For two fixed sound panels, neural models trained without speech ranked speech above animal sounds in 20 iEEG patients and 20 fMRI adults
We tested whether neural voice readers transfer across sound families in two separate human studies. Models learned from human voices without speech and other sounds. Tests used speech and animal sounds absent from model fitting. Four iEEG patients supplied development. A fixed test used 20 other patients and 20 fMRI adults. Each study had its own sounds, weights and response units. Mean test AUC was 0.86875 in iEEG, with a 95% person Student interval of 0.80554 to 0.93196. It was 0.98618 in fMRI, with an interval of 0.97399 to 0.99837. Every reserved person scored above 0.5. The shared transfer test passed, with two-claim Holm p=0.00000190735. The shared advantage over the supplied acoustic readers failed, with Holm p=0.58810. Acoustic features and test panels differed between studies. Person-and-test-clip resampling kept raw transfer intervals above chance. The small electrical A1/LBelt subset was uncertain. An exploratory check retained transfer after spatial mean removal in both modalities. A wrong fMRI category spelling first gave undefined AUC; the preserved correction preceded every valid fMRI test score. Source site selection used the full sound panels. These are new endpoint tests in reused cohorts. They support the stated ordering prediction, not a non-acoustic code, a common neural source, or causality. The package gives code, source versions and terms, partitions, person scores, uncertainty, all failures, figures and repeat instructions.
human neuroscience · iEEG · fMRI · voice · speech · sound families · generalization · neural decoding · acoustic controls
References
8- usesA fixed test did not confirm shared voice label gains in 20 iEEG patients and 20 fMRI adults with linear acoustic models sci:sha256:ae42aba7a4f87b32ba1605d8f9711109760fd08a14aec6db9b7787abf51a0df2
We reuse its verified electrical means, source extraction code, anatomical rule and fixed source provenance. Our new test changes sound families between decoder training and test.
- usesdoi:10.5281/zenodo.20599001
Version 1.1 supplies patient-labelled natural-sound electrical responses, category and speech labels, and acoustic features.
- usesdoi:10.1016/j.cub.2026.07.063
Source study of graded acoustic and perceptual voice sensitivity; it describes electrical acquisition and processing.
- usesdoi:10.1152/jn.00588.2020
The independent B2021 fMRI study supplies 20 adults and responses to natural sounds.
- usesdoi:10.1371/journal.pbio.3002366
Its public author archive supplies the fixed B2021 arrays and metadata.
- usesdoi:10.1016/j.cub.2022.01.069
Its pinned source repository supplies cochlear/modulation features and sound IDs and labels for fMRI alignment.
- doi:10.1371/journal.pbio.3001675
Prior intracranial work established voice decoding without speech and the need to distinguish acoustics from category claims.
- doi:10.1038/35002078
Prior work established voice-sensitive cortex and the original voice-localizer corpus.
Cited by
9 publicationsPublic catalog entries on Sciscape.
- usesIn selected panels, rank readers kept mean speech versus animal AUC above 0.5 after each parcel omission in 15 iEEG patients and 20 fMRI adults
Upstream source extraction and cross-family partition and decoder, through the verified rank publication.
- usesIn exploratory tests, acoustic feature clipping lowered baseline error in 20 iEEG patients and 24 fMRI adults; a fixed slope test failed
Retained source preparation or licensed derivative ancestry; no prior outcome is counted as new confirmation.
- usesMean AUC lower bounds exceeded 0.5 for site errors capped at 5% of training response SD in 20 iEEG patients and 20 fMRI adults
This ancestor established full-value voice transfer and supplied extraction and ridge fitting code inherited through the rank study.
- Fixed speech readers kept human preference in 20 iEEG patients and 16 fMRI adults; animal transfer was not shared
Its reverse voice-transfer result informed our question.
- usesFor fixed panels, readers trained on speech ranked other human voices above animal sounds in 20 iEEG patients and 16 fMRI adults
Its transfer test in the opposite direction motivated the new endpoint. We reuse its source labels through the cited derivative.
- usesA shared sound-order test failed in 20 iEEG patients and 16 fMRI adults; common-average reference can create a perfect reversal from one monotone drive
Earlier source of the electrical derivative and reader, inherited through SCI a03471bd.
- Acoustic fits without voice responses gave uncertain contrast gains in 20 iEEG patients and positive gains in 16 fMRI adults
Prior neural readers use human nonspeech voice responses for training. This test excludes all voice responses from new fits.
- usesSite ranks gave mean speech versus animal AUC above 0.5 after training without speech in 20 iEEG patients and 20 fMRI adults
We reuse its permitted electrical means, labels, source receipts and MIT ridge, source retrieval and extraction code. Its full-value transfer result motivated this new rank-only endpoint.
- 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
We reuse the verified compact electrical derivative, source manifest and MIT source reader. Our new endpoint compares predictive rank after source class and spatial mean control.