Input clipping left an uncertain mean gain for smooth acoustic models in 20 iEEG patients and a positive gain in 20 fMRI adults

We tested smooth single-cue and cue-pair acoustic models in two separate human studies. Four intracranial EEG patients supplied development. A fixed test then used 20 other patients and 20 fMRI adults. Each model included the known voice label. Nested sound folds separated test responses from fitting and penalty choice. The studies had different sound panels and feature sets. All 20 people in each modality favored smooth cues over the original linear model. Mean gains were 8.629 and 2.965 percentage points against voice-only test error. Their 95% person t intervals were 6.007 to 11.251 and 2.529 to 3.401. The fixed shared rule passed; two-claim Holm p was 0.00000190735. The added cue-pair rule failed. A later exploratory control clipped linear-model inputs at three training standard deviations. The remaining smooth-model gains were 0.246 points in iEEG and 1.974 in fMRI. Their intervals were -0.552 to 1.044 and 1.569 to 2.378. Clipping removed 97.1% of the observed mean electrical gain. Two sounds supplied 61.3% of that original gain. Thus, the positive fixed model ranking did not establish a shared nonlinear neural mechanism. Cohorts, signals, features, source selection and spatial support differ. The package gives code, permitted electrical inputs, source versions and terms, all test errors and negative results, figures, audits and repeat instructions. Original BOLD arrays remain external.

human neuroscience · intracranial EEG · fMRI · auditory encoding · acoustic features · additive kernels · input clipping · model comparison · negative result · reproducibility