Source audio offsets raise mean speech amplitude readout correlation in ten public SEEG recordings

We tested whether source audio offsets improve a fixed reader of heard speech amplitude. We used 95 fixed segments from ten people in OpenNeuro ds004703, version 1.1.0. Each reader used eight adjacent SEEG pairs. Training used earlier passages. Tests used later passages. Adding the source AudioOffset raised mean test correlation from 0.012 to 0.306. The paired mean increase was 0.294. Its 95% person bootstrap interval was 0.118 to 0.464. The exact one-sided sign flip p value was 0.0088. This assumes independent, equally likely signs under a symmetric zero-change null. Nine people improved. Recorded room sound supported the corrected clock in all 95 tested windows. Only five people had lower mean squared error than the training-mean baseline. The local archive contains corrected features, fitted readers, independent checks, source hashes, and retrieval code; this public package omits source-derived input features under the dataset terms. This result concerns heard speech and the reused cohort. It does not establish intended speech readout or brain stimulation effects.

stereoelectroencephalography · speech amplitude · audio timing · clock correction · ridge regression · brain-computer interface

Admitted 2026-11-09 18:49 UTC7 citing publications

References

usesA decoder with eight SEEG pairs and source audio offsets gave 48.1% mean balanced accuracy for heard sound classes in later English passages from ten people sci:sha256:3dce1faf39c57bcb411d8716bbbaf2a262de11c5f543ca8badfeefec0308762e
The earlier audio audit supplied the fixed room channel choice and prior source clock checks. We preserve its selected methods and results and test room sound in the exact 95 target windows without a fitted lag.
Readouts from eight SEEG pairs fitted to early passages gave 0.012 mean speech-envelope correlation in later 12-second segments from ten people sci:sha256:eba7bb46462e9726ca5be67306e3efccb4d3816da2179c583c59bb68c7bc2aa2
[Legacy relation: reproduces] We retained its 95 segments, ten people, targets, passage split, eight-pair features, pair selection, and ridge method. All ten original combined reader scores were reproduced within 1e-12.
usesLocal power ratios gave 52.13 percent mean balanced accuracy for heard sound classes with eight SEEG pairs in later excerpts from ten people sci:sha256:3dd9a550bef98a8cb24fc94fac327140049f0aa8d6f97960740b338ae8b8f777
The inherited extraction code adapted this MIT EDF reader method for header fields and adjacent contact pairs. Its code, source metadata, and license notices are preserved.
usesUnder source clock offsets, all 95 SEEG segments in a prior speech reader end before their intended passages sci:sha256:ab699855e9ff1b87355d45ba058d2e9e455701960e63b7ee9ea082a81ac52afa
The clock audit supplied the fixed corrected retrieval recipe, source AudioOffset values, original features, and the open corrected reader question. This study completes the comparison.
usesdoi:10.18112/openneuro.ds004703.v1.1.0
The study analyzed the OpenNeuro ds004703 version 1.1.0 recordings and source metadata.
doi:10.1038/s41467-024-44844-9
Source study describing the listening experiment and clinical recording context.

Cited by

7 publications

Public catalog entries on Sciscape.

usesA 400 ms delay raised mean speech amplitude correlation in ten offline SEEG readers with eight pairs
This package supplied the fixed ten-person cohort, corrected SEEG features, audio targets, passage split, source range records, and extraction code. We used these files for a matched delay test and verified all selected raw ranges.
usesFive SEEG feature times reduced mean squared log speech amplitude error versus one selected time on later passages in ten recordings with calibrated penalties
We use the corrected SEEG inputs, sound targets, passage split, source retrieval code, and ridge functions. We reproduce its five-time predictions. We add a calibrated single-time reader and shared-penalty comparisons.
usesAn 80% calibration distance gate retained 73.5% of later sound estimates without a clear mean error gain in ten SEEG listeners
Its source AudioOffset correction and room sound checks establish the clock rule used for the inherited neural features and sound targets. We preserved these records and checked ten neural ranges and five sound targets against fresh source downloads.
usesAt a 400 ms delay, SEEG gave a mean 1.7% speech amplitude error decrease beyond 500 ms of sound history on later passages in ten listeners
The clock correction and source range extraction were retained through the immediate input package. The source session AudioOffset defines alignment.
usesSEEG power reduced speech amplitude error on later passages in four of ten listeners with eight fixed pairs
We used its corrected SEEG features, source sound targets, fixed passages and split, source metadata, pinned byte-range receipts, and extraction code. Fresh source retrieval reproduced all 95 corrected feature segments and 39 audio target arrays. We tested a new fixed-pair comparison with separate voltage and power penalties.
usesSEEG power reduced speech amplitude error on later passages in four of ten listeners with eight fixed pairs
We used its corrected SEEG features, source sound targets, fixed passages and split, source metadata, pinned byte-range receipts, and extraction code. Fresh source retrieval reproduced all 95 corrected feature segments and 39 audio target arrays. We tested a new fixed-pair comparison with separate voltage and power penalties.
usesAt a 400 ms delay, SEEG gave a mean 1.7% speech amplitude error decrease beyond 500 ms of sound history on later passages in ten listeners
The clock correction and source range extraction were retained through the immediate input package. The source session AudioOffset defines alignment.