Fixed speech/song readers kept minimum AUC above 0.5 after one positive and six negative score replacements in six SEEG patients and 16 fMRI adults

We tested whether a few easy sound clips support neural category reading. Separate public studies supplied human SEEG high-gamma responses and BOLD fMRI estimates. Each person had a simple reader. Fitting used 14 speech/song and 64 source nonvoice clips from repeat estimate 1. Testing used different clips and the mean of estimates 2 and 3. We computed the exact minimum AUC under test score replacement. Any one positive and six negative scores could take arbitrary values. The reader stayed fixed. Development used two patients and four MRI adults. A local plan and code lock preceded reserved scores. Mean minimum AUC was 0.6765 in six reserved patients and 0.8202 in 16 reserved adults. Their 95% person Student limits were [0.5484, 0.8046] and [0.8123, 0.8280]. All 22 values exceeded 0.5. The fixed shared mean, majority and label-reference rule passed, p=0.015625. Eight records from another MRI study gave mean 0.8210; possible person overlap between MRI studies was not auditable. Separate SEEG test estimates had mean intervals that included 0.5. Controls, lower scores, source pins, code, exact-bound checks and replay steps remain in the package. The result is a finite test-library guarantee. It does not cover training errors or unknown sounds. Acoustics, source site selection, sparse clinical coverage, task and cohort differences limit interpretation. No common neural cause was shown.

auditory cortex · SEEG · fMRI · AUC · robustness · stimulus generalization · speech · song