Band change magnitudes did not confirm a shared speech prediction gain in six iEEG patients and 25 fMRI listeners

We tested whether separate weights for rising and falling sound levels improve later neural prediction. Public inputs came from separate human intracranial EEG (iEEG) and fMRI speech studies. The baseline had eight band levels, signed band changes, total sound controls and word timing. The larger model added absolute 100 ms band changes before each measurement filter. Each model selected its own ridge penalty with earlier validation data. Three iEEG patients supplied development. A fixed test used six other patients and 25 fMRI listeners. Mean added test R2 was +0.000526 in iEEG, with 95% person t interval -0.000201 to +0.001252, and -0.008798 in fMRI, with interval -0.014252 to -0.003343. Gains were positive in 5/6 patients and 7/25 listeners. The shared positive test failed. All fixed controls and negative scores remain. This gives a reproducible benchmark with nested penalty selection. It does not establish absent edge encoding, cross-modal neural generalization, or causality. Small clinical samples, different stories and spatial coverage, source processing and full-run fMRI nuisance removal limit the test. The package contains inputs, source terms and pins, code, fitted models, all scores, figures and repeat instructions.

human iEEG · fMRI · speech · sound changes · negative result · nested validation · prediction