set with other sounds gave lower mean coverage for voices in 20 iEEG patients and 24 fMRI adults

We tested whether error limits learned from sounds without a human voice cover omitted human voices. Separate public studies supplied signed intracranial local field potentials and BOLD fMRI. Four patients and four MRI adults supplied development. A local code lock preceded tests in 20 other patients, 16 B2021 adults, and eight NH2015 adults. Each person had twelve fixed fitting, calibration, and test splits. Models used acoustic features and removed the mean across scaled sites. Error limits had a nominal 80% target. Mean other-sound minus voice coverage was 11.35, 60.92, and 60.81 percentage points. Their 95% person Student intervals were [6.20, 16.51], [57.63, 64.21], and [55.12, 66.49]. Positive counts were 16/20, 16/16, and 8/8. The shared mean and sign p was 0.0059089661. The fixed shared test passed. Crossed person and sound resample ranges were also positive in each cohort. They hold forecasts fixed and reset calibration limits. All negative people and fixed checks remain. The three-site electrical check had only 13/20 positive people and failed its majority component. A constant-forecast check and an exact error-scale example limit neural interpretation. The result concerns a fixed measurement and prediction rule. It does not identify a voice module, nonacoustic coding, or a common cause. Cohorts, tasks, selected sites, acoustic features, trial averages, and measurement filters differ. The package gives code, 96624 primary scalar errors, source pins and terms, all outcomes, figures, audits, and repeat steps. Original MRI arrays remain external.

human neuroscience · intracranial EEG · fMRI · voice · neural encoding · prediction sets · calibration · distribution shift · public data