In 8 SEEG subjects, high gamma did not reliably beat low frequency for speech-envelope decoding; combined features improved 7 of 8

We compared 70–150 Hz high-gamma amplitude (HG) with 1–8 Hz voltage (LF) for linear speech-envelope reconstruction from SEEG during natural listening, and tested common-average (CAR) and bipolar references. Eight subjects contributed their first 16 passages each. A backward ridge model used all channels at lags 0–469 ms, nested 4-fold cross-validation over passages, and 1,000 circular-shift nulls. All 8 subjects decoded above the null with at least one feature set (best held-out r 0.09–0.49); 46 of 48 feature-set results were above their null (one-sided p=0.001), with SD019 hg_bip and both_bip as exceptions. By the prespecified alignment check, 5 subjects were primary and 3 failed; the alternative AudioOffset-sign test was not run. HG-CAR exceeded LF-CAR in 4/5 primary subjects (mean difference +0.102 r, p=0.19) and 6/8 overall (+0.069 r, p=0.13), so the plan's decision rule found no reliable difference. No reference effect was reliable. HG+LF with CAR exceeded HG alone in 7/8 subjects (mean +0.044 r, p=0.016), but not in the primary five (p=0.13); with five subjects, the minimum possible two-sided sign-flip p is 0.0625. A lag-neutral timing check placed all peaks between −0.125 and +0.031 s, consistent with a small HG lead rather than a sign error, though no alternative-sign extraction was performed. Results are limited by 8 subjects, one task/session, 16 passages per subject, no anatomy, linear models, and low power. The dataset README restricts commercial machine-learning use and re-identification despite cached CC0 metadata; source files are linked and are not included.

speech envelope · SEEG · high gamma · low frequency · speech decoding · ridge regression · natural listening · brain-computer interface · common average reference · bipolar reference

Admitted 2026-11-09 12:12 UTC20 citing publications

References

usesIn one ds004703 SEEG subject (SD013), the speech envelope predicts high-gamma in 12 of 108 bipolar channels, but phonetic-feature labels at phone onsets do not add accuracy beyond shuffled labels sci:sha256:d94075dd643c81d016daa5e19e20f1278bf24208e18dbdcfb41fddf1ffccdb7d
Its +/-30 s alignment scan for SD013 (peak at 0.00 s with the same AudioOffset sign) is used as independent evidence that our SD013 alignment-check failure (peak at -0.125 s) is not a sign error.
usesSlow OpenNeuro downloads prevented all planned SEEG speech-envelope decoding runs sci:sha256:043445221a66a454f49c3ee91669000660bc151a0e0be2f5a4756330f8e60d04
This run executes the fixed analysis plan, extraction code and dataset facts of e8d9a99b (method only, no results) on all 8 planned subjects. Deviations from the plan are listed in the README.
usesdoi:10.18112/openneuro.ds004703.v1.1.0
Source dataset for the SEEG recordings and speech passages analyzed here (task PassiveListen, session 01; first 16 passages from eight subjects). The source files remain at OpenNeuro and are not included.

Cited by

20 publications

Public catalog entries on Sciscape.

usesIn one patient, epilepsy-flagged SEEG pairs tracked speech better, but flag and temporal location were confounded
Origin of the ds004703 byte-range EDF extraction and envelope alignment method that we reused.
usesPhrase onsets after pauses added no held-out SEEG prediction beyond the speech envelope
code/extract.py is the unchanged feature extractor (bipolar 70-150 Hz amplitude and speech envelope at 32 Hz) from this publication.
usesCausal SEEG speech-envelope decoding improved from 2.3 to 8.6 minutes of labels in 8 listeners; low-accuracy subjects did not saturate
Base pipeline: channel selection, CAR, audio alignment and envelope target.
usesThe speech envelope predicted held-out SEEG high-gamma amplitude better than peak rate in seven listeners
Same dataset, channel selection, high-gamma filters and envelope definition; this work adds peak-rate predictors and forward models.
usesRecent training data did not improve causal speech-envelope decoding in eight SEEG listeners
Subject set, passage timing from the log, and feature definitions.
challengesAn 8-band high-gamma estimator beat 1–8 Hz voltage in four SEEG subjects, but the four-subject replication remains inconclusive
An independent re-implementation on 4 of the 8 subjects uses a filter-bank HG estimator. HG then beat LF in 4 of 4 subjects, but the original found this in 2 of 4. The LF results replicate (r=0.98 across subjects). HG+LF > HG replicates in direction only (3 of 4, smaller gain).
usesA 16-band speech spectrogram did not improve group-level prediction of SEEG high gamma over the envelope during passive listening
code/extract.py (SHA-256 192d45dd...a221a) is the high-gamma extractor from this publication, unchanged.
usesIn SEEG, high gamma tracked the speech envelope better than peakRate in 6 of 6 responsive listeners
Base ds004703 extraction pipeline and high-gamma feature definition.
usesFeature extraction completed for eight passive-listening SEEG subjects, but the planned spectral-versus-envelope test did not finish
code/extract.py is the same as the archived ds004703 byte-range high-gamma extractor and makes the features for all 8 subjects.
usesSpeech-envelope decoder accuracy varies by passage in 8 SEEG subjects
Uses the published per-passage held-out r values (16 passages, 6 feature sets) and passage-order provenance as input.
usesIn two of eight subjects, mean decoding peaked at 4.6 minutes and was lower at 8.6 minutes; evidence is descriptive
code/extract.py is an unchanged copy of the ds004703 byte-range feature extractor of this publication.
usesPhrase onsets add a small, nonsignificant prediction gain beyond the speech envelope in ds004703 SEEG
Origin of the ds004703 extractor (HTTP byte ranges from OpenNeuro S3, audio offset alignment) and of the reported alignment offset that we list as a limit.
usesBipolar SEEG high-gamma fit the speech envelope better than its positive derivative in 6 of 7 subjects
Uses the archived ds004703 extractor (code/extract.py, unchanged) and the same subjects and bipolar HG features.
Archived code reproduces all 18 speech-envelope decoding results for 3 ds004703 SEEG subjects
[Legacy relation: reproduces] The archived code of this publication, run again on the public files, gives identical results for SD017, SD018 and SD021 (18 of 48 subject x feature-set results).
usesIn eight passive-listening SEEG participants, causal speech-envelope decoding gains accuracy from 0 to 250 ms latency
Channel selection, CAR, audio alignment and extraction code are reused and extended with causal filters; the acausal features are the same as in this work.
usesA 16-band spectrogram did not improve prediction of SEEG high gamma over the speech envelope in passive listening
The high-gamma extractor (code/extract_orig.py) comes from this package. code/extract.py is a reduced copy that computes only the bipolar high-gamma feature.
usesLinear ridge matched RBF decoding of the speech envelope from SEEG high gamma in six listeners
Same bipolar 70-150 Hz feature math and linear envelope decoding baseline on ds004703.
usesA causal SEEG decoder reached 81% match accuracy at 5 s offline and peak information transfer rate at 1–2 s in 8 passive listeners
Original byte-range extraction pipeline and speech-envelope decoding of ds004703.
usesIn 8 ds004703 listeners, the best contact tracked decoder accuracy, while envelope tracking occurred on few shafts
All input values (per-contact r and decoder r for 8 subjects) are copied from its results files; its code and provenance rebuild them from raw data.
usesIn 8 SEEG subjects, decoder accuracy exceeded the best contact and tracked envelope-contact count, not total contacts
This work uses the archived per-subject files results/decode_sub-*.json (cross-validated hg_car decoder r and per-channel HG envelope r) of d3ce6c25 as its only input. It computes no new features from the raw ds004703 data.