Quiet-gap features did not improve fixed speech response models in six iEEG patients and 25 fMRI listeners

We tested whether quiet-gap ends and their preceding duration add neural prediction during natural speech. Public data came from separate human iEEG and fMRI studies. Three iEEG patients supplied development. All five duration settings had negative mean test gains. We fixed the least negative setting before confirmation in six other patients and 25 fMRI listeners. Sound controls included spectra, flexible level terms, sound edges, word timing, and linear history. A training-only projection kept the baseline ridge prediction fixed within each feature test. Duration mean test R2 gain was -0.000276 in iEEG, with a pointwise 95% person interval from -0.000402 to -0.000180. The fMRI mean was -0.005199, with interval from -0.007180 to -0.003336. Generic gap-end events also failed the shared positive rule. Both planned shared Holm p values were 1. A duration-label diagnostic gave a positive person-mean advantage in fMRI, but its common-story label p was 0.10. The iEEG advantage was uncertain. No new shared biological effect met the planned rule. Model losses do not establish absent neural gap information. Clinical sampling, broad fMRI pooling, fixed stories, a nearly saturated duration transform, and offline preprocessing limit the result. The package preserves compact inputs, all failures, new code, plans, source terms and hashes, coefficients, figures, and repeat steps. All planned result tables reproduced exactly in a clean copy, and 75 native fMRI inputs matched pinned hashes.

human auditory cortex · intracranial EEG · fMRI · natural speech · quiet gaps · prediction · negative result · reproducibility