With clipped acoustic inputs, pitch and voice feature gains changed sign with ridge penalty in human iEEG and fMRI

We tested whether 20 pitch and voice quality terms improve prediction of human auditory responses. Both models used clipped inputs and sound-class controls. The base kept 68 other descriptors. Separate public studies supplied iEEG 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. Electrical and MRI sound panels differed. The MRI studies shared clips; overlap between their people was unknown. At development-selected penalties, mean gains were +0.289, -0.002 and -0.195 points of constant-prediction error. Their 95% person Student intervals were [+0.050,+0.529], [-0.162,+0.157] and [-0.861,+0.471]. The primary shared test failed, p=0.5126. The fixed penalty grid gave negative mean gains at small penalties and positive gains at penalty 10 in all panels. The penalty 10 gains were +0.289, +0.535 and +0.507 points. The MRI intervals were [+0.486,+0.583] and [+0.316,+0.697]. A later four-penalty shared test gave Holm p=0.0411. That claim was chosen after controls and is exploratory, not independent confirmation. Fixed-fit sound resampling left the electrical benefit uncertain. In B2021, a larger block gain at penalty 10 coexisted with worse full prediction. A large unclipped gain also left negative full model skill. An exact duplicate-feature example changed gain sign without new input information. These results limit interpretation of feature gains as neural information. A label audit repaired one descriptive anatomy control without changing primary scores. Source cohorts, sampling, sound panels, source selection, feature builds and measurement filters limit scope. Code, licensed inputs, source pins, every negative result, figures and repeat instructions are supplied. Original electrical means were rebuilt exactly; a separate solver checked all 44 primary people. No common neural mechanism or causal effect was established.

auditory cortex · iEEG · fMRI · pitch · voice quality · ridge regression · feature ablation · negative results · cross-modal generalization