Under person t tests, mean voice transfer was above chance in six SEEG patients and two fMRI groups; the test for all nine clips failed

We tested whether a speech/song-trained neural reader ranks excluded nonspeech human voices above excluded source nonvoice sounds. Separate public studies supplied high-gamma SEEG and BOLD fMRI. Two patients and four MRI adults supplied development. A local code lock preceded new endpoints in six patients, 16 B2021 adults and eight NH2015 adults. Each reader used scales from training data. We removed the spatial mean and set response length to one. We subtracted the pre-event mean from electrical responses. Mean transfer AUCs were 0.63385, 0.82835 and 0.81648. Their 95% person Student limits were [0.57936,0.68834], [0.80462,0.85209] and [0.79383,0.83913]. All 30 person means exceeded 0.5. A shared mean test also required held-speech validation and an electrical gain over a pre-event reader. It passed two-claim Holm correction, p=0.001533. A stronger fixed test required positive mean AUC for every one of nine voice clips in each study. It failed, p=0.789458. The weakest electrical clip had mean 0.45681 [0.33010,0.58353]; its name is unavailable. The scream clip was uncertain in the added MRI study. The minimum clip mean equals the worst AUC over offline mixtures of fixed voice score distributions against the fixed background. This elementary result gives no guarantee for new voices or different background weights. A small-site control left electrical and added-MRI two-sided intervals across chance. Independent arithmetic checked all 30 test people. Source hashes and an offline audit passed. These are new endpoints in reused studies, with one added MRI acquisition and no new electrical study. Small clinical samples, source screens, common task schedules, missing common acoustics, different cohorts, unequal spatial sampling and measurement filters limit scope. An exact single-factor example gave AUC 1 under the same method. No common circuit or causal effect was shown. Code, source pins and terms, all failures, figures and repeat steps are supplied.

human neuroscience · SEEG · fMRI · voice perception · cross-context transfer · sound mixtures · subject-level inference · negative result · reproducible benchmark