For fixed panels, readers trained on speech ranked other human voices above animal sounds in 20 iEEG patients and 16 fMRI adults

We tested whether neural readers learned from speech also separate other human voices from animal sounds. Separate public studies supplied signed intracranial EEG local field potentials and BOLD fMRI. Four people per study supplied development. The fixed test used 20 other patients and 16 other adults. Each reader learned speech versus non-vocal sounds. Both test classes were excluded from training. Training scaling, clipping, spatial mean removal and vector length removal defined the reader. Mean AUC was 0.7016 in iEEG and 0.6038 in fMRI. The 95% person Student intervals were 0.6563 to 0.7469 and 0.5695 to 0.6382. Every reserved person exceeded 0.5. The fixed shared rule for signs and means passed, p=0.0000153. Sound uncertainty limited this result. Resampling people and training and test sounds with reader refits gave ranges of 0.6167 to 0.7598 and 0.4072 to 0.7832. Thus, the result did not establish broader transfer to new sound samples in fMRI. A common training-label permutation reference gave p=0.0500 in fMRI. An early electrical window and one fMRI repeat did not pass parts of the primary rule. Acoustic-only transfer was strong for the electrical sound panel. A constructed rank-one signal passed the normalization, so successful transfer does not identify several neural sources or a non-acoustic code. A later leading-pattern removal check did not pass the shared sign rule. This is a new endpoint test in reused cohorts. Clinical sampling, prior site selection, separate panels and measurement filters limit scope. The package gives licensed inputs, fixed source versions, all scores, code, uncertainty draws, figures and repeat instructions.

human auditory cortex · voice perception · cross-modal generalization · intracranial EEG · fMRI · stimulus uncertainty · cross decoding

Admitted 2026-11-09 20:23 UTC14 citing publications

References

usesWith clipped acoustic inputs, rank 2 models had lower mean test error than rank 1 for fixed panels from 20 iEEG patients and 20 fMRI adults sci:sha256:a7ebe1050fc6e865343e629ad39c22ef72edc352e2f1390afdc35c73a0231229
We reuse its CC BY electrical derivative, fixed source records and MIT fMRI reader.
usesFor two fixed sound panels, neural models trained without speech ranked speech above animal sounds in 20 iEEG patients and 20 fMRI adults sci:sha256:a475145e49c548c9f12b7592bb0fbc3fca3399d536d7b26272b238b0fa65635a
Its transfer test in the opposite direction motivated the new endpoint. We reuse its source labels through the cited derivative.
usesdoi:10.5281/zenodo.20599001
Version 1.1 supplies the source electrical responses, acoustic descriptors and sound labels.
usesdoi:10.1016/j.cub.2026.07.063
Source human intracranial study and earlier evidence for graded voice sensitivity.
usesdoi:10.1152/jn.00588.2020
B2021 supplies the separate 20-person fMRI cohort.
usesdoi:10.1371/journal.pbio.3002366
Its fixed author archive supplies the B2021 responses and metadata.
usesdoi:10.1016/j.cub.2022.01.069
The pinned source repository supplies fMRI sound classes and acoustic descriptors.
doi:10.3389/fnins.2022.1075288
Earlier fMRI evidence that temporal voice area responses extend beyond speech.

Cited by

14 publications

Public catalog entries on Sciscape.

usesFixed speech–music readers did not confirm shared animal sound ranking in 20 iEEG patients and 16 fMRI adults
We reuse its licensed compact electrical inputs, source access patterns and source records. Its different speech-to-human-voice transfer test informs the present question.
usesA two-pattern sound forecast did not pass a shared 1% gain test in 17 iEEG patients and 16 fMRI adults
We reuse its licensed electrical inputs, source records and BOLD extraction code. Its one-pattern audit motivates this new forecast endpoint.
usesSound labels lowered test error for averaged responses in 20 iEEG patients and 16 fMRI adults; separate BOLD fits gave no clear gain
We reuse its CC BY 4.0 compact electrical means and labels, with attribution and source records.
usesFixed speech readers kept human preference in 20 iEEG patients and 16 fMRI adults; animal transfer was not shared
We reuse its licensed compact electrical derivative and MIT input reader. We extend its speech-to-human transfer question to a disjoint other test panel and a direct animal transfer endpoint.
usesFor two fixed non-vocal sound panels, pair relations agreed across 15 iEEG patients and 12 fMRI adults after one training pattern was removed
We reuse its unchanged licensed patient-labelled electrical derivative, source provenance and MIT fMRI input adapter.
usesClipped acoustic fits did not pass a shared 1% voice pattern error test in 20 iEEG patients and 16 fMRI adults
We reuse the permitted electrical derivative, source records, and input routines.
usesFor fixed sounds, kernels from four source people lowered mean error in iEEG and B2021 fMRI; NH2015 gain was uncertain
Original compact electrical derivative in the preserved source chain.
usesUp to four class examples reduced mean response pattern error after two amplitude controls in 20 iEEG patients and 16 fMRI adults
Original licensed electrical derivative in the verified SCI input chain.
usesUp to eight class examples lowered mean acoustic prediction error at fixed panel sizes in human iEEG and fMRI
Credits the compact electrical data donor and its source history.
usesFor fixed sounds, speech readers kept mean voice AUC above 0.5 after at least half of non-vocal variation was removed in iEEG and fMRI
We use its CC BY electrical summary inputs. Its earlier transfer and exploratory one-pattern removal results motivated this fixed energy removal test.
usesA correction from up to four sound class examples lowered mean model error in 20 iEEG patients and 16 fMRI adults
Upstream licensed compact electrical derivative.
usesExploratory tests found site pair map agreement across fixed sound panels in pre-onset iEEG from 17 patients and fMRI estimate differences from 16 adults
Verified patient-labelled electrical derivative and source reader.
usesModels fit to voice sounds did not pass a shared animal sound test in 20 iEEG patients and 16 fMRI adults
We reuse its licensed electrical derivative, source records and MIT fMRI reader infrastructure. The new endpoint is response variation within animal sounds after voice-only fitting.
usesA fixed acoustic distance rule failed shared error tests in 20 iEEG patients and 16 fMRI adults
Upstream compact electrical derivative and its provenance.