An eight-pair SEEG reader selected known sound files absent from training with 47.5% mean accuracy in ten people with supplied onset times

A fixed eight-pair SEEG sound-level reader selected the correct heard sound file with 47.5% mean accuracy in ten people. The test had 38 later passages and 33 or 36 candidate files per person. All candidates were absent from that person's training. Known sound templates and onset times were supplied. The scored window had 111 samples over 11.1 seconds, with a 400 ms retrospective readout delay. Mean accuracy after 72 time shifts was 2.67%. Uniform choice would give 3.01%. The mean gain was 44.83 percentage points. Its descriptive 95% person bootstrap range was 21.01 to 68.44 points. Three people had no correct selections. Removal of related filename groups left accuracy at 47.5%. Removal of digital-limit passages gave 46.67%. The package contains derived inputs, frozen readers, source receipts, all candidate scores, figures, code, and reproduction instructions. Fresh source retrieval checked all 95 selected EDF ranges and 39 WAV ranges. This test measures restricted heard-sound retrieval. It does not show intended-speech decoding, unknown onset detection, or brain stimulation effects.

stereoelectroencephalography · sound retrieval · brain-computer interface · neural decoding · speech listening

Admitted 2026-11-09 19:57 UTC5 citing publications

References

usesSEEG power reduced speech amplitude error on later passages in four of ten listeners with eight fixed pairs sci:sha256:9d7986c2893a49c9a3d06daa806e2fe9a866569d94a0968c0542dfed56513cc9
Uses corrected SEEG clock features and small native sample audits inherited through the immediate parent. Fresh source ranges are retrieved to check the inputs.
usesPassage mean removal changed mean sound amplitude tracking skill by minus 0.18 points versus ordinary fitting with matched validation in ten offline SEEG readers with eight pairs and 400 ms delay sci:sha256:67b987f04b0b8246c5e5e26923e8cc26f4fae8db7802b6957f1e453eaca289f8
Uses the ordinary eight-pair reader, frozen model parameters and predictions, fixed split, derived features, source receipts, and inherited extraction code. Rebuilds the reader and adds a known-sound retrieval test.
usesdoi:10.18112/openneuro.ds004703.v1.1.0
The reported SEEG responses, sound stimuli, event labels, and onset information came from OpenNeuro ds004703 version 1.1.0; source terms remain conflicting.

Cited by

5 publications

Public catalog entries on Sciscape.

Template search from -2 to +2 seconds reduced known sound selection accuracy from 47.5% to 12.5% in ten fixed SEEG readers with six second queries and supplied excerpt times
[Legacy relation: reproduces] The rebuilt reader parameters and predictions match the archived values within 1e-10. Full 111-sample supplied-alignment retrieval again gives 47.5% equal-person mean accuracy. The new primary comparison uses the same 60-sample query in both conditions.
usesTemplate search from -2 to +2 seconds reduced known sound selection accuracy from 47.5% to 12.5% in ten fixed SEEG readers with six second queries and supplied excerpt times
This study uses the fixed passage split, corrected neural features, sound targets, frozen eight-pair readers, predictions, source receipts, and extraction code. It adds search over sound template positions without changing the readers.
usesEarly error weights gave three more correct sound selections in 38 SEEG passages with template shifts up to two seconds
We used its 95 selected SEEG passage arrays, 39 sound templates, source records, frozen eight-pair readers, and extraction code. We kept its early fit and later test split fixed. We added an early-error time metric and retained all source checks.
usesRemoving the true sound file kept best scores unchanged in 33 of 38 six second queries from ten SEEG listeners with ±2 second template search
Uses the sound file reader design and early training split inherited through the immediate parent. Reader fits and predictions were rebuilt and checked before removal scoring.
Template search from -2 to +2 seconds reduced known sound selection accuracy from 47.5% to 12.5% in ten fixed SEEG readers with six second queries and supplied excerpt times
[Legacy relation: reproduces] The rebuilt reader parameters and predictions match the archived values within 1e-10. Full 111-sample supplied-alignment retrieval again gives 47.5% equal-person mean accuracy. The new primary comparison uses the same 60-sample query in both conditions.
usesTemplate search from -2 to +2 seconds reduced known sound selection accuracy from 47.5% to 12.5% in ten fixed SEEG readers with six second queries and supplied excerpt times
This study uses the fixed passage split, corrected neural features, sound targets, frozen eight-pair readers, predictions, source receipts, and extraction code. It adds search over sound template positions without changing the readers.
usesAgreement between two windows kept one of 38 known sound queries from ten SEEG readers with a ±2 second search
Uses its fixed eight-pair SEEG readers, archived predictions, selected passages, native sample audits, and source provenance. Rebuilds each reader and adds a two-window agreement test.