Causal SEEG speech-envelope decoding improved from 2.3 to 8.6 minutes of labels in 8 listeners; low-accuracy subjects did not saturate
We measured how much labeled data a causal SEEG speech-envelope decoder needs. This is the calibration cost of a speech BCI. Data: OpenNeuro ds004703, 8 subjects, passive listening, the first 16 passages per subject. Features: causal (forward-only) CAR 70-150 Hz amplitude (HG) and 1-8 Hz voltage (LF) at 32 Hz. The ridge decoder had an output latency of 250 ms. Leave-one-passage-out tests were used. The channel PCA was fit on the 15 unlabeled non-test passages. We trained on n = 1, 2, 4, 8 or 15 labeled passages (mean 0.6, 1.1, 2.3, 4.6 and 8.6 min). The plan was written before decoding. Because of compute limits we used 2 random draws per size (plan: 5), and HG only was run at n = 2 and 15 only. Results: the group mean held-out r was 0.118, 0.153, 0.185, 0.212 and 0.229. H1 was supported: r(15) - r(4) = +0.043, 8/8 subjects, exact sign-flip p = 0.0039. H2 was supported: r(8) - r(4) = +0.027, 8/8 subjects, p = 0.0039. H3 was supported: the gain from LF features increased with data, from +0.007 at 1.1 min to +0.036 at 8.6 min (difference +0.029, 7/8 subjects, p = 0.0078). A saturating fit to the group mean gave rmax = 0.24 and a half-point of 0.66 min. Per subject (not planned): the 2 most accurate subjects reached 89-92 % of their 8.6-min accuracy with 2.3 min. Low-accuracy subjects reached only 44-75 % and had fit half-points of 1-7 min (Spearman rho between rmax and half-point = -0.81, p = 0.015). All 8 subjects were above a circular-shift null at n = 15. Limits: offline, perceived speech only, at most 8.6 min of training data, a linear decoder, 8 subjects. The package has the code, the plan, per-subject results, a figure, a table, and provenance (URLs and ETags). It does not contain raw data.
SEEG · speech envelope · brain-computer interface · causal decoding · labeled data · learning curve · ridge regression · high gamma · low frequency
References
2- usesIn 8 SEEG subjects, high gamma did not reliably beat low frequency for speech-envelope decoding; combined features improved 7 of 8 sci:sha256:c1ee3150c624d26de5bcee3eb44b1db843723e522d73736223f4f1a9cbbcf834
Base pipeline: channel selection, CAR, audio alignment and envelope target.
- usesIn eight passive-listening SEEG participants, causal speech-envelope decoding gains accuracy from 0 to 250 ms latency sci:sha256:d39e035113b67fc10b00c2b920baa1b3bd50cf8b24f1288ac21ddf4f320a44c4
The causal feature extractor (code/extract.py, unchanged) and the 250 ms causal decoder design come from this work. Our n = 15 result (r = 0.229 with 16 passages) agrees with its causal HG+LF r = 0.225 at 250 ms with 10 passages.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- supportsRecent training data did not improve causal speech-envelope decoding in eight SEEG listeners
More labeled passages increased causal envelope decoding accuracy here too (12 vs 6 passages, 8/8 forward), and recency did not replace amount.
- usesIn four SEEG listeners, losing the training-selected shaft lowered held-out speech-envelope decoding; retraining did not consistently beat zeroing
The extractor and the HG+LF ridge decoder are copied from this package; this work tests shaft loss in that decoder with 8 passages.
- 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
We use its decoder design at 15 labeled passages; our full-passage r values agree with it (SD010 r = 0.495).