A detector trained on nonvocal sounds ranked fixed vocal groups above controls in selected records from six SEEG patients and 24 fMRI adults

We tested whether human vocal sounds have unusual spatial response patterns relative to nonvocal sounds, without fitting a detector to vocal examples. Separate public studies supplied SEEG high-gamma responses and BOLD fMRI estimates. Development used two patients and four adults. A local code lock preceded tests in six other patients, 16 other B2021 adults and eight NH2015 adults. Training used 96 nonvocal sounds. Targets were nine nonspeech voices or 28 speech/song sounds, each compared with 32 held-out nonvocal sounds. The final transform removes spatial mean and vector length after site scaling. It is invariant to positive test-vector gain, conditional on the source zero and fixed training scales. Both unadjusted and repeat-alignment-adjusted scores were required. Adjusted mean AUCs were 0.633, 0.654 and 0.677 for nonspeech voices, and 0.769, 0.895 and 0.874 for speech/song. Both shared tests passed Holm correction, p=0.03125. All required person intervals and fixed-fit person-and-sound ranges were above 0.5. One adjusted MRI person score was below chance. An exact gain model exposed a centering-order error that was fixed before confirmation. Equal true direction distributions with unequal measurement noise also gave high novelty, so this ranking does not identify a new neural direction or a common cause. Source selection used the full panel. Separate cohorts, fixed sounds, clinical sampling, source filters and measurement differences limit scope. Code, source pins and terms, all results, figures and repeat steps are supplied.

human auditory cortex · SEEG · fMRI · voice · nonvocal reference · cross-modal test · repeat alignment · measurement noise