Mean residual voice AUC was 0.614 in six SEEG patients and 0.746 in 16 fMRI adults; the shared test failed

We tested voice reading after two fitted response patterns were removed. Separate public human studies supplied SEEG high-gamma activity and fMRI BOLD estimates. Each person had a separate centroid reader. Four nonspeech voice clips and 64 other clips fit it. Five voice clips and 64 other clips tested it in separate response estimates. We removed uniform activity and a speech/song preference pattern in standardized site space. The primary SEEG input lost its short trial baseline. Development used two patients and four adults. A local code lock preceded scores in six other patients and 16 other adults. Mean residual AUC was 0.61354 and 0.74570. Their 95% person Student limits were [0.46867, 0.75841] and [0.66907, 0.82233]. The fixed shared rule failed, p=0.0625. All three below-chance records remain. Eight complete records from another MRI acquisition gave mean 0.70781; possible person overlap was unknown. At 24 sites, both main-study intervals included chance. An exact model with one true signal direction gave residual AUC 0.76025 after a 30-degree direction error. Signal difference was twice noise SD. Thus favorable residual reading would not establish an extra neural signal dimension. Source screens, sparse clinical coverage, separate cohorts, short baseline, acoustic confounds and measurement filters limit scope. No positive neural generalization across both modalities or causal effect was confirmed. The package gives code, source versions and terms, all scores, uncertainty draws, audits, figures and replay steps. Original neural arrays remain external.

human neuroscience · SEEG · fMRI · voice decoding · measurement error · projection · cross-modal validation