2026 · arXiv · Field expert review · confidence medium

Do EEG Foundation Models Transfer to Speech? A Benchmark on Overt and Imagined Speech Decoding

Owais Mujtaba Khanday, Mohamed Baha Ben Ticha, Sanae Belfrouh, Marc Ouellet, Jose A. Gonzalez-Lopez

BibTeX

@misc{do-eeg-foundation-models-transfer-to-speech-a-benchmark-on-overt-and-imagined-speech-decoding,
  title = {Do EEG Foundation Models Transfer to Speech? A Benchmark on Overt and Imagined Speech Decoding},
  author = {Owais Mujtaba Khanday and Mohamed Baha Ben Ticha and Sanae Belfrouh and Marc Ouellet and Jose A. Gonzalez-Lopez},
  year = {2026},
  note = {arXiv},
  eprint = {2607.27268},
  archivePrefix = {arXiv},
  url = {http://arxiv.org/abs/2607.27268v2},
}

General EEG pretraining shows no consistent speech-decoding advantage here; word decoding remains weak, and test-subject-informed early stopping limits the unseen-user claim.

Verdict: full-text draftPriority: highConfidence: mediumBasis: full text + summaryCoverage: high

Reading guidance

Verdict: full-text draft · priority high · confidence medium
Why it matters: A speech-focused negative-result benchmark that separates mode detection from lexical decoding and makes the limited cross-subject utility of the tested models visible.
What to trust: Basis: full text + summary. Coverage: high. 11 evidence records back the review.
What is weak: Low lexical signal, small cohorts and fixed vocabulary; EEGConformer collapses to chance on all UGR tasks under this training configuration. No random-initialization ablation of the same foundation architectures establishes the causal benefit or harm of pretraining. Section 2.4 explicitly uses the held-out test subject for early stopping and restores the best checkpoint; although gradients do not use that subject, model selection is test-informed. BCIC results use the official validation split, not a hidden test set. Multiple training seeds are left to future work. Table 1 lists 3/6/30 classes whereas Table 2 and the word-analysis text use 60 word classes. Only offline classification is reported. No online communication interface, assistive-user trial, walking test, latency, power consumption, or mobile execution measurement is provided. Two EEG datasets, two foundation-model families, and finite-label offline tasks. Spanish lexical tasks combine overt and covert trials; they do not establish a silent-only continuous decoder. No free-form thought reading, audio reconstruction, or open-vocabulary communication is tested. Overclaim risk: High if speech-mode accuracy is described as word recognition or the protocol as leakage-free unseen-user evaluation; moderate for the narrower comparative negative result..
Read before: SSI review rubric
Read next: SSI archive

Axes

Task: speech-recognition
Modality: eeg
Hardware: Existing 64-channel scalp EEG recordings. UGR-MINDVOICE uses a 10–20 layout, 1000 Hz acquisition and FCz reference, retaining 62 scalp channels after preprocessing; both datasets are resampled to 200 Hz. No new hardware.
Body site: brain
Output: labels
Vocabulary: closed-set word/phrase and category labels; no open vocabulary
Metrics: UGR-MINDVOICE Table 2: speech-mode accuracy EEGNet 61.3% (95% CI 56.5–66.1), LaBraM 57.3% (53.0–61.7), EEGMamba 56.6% (51.8–61.3), versus 33.3% chance; EEGNet semantic-category accuracy 19.9% versus 16.7% chance; 60-word accuracy across models 1.7–2.9% versus 1.7% chance. BCIC2020-T3 Table 3, mean ± SD across 15 participants: within-subject balanced accuracy EEGNet 26.7 ± 3.3%, EEGConformer 30.2 ± 4.4%, LaBraM 27.7 ± 2.7%, EEGMamba 25.7 ± 2.5%; cross-subject range 18.9–20.2%, all p > 0.18 against 20% chance. These are reported results, not independently reproduced measurements.
Evaluation mode: Offline UGR-MINDVOICE leave-one-subject-out evaluation over 14 participants, with test-subject-informed early stopping; semantic and word tasks pool overt and covert examples. BCIC2020-T3 reports within-subject and leave-one-subject-out results over 15 participants. Accuracy or balanced accuracy, weighted F1, Cohen kappa, and Wilcoxon tests are reported.
Review confidence: medium
Overclaim risk: High if speech-mode accuracy is described as word recognition or the protocol as leakage-free unseen-user evaluation; moderate for the narrower comparative negative result.

Expert take

This is a useful caution for EEG-based silent communication: scaling a model pretrained on general EEG does not automatically improve speech-content decoding. In UGR-MINDVOICE, EEGNet reaches 61.3% accuracy on overt/covert/rest classification, compared with 57.3% for LaBraM and 56.6% for EEGMamba. Those are speech-mode results, not word-recognition accuracy. On the 60-word task, all models reach only 1.7–2.9% accuracy against 1.7% chance. In BCIC2020-T3, within-subject balanced accuracy is 25.7–30.2% for five imagined words/phrases, but cross-subject results are 18.9–20.2%, with no model significantly above 20% chance. The value is therefore a controlled comparison and a warning against importing foundation-model expectations into SSI, rather than a usable communication system. A major qualification is that Section 2.4 uses the held-out test subject to choose when training stops and which checkpoint is evaluated. This is test-informed model selection even without gradient updates, so UGR-MINDVOICE results are not a clean estimate of performance on a wholly untouched user. The paper also relies on official BCIC validation data, leaves multiple training seeds to future work, and contains a 30-versus-60-class inconsistency between its dataset table and word results. The experiments support a lack of consistent advantage for these two models under the reported conditions; they do not prove that all EEG pretraining fails or that speech-specific pretraining would solve the problem.

True value

A speech-focused negative-result benchmark that separates mode detection from lexical decoding and makes the limited cross-subject utility of the tested models visible.

What changed

Canon before

The paper situates its comparison against general EEG pretraining successes on non-speech tasks and small-vocabulary imagined-speech studies. Such results do not by themselves establish transfer to speech production.

Delta from canon

Evaluates whether generic pretrained EEG representations help speech-related classification, with compact baselines and separate speech-mode, semantic-category, word-identity, within-subject, and cross-subject results. This is an adjacent imagined-speech BCI benchmark, not articulatory SSI recognition.

Position in field

An adjacent imagined-speech EEG/BCI benchmark with negative transfer findings; useful for SSI evaluation methodology but not a demonstrated articulatory or deployable silent-speech interface.

Evidence

“ UGR-MINDVOICE starts with 15 participants and 22 sessions; excluding subject 15 for ICA errors leaves 14 evaluated participants. BCIC2020-T3 contains 15 healthy participants and five imagined words/phrases. ”

validation_scope · Sections 2.1 and 2.3; Table 1; PDF p. 2 · confidence 0.99

“ Table 2 reports speech-mode accuracy of 61.3% for EEGNet, 57.3% for LaBraM, and 56.6% for EEGMamba against 33.3% chance. This task distinguishes overt, covert, and rest states rather than individual words. ”

metric · Section 3.1; Table 2; PDF pp. 3–4 · confidence 0.99

“ The UGR-MINDVOICE 60-word task yields 1.7–2.9% accuracy across the six models against 1.7% chance. EEGNet reaches 2.8%, LaBraM 1.9%, and EEGMamba 2.9%. ”

metric · Table 2, Word (60-cl) columns; PDF p. 4 · confidence 0.99

“ BCIC2020-T3 within-subject balanced accuracy is 26.7 ± 3.3% for EEGNet, 30.2 ± 4.4% for EEGConformer, 27.7 ± 2.7% for LaBraM, and 25.7 ± 2.5% for EEGMamba; these are means and standard deviations over 15 participants for five classes. ”

metric · Table 3, Within-subject columns and caption; PDF p. 4 · confidence 0.99

“ BCIC2020-T3 cross-subject balanced accuracy ranges from 18.9% to 20.2%; no model significantly exceeds 20% chance, with all reported one-sided Wilcoxon p-values greater than 0.18. ”

metric · Table 3 and Section 3.1 continuation; PDF p. 4 · confidence 0.99

“ Section 2.4 says the held-out UGR test subject determines early stopping and that the best checkpoint is restored. Reviewer assessment: excluding this subject from gradient updates does not make the evaluation independent, because the test subject informs model selection. ”

limitation · Section 2.4, first two paragraphs; PDF p. 2 · confidence 0.99

“ The authors report BCIC2020-T3 results on the official validation split because official test labels are unavailable. Absence of this dataset from pretraining, as asserted by the authors, does not resolve the separate issue of validation/test separation during model selection. ”

limitation · Section 2.1 final paragraph; Section 2.4; PDF pp. 2–3 · confidence 0.99

“ Table 1 lists UGR classes as 3/6/30, but Table 2 and Figure 1 label word identity as 60 classes. This review reports Table 2 values and preserves the inconsistency for clarification. ”

limitation · Table 1, PDF p. 2; Figure 1, PDF p. 3; Table 2, PDF p. 4 · confidence 0.99

“ UGR semantic-category and word models train on pooled overt/covert labels. Their separate overt-only and covert-only evaluation shows no statistically significant difference, all p > 0.11; this is not a silent-only training protocol. ”

validation_scope · Section 2.4, PDF p. 2; Section 3.2, PDF p. 4 · confidence 0.98

“ Multiple-seed experiments and speech-specific pretraining are explicitly future work. Reviewer assessment: speech-specific pretraining is a proposed direction, not an experimentally demonstrated remedy. ”

limitation · Section 4, final sentences; PDF p. 4 · confidence 0.98

“ The paper compares pretrained LaBraM and EEGMamba with established baselines across speech-mode and lexical tasks on two datasets. Reviewer assessment: its contribution is comparative evidence about transfer, not a new decoder or usable speech interface. ”

actual_novelty · Abstract; Sections 2.2 and 3; Tables 2–3; PDF pp. 1–4 · confidence 0.96

Limits

Technical limits

Low lexical signal, small cohorts and fixed vocabulary; EEGConformer collapses to chance on all UGR tasks under this training configuration. No random-initialization ablation of the same foundation architectures establishes the causal benefit or harm of pretraining.

Evaluation limits

Section 2.4 explicitly uses the held-out test subject for early stopping and restores the best checkpoint; although gradients do not use that subject, model selection is test-informed. BCIC results use the official validation split, not a hidden test set. Multiple training seeds are left to future work. Table 1 lists 3/6/30 classes whereas Table 2 and the word-analysis text use 60 word classes.

Deployment limits

Only offline classification is reported. No online communication interface, assistive-user trial, walking test, latency, power consumption, or mobile execution measurement is provided.

Scope limits

Two EEG datasets, two foundation-model families, and finite-label offline tasks. Spanish lexical tasks combine overt and covert trials; they do not establish a silent-only continuous decoder. No free-form thought reading, audio reconstruction, or open-vocabulary communication is tested.