Pith. sign in

REVIEW 3 major objections 6 minor 32 references

EEGAlign claims that jointly aligning EEG to text embeddings and speech features—not either alone—enables closed-set Chinese sentence decoding from scalp EEG, reaching 82.37% Top-1 on reading-aloud and 41.43% on passive-listening data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 01:49 UTC pith:3OKIFE76

load-bearing objection Genuine complementarity finding under a load-bearing evaluation flaw: the headline accuracies come from best-epoch-on-test selection, so the margins over baselines won't hold as reported. the 3 major comments →

arxiv 2607.25626 v1 pith:3OKIFE76 submitted 2026-07-28 cs.AI

Joint Text-Audio Alignment for EEG-to-Text Decoding in Chinese Speech Production and Perception

Tian Zheng , Xurong Xie , Xinxin Zhu , Xiaolan Peng , Feng Tian This is my paper
classification cs.AI
keywords EEG-to-text decodingChinese speech decodingcontrastive learningtext-audio alignmentCTC decodingclosed-set sentence classificationBGE-M3wav2vec 2.0
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces EEGAlign, a framework that decodes Chinese sentences from non-invasive EEG by aligning brain signals to two complementary supervision axes: text semantics (via BGE-M3 embeddings) and audio acoustics (via wav2vec 2.0 features), alongside a CTC character decoder. It claims that neither axis alone is sufficient: in ablations on the reading-aloud condition, text-only alignment drops to 56.92% and audio-only to 62.63%, both below the 66.28% achieved by CTC alone, whereas the combined model reaches 69.62% on the lightweight Conformer and 82.37% with a pretrained LaBraM encoder. The paper further claims the framework is backbone-agnostic and parameter-efficient, using per-subject adapters of about 16K parameters, and that the two alignment signals are complementary rather than redundant. The central significance is a step toward non-invasive brain-to-text communication for Chinese, which has a large character space and tonal complexity that prior English-focused methods do not handle.

Core claim

On the ChineseEEG-2 dataset, EEGAlign achieves state-of-the-art closed-set sentence classification: up to 82.37% Top-1 accuracy on Reading Aloud EEG and 41.43% on Passive Listening EEG out of 101 candidate sentences, with chance-ratio accuracies of 83.20× and 41.84× respectively. The paper's key claim is that text-semantic alignment and audio-acoustic alignment provide complementary supervision: combining them yields consistently better performance than either alone, and the full model outperforms baselines (LaBraM-MSE, DeWave, SMM-Challenge) by large margins across two different encoder backbones. It also shows that the semantic branch contributes mainly as an inference-time corrective scor

What carries the argument

The central mechanism is a three-way training objective that combines (1) CTC character-sequence decoding, (2) sentence-level InfoNCE contrastive alignment to frozen BGE-M3 text embeddings, and (3) chunk-level bidirectional InfoNCE alignment to frozen wav2vec 2.0 acoustic features, with a learned upsampler that restores temporal density before the three prediction heads. Subject-specific bottleneck adapters (zero-initialized residual projections, db=32) absorb inter-subject variability, and a three-stage curriculum (semantic warm-start, bootstrapping with CTC, joint refinement) stabilizes the multi-objective optimization. The alignment heads are applied on top of any temporally ordered EEG e

Load-bearing premise

The headline results depend on the within-subject stratified random split (90% train, 10% evaluation, seed 42) and on candidate sampling that draws negative sentences from the full dataset; if the split does not prevent the same sentence text from appearing in both training and evaluation for the same subject, the model could exploit memorized sentence-embedding associations rather than learning a generalizable EEG-to-sentence decoder.

What would settle it

Re-run the evaluation with a sentence-disjoint split, ensuring that no sentence appearing in the training set also appears in the evaluation set for the same subject (i.e., group trials by sentence identity before splitting). If Top-1 accuracy drops substantially (e.g., from 82% to near chance), the current numbers would be inflated by sentence-level leakage rather than reflecting genuine EEG-to-text decoding.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If EEGAlign's complementarity claim holds, future EEG-to-text systems for tonal languages can jointly exploit semantic and acoustic information rather than choosing one supervisory signal, potentially improving decoding robustness in low-signal conditions.
  • The closed-set classification results (up to 82% Top-1 out of 101 candidates) suggest that non-invasive EEG could support practical sentence-selection communication aids for users with severe speech impairments, at least within a constrained candidate vocabulary.
  • The CTC-based decoder avoids the hallucination problem of autoregressive language models, providing a monotonic, character-aligned decoding path that may be more reliable for noisy neural inputs.
  • The framework's backbone-agnostic design and parameter-efficient adapters (≈16K per subject) imply that high-performing EEG decoding can be achieved without large per-subject models, easing deployment to new users.
  • The finding that PL (passive listening) decoding is weaker but still far above chance (41% Top-1) indicates that even without overt articulation, neural signals carry recoverable sentence information, relevant for locked-in users who cannot speak.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The non-monotonic ablation pattern—where adding one alignment axis alone hurts accuracy but adding both helps—suggests that each single axis imposes an incomplete constraint on the encoder, and that the two axes act as mutual regularizers preventing representational collapse; this interaction may generalize to other cross-modal EEG decoding tasks.
  • A natural testable extension would be to evaluate EEGAlign on other tonal or logographic languages (e.g., Japanese kanji or Vietnamese), where the acoustic-semantic complementarity may be even more pronounced than in Chinese.
  • The paper explicitly frames its evaluation as a closed-set candidate-selection task; an open-vocabulary setting would require a different decoding paradigm, and the current CTC head might serve as a building block for beam-search decoding over a language model prior.
  • The strong paradigm-specific collapse of alignment across RA and PL (cross-paradigm retrieval near chance) implies that production and perception brain signals are sufficiently different that unified multi-paradigm decoding would need adaptation mechanisms beyond the current per-subject adapters.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. EEGAlign is a framework for closed-set Chinese sentence identification from scalp EEG, evaluated on the ChineseEEG-2 dataset in both reading-aloud (RA) and passive-listening (PL) conditions. The method combines a CTC character decoder with two contrastive auxiliary supervision signals: alignment of a mean-pooled EEG sentence vector to frozen BGE-M3 text embeddings, and chunk-level alignment of upsampled EEG features to frozen wav2vec 2.0 audio features. Subject-specific bottleneck adapters and a three-stage curriculum are used. The paper reports state-of-the-art Top-1 accuracies of 82.37% on RA and 41.43% on PL out of 101 candidates, with ablations showing that both alignment axes contribute, and includes additional analyses of fusion gains, candidate-set sensitivity, CKA representation geometry, and per-subject results.

Significance. If the reported numbers hold, this is a meaningful step for non-invasive Chinese speech decoding: it is the first system to combine text-semantic and audio-acoustic supervision for sentence-level EEG decoding in both overt production and passive listening, and it demonstrates backbone-agnostic gains on both a lightweight Conformer and a pretrained LaBraM encoder. The paper also makes a useful empirical contribution by showing that the two alignment axes are complementary, and it includes several careful analyses (five-fold cross-validation, candidate-set scaling, CKA) that go beyond a single accuracy number. However, the central claim depends on an evaluation protocol whose checkpoint-selection rule and split unit are not adequately controlled; this must be fixed before the reported performance can be taken at face value.

major comments (3)
  1. [Appendix D / Section 4.1] Best-epoch model selection on the evaluation split. Appendix D states that per-subject Top-1 accuracies are taken 'at the best evaluation epoch for each training configuration' and that per-subject metrics are emitted on every eval1/101 line during training. Since the main experiments use a single 90/10 split (seed=42) with no separate validation set, the headline numbers (82.37% RA, 41.43% PL) are maxima over training epochs on the test split, not the accuracy of a pre-specified checkpoint. This inflates absolute accuracy and makes the comparison with baselines ambiguous unless every baseline uses the identical rule. The five-fold cross-validation in Appendix E does not remove this concern because it does not state whether each fold selects its checkpoint on the evaluation fold. Please re-run with a validation-based epoch selection and report test-set numbers, or clearly show that the r
  2. [Section 4.1, Data Split Protocol] The split unit is unspecified. The authors write 'within-subject stratified random split (90% train, 10% evaluation, seed=42)' but do not say whether the split is at the trial level or at the level of unique sentence texts. The candidate pool C contains 'all unique sentences in the dataset', so if EEG trials of the same sentence text appear in both training and evaluation for the same subject, the model can memorize associations between those EEG trials and the corresponding frozen sentence embedding during training, artificially boosting closed-set retrieval. Please explicitly state the split unit, and ideally report results under a sentence-disjoint split to rule out this memorization path.
  3. [Section 4.4 / Table 1] Baseline selection and tuning are not specified. The text says all methods are evaluated under 'the same closed-set protocol', but the protocol does not define how checkpoints are selected for the three external baselines (LaBraM-MSE, DeWave, SMM-Challenge). If, for example, the baselines use a fixed final-epoch checkpoint while EEGAlign uses the best evaluation-epoch checkpoint, the large margins in Table 1 are not a fair comparison. Please state the selection rule and, ideally, report baselines under the same rule used for EEGAlign.
minor comments (6)
  1. [Section 1 / Abstract] The phrase 'state-of-the-art' is used repeatedly although the comparison set is narrow and evaluation protocols differ across prior work. Consider qualifying it as 'state-of-the-art under the closed-set, within-subject protocol used here.'
  2. [Appendix D] The notation 'eval1/101 line' and the logging format are opaque; explain what '1/101' means or use a less cryptic notation.
  3. [Section 4.2 / Eq. (5)] The z-normalization z_c(·) is applied using 'training-set statistics,' but it is not stated whether these statistics are computed per candidate c over the training set, and whether the same statistics are used for the CTC log-likelihood and the cosine score. Clarify the normalization details.
  4. [Appendix E] The five-fold cross-validation reports only RA and only the LaBraM encoder. The main text says this supports the stability of the PL results, but no equivalent PL cross-validation is presented; please either add it or soften the claim.
  5. [Appendix G, Table 12] In the header 'cospos−cosneg' there are missing spaces and the subscript is unclear; typeset as cos_pos − cos_neg for readability.
  6. [Section 3.3] The audio chunk size is given as chunk_size=20 frames in the text and Table 5, which at 50 Hz is 400 ms. It would be helpful to state explicitly that this was chosen to match the wav2vec 2.0 feature frame rate and whether other chunk sizes were tried.

Circularity Check

1 steps flagged

Headline Top-1 accuracy is selected on the evaluation split (best-epoch checkpoint), so the reported SOTA numbers are test-set-optimized maxima rather than fixed-model held-out predictions.

specific steps
  1. fitted input called prediction [Appendix D, Tables 9–10 caption; Section 4.1 data split protocol]
    "Tables 9–10 and Figure 3 report the per-subject Top-1 accuracy (closed-set, K=100, seed=42) at the best evaluation epoch for each training configuration, extracted directly from the training logs (the per-subject breakdown is emitted alongside every eval1/101 line as [eval][per_subject] sub-X top1=...n=...)."

    The paper uses a single 90/10 within-subject split and reports all main-text accuracies on the held-out evaluation trials, but Appendix D states that the reported per-subject Top-1 values are taken at the best evaluation epoch for each configuration. Thus the evaluation split is used for model/checkpoint selection, and the same split is then used to report final accuracy. The headline 82.37% RA / 41.43% PL is therefore the maximum over training epochs on the evaluation set—a test-set-selected quantity—rather than the accuracy of a fixed pre-specified model on held-out data. This is a fitted selection criterion renamed as a prediction: the reported number is statistically forced upward by construction, and comparisons with baselines are only meaningful if the baselines use the identical bes

full rationale

The core training pipeline is not equation-level circular: the CTC, InfoNCE text, and chunk-level audio losses are standard; the text and audio targets (BGE-M3, wav2vec 2.0) are frozen external features; and candidate scoring at inference is a straightforward combination of cosine similarity and CTC log-likelihood. Ablations and the backbone comparison provide independent content, and citations to LaBraM and ChineseEEG-2 are external data/models rather than self-referential support. The one load-bearing circularity is in the evaluation protocol: Appendix D explicitly selects the checkpoint by best evaluation epoch on the same 10% evaluation split used to report the headline accuracy. This makes the reported SOTA numbers partially circular—they are optimized with respect to the test labels, then presented as the model's predictive accuracy. The five-fold cross-validation in Appendix E does not resolve this unless each fold also avoids best-epoch selection on the evaluation fold, which is not documented. The unspecified trial-vs-sentence split unit is a further leakage risk but is not established by a quoted statement, so I do not count it as a separate circular step.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim is empirical and rests on standard ML components (CTC, InfoNCE, adapters) plus a set of hand-chosen hyperparameters and dataset assumptions. No new physical entity, mediator, or conserved quantity is introduced.

free parameters (6)
  • Loss weights λ_ctc, λ_text, λ_audio = λ_audio=0.1 (default), λ_audio=0.5 swept; λ_ctc and λ_text not explicitly stated in text
    Equation (4) weights three objectives; only λ_audio is reported and swept, so the relative objective weighting is partly unspecified and appears tuned on the same data.
  • Fusion weight α = 0.4
    Equation (5) uses α=0.4 by default; no sensitivity analysis for α is reported.
  • InfoNCE temperatures τ and τ_a = 0.05 (text), 0.07 (audio)
    Chosen by hand in Section 3.3; they control the sharpness of the contrastive losses and affect the reported retrieval performance.
  • Audio chunk size = 20 frames = 400 ms
    Window length for the chunk-level acoustic alignment, chosen in Section 3.3; it defines the temporal granularity of audio supervision.
  • Candidate-set size K = 100 default, swept over 5, 20, 50, 100, 500
    The closed-set protocol parameter; Figure 2(b) shows Top-1 accuracy depends strongly on K, so headline numbers are protocol-bound.
  • Data split seed and train ratio = seed=42, 90/10
    Main results use a fixed within-subject split; five-fold CV is given only for RA-LaBraM, and PL has no cross-validation.
axioms (4)
  • standard math CTC assumes monotonic frame-to-character alignment between EEG and text tokens.
    Section 3.1 introduces CTC negative log-likelihood as the output decoder; if EEG-to-text alignment is not monotonic, the CTC path cannot recover the sequence.
  • domain assumption Scalp EEG contains decodable sentence-level semantic and acoustic information at the granularity assumed by the alignment targets.
    Section 3.2 states that text alignment provides sentence-level discriminability and audio alignment provides temporal structure; if this premise is false, the dual-alignment recipe cannot work.
  • domain assumption ChineseEEG-2 labels, segmentation, and the within-subject 90/10 split are valid and do not leak sentence identity between training and evaluation.
    Section 4.1 defines the data and split protocol; the headline accuracies depend on the split being trial-independent or otherwise leak-free.
  • domain assumption Frozen BGE-M3 text embeddings and wav2vec 2.0 speech features are appropriate, stable external anchors for EEG alignment.
    Section 3.3 uses these external features as fixed targets; if they mismatch the information actually present in EEG, the alignment losses would fit noise.

pith-pipeline@v1.3.0-alltime-deepseek · 17765 in / 12533 out tokens · 140038 ms · 2026-08-01T01:49:48.209742+00:00 · methodology

0 comments
read the original abstract

Decoding speech information directly from scalp electroencephalography (EEG) into text provides a potential non-invasive neural communication pathway for individuals with severe speech and motor impairments. Compared with invasive approaches such as electrocorticography, EEG is safer and more widely deployable, yet substantially more challenging to decode.This challenge is exacerbated for Chinese sentence decoding, which must handle a high-dimensional output space with thousands of characters, severe inter-subject variability, and low signal-to-noise ratios for text alignment.Existing methods commit to a single supervisory axis---either text semantics or audio acoustic features---yet neither can simultaneously satisfy the demands of sentence-level discriminability and fine-grained temporal resolution required for large-vocabulary Chinese decoding. We introduce EEGAlign, a novel parameter-efficient framework that jointly aligns EEG with two axes---text alignment with BGE-M3 text embeddings and audio alignment with wav2vec~2.0 speech features via contrastive learning followed by CTC character-sequence decoding. On ChineseEEG-2 data, EEGAlign yields state-of-the-art closed-set sentence classification performance, reaching up to 82.37% Top-1 accuracy on Reading Aloud EEG and 41.43% on Passive Listening EEG out of 101 candidates. Ablation studies show that the two alignment axes are highly complementary: combining them yields consistently better performance than either alone. To the best of our knowledge, this is the first study on decoding large-vocabulary Chinese sentences from non-invasive EEG during overt speech production, and achieving strong classification performance with relatively large closed-set candidate-sentence setting.

Figures

Figures reproduced from arXiv: 2607.25626 by Feng Tian, Tian Zheng, Xiaolan Peng, Xinxin Zhu, Xurong Xie.

Figure 1
Figure 1. Figure 1: Overall architecture of EEGAlign. A shared sequence encoder with lightweight subject-specific adapters maps EEG to latent representations. A learned upsampler restores temporal density before applying three joint heads: CTC for decoding, text semantic alignment to BGE-M3 text embeddings, and audio acoustic alignment to wav2vec 2.0 features. Lctc = − log Pctc(y | gctc(H)). (1) Although this decoder avoids a… view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of fusion gain, hyperparameter sensitivity, and qualitative case studies. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-subject Top-1 accuracy. (a) RA (4 subjects). (b) PL (8 subjects). Dashed lines mark the across-subject mean for EEGAlign-LaBraM [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-paradigm transfer analysis. (a) 200-way in-batch Top-1 using the text head only, for every (training paradigm, evaluation paradigm) pair. Diagonal cells are the values reported in [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Encoder-feature text-alignment analysis for RA and PL paradigms. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: CKA alignment analysis across all four ablation conditions. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 7 linked inside Pith

  1. [1]

    A high-performance speech neuroprosthesis.Nature, 620(7976):1031–1036, 2023

    Francis R Willett, Erin M Kunz, Chaofei Fan, Donald T Avansino, Guy H Wilson, Eun Young Choi, Foram Kamdar, Matthew F Glasser, Leigh R Hochberg, Shaul Druckmann, et al. A high-performance speech neuroprosthesis.Nature, 620(7976):1031–1036, 2023

    2023

  2. [2]

    Transfer learning in brain-computer interfaces.IEEE Computational Intelligence Magazine, 11(1):20–31, 2016

    Vinay Jayaram, Morteza Alamgir, Yasemin Altun, Bernhard Scholkopf, and Moritz Grosse- Wentrup. Transfer learning in brain-computer interfaces.IEEE Computational Intelligence Magazine, 11(1):20–31, 2016

    2016

  3. [3]

    Reducing the subject variability of eeg signals with adversarial domain generalization

    Bo-Qun Ma, He Li, Wei-Long Zheng, and Bao-Liang Lu. Reducing the subject variability of eeg signals with adversarial domain generalization. InInternational Conference on Neural Information Processing, pages 30–42. Springer, 2019

    2019

  4. [4]

    Dewave: Discrete encoding of eeg waves for eeg to text translation.Advances in Neural Information Processing Systems, 36:9907–9918, 2023

    Yiqun Duan, Jinzhao Zhou, Zhen Wang, Yu-Kai Wang, and Chin-teng Lin. Dewave: Discrete encoding of eeg waves for eeg to text translation.Advances in Neural Information Processing Systems, 36:9907–9918, 2023

    2023

  5. [5]

    Enhancing eeg-to-text decoding through transferable representations from pre-trained contrastive eeg- text masked autoencoder

    Jiaqi Wang, Zhenxi Song, Zhengyu Ma, Xipeng Qiu, Min Zhang, and Zhiguo Zhang. Enhancing eeg-to-text decoding through transferable representations from pre-trained contrastive eeg- text masked autoencoder. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7278–7292, 2024

    2024

  6. [6]

    Eeg2text: Open vocabulary eeg-to-text decoding with eeg pre-training and multi-view transformer.arXiv preprint arXiv:2405.02165, 2024

    Hanwen Liu, Daniel Hajialigol, Benny Antony, Aiguo Han, and Xuan Wang. Eeg2text: Open vocabulary eeg-to-text decoding with eeg pre-training and multi-view transformer.arXiv preprint arXiv:2405.02165, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Neurolm: A universal multi-task foundation model for bridging the gap between language and eeg signals.arXiv preprint arXiv:2409.00101, 2024

    Wei-Bang Jiang, Yansen Wang, Bao-Liang Lu, and Dongsheng Li. Neurolm: A universal multi-task foundation model for bridging the gap between language and eeg signals.arXiv preprint arXiv:2409.00101, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Guiding llms to decode text via aligning semantics in eeg signals and language.Expert Systems with Applications, page 130300, 2025

    Huanran Zheng, Yuanbin Wu, Tianwen Qian, Wenjing Yue, and Xiaoling Wang. Guiding llms to decode text via aligning semantics in eeg signals and language.Expert Systems with Applications, page 130300, 2025

    2025

  9. [9]

    Braindec: A multimodal llm for the non-invasive decoding of text from brain recordings

    Youssef Hmamouche, Ismail Chihab, Lahoucine Kdouri, and Amal El Fallah Seghrouchni. Braindec: A multimodal llm for the non-invasive decoding of text from brain recordings. INFORMATION FUSION, 127, 2026

    2026

  10. [10]

    Ssm2mel: State space model to reconstruct mel spectrogram from the eeg

    Cunhang Fan, Sheng Zhang, Jingjing Zhang, Zexu Pan, and Zhao Lv. Ssm2mel: State space model to reconstruct mel spectrogram from the eeg. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025

    2025

  11. [11]

    Dmf2mel: A dynamic multiscale fusion network for eeg-driven mel spectrogram reconstruction

    Cunhang Fan, Sheng Zhang, Jingjing Zhang, Enrui Liu, Xinhui Li, Gangming Zhao, and Zhao Lv. Dmf2mel: A dynamic multiscale fusion network for eeg-driven mel spectrogram reconstruction. InProceedings of the 33rd ACM International Conference on Multimedia, pages 6977–6985, 2025

    2025

  12. [12]

    Convconcatnet: a deep convolutional neural network to reconstruct mel spectrogram from the eeg

    Xiran Xu, Bo Wang, Yujie Yan, Haolin Zhu, Zechen Zhang, Xihong Wu, and Jing Chen. Convconcatnet: a deep convolutional neural network to reconstruct mel spectrogram from the eeg. In2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pages 113–114. IEEE, 2024

    2024

  13. [13]

    Biot: Biosignal transformer for cross-data learning in the wild.Advances in Neural Information Processing Systems, 36:78240–78260, 2023

    Chaoqi Yang, M Westover, and Jimeng Sun. Biot: Biosignal transformer for cross-data learning in the wild.Advances in Neural Information Processing Systems, 36:78240–78260, 2023

    2023

  14. [14]

    Large brain model for learning generic representations with tremendous eeg data in bci.arXiv preprint arXiv:2405.18765, 2024

    Wei-Bang Jiang, Li-Ming Zhao, and Bao-Liang Lu. Large brain model for learning generic representations with tremendous eeg data in bci.arXiv preprint arXiv:2405.18765, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Zuco, a simultaneous eeg and eye-tracking resource for natural sentence reading

    Nora Hollenstein, Jonathan Rotsztejn, Marius Troendle, Andreas Pedroni, Ce Zhang, and Nicolas Langer. Zuco, a simultaneous eeg and eye-tracking resource for natural sentence reading. Scientific data, 5(1):180291, 2018. 11

    2018

  16. [16]

    Effects of eeg analysis window location on classifying spoken mandarin monosyllables

    Mingtao Li, Shangdi Liao, Sio Hang Pun, and Fei Chen. Effects of eeg analysis window location on classifying spoken mandarin monosyllables. In2023 11th International IEEE/EMBS Conference on Neural Engineering (NER), pages 1–4. IEEE, 2023

    2023

  17. [17]

    3m-cpseed, an eeg-based dataset for chinese pinyin production in overt, mouthed, and imagined speech.Scientific Data, 2025

    Xinyu Ma, Yi Jiang, and Ning Jiang. 3m-cpseed, an eeg-based dataset for chinese pinyin production in overt, mouthed, and imagined speech.Scientific Data, 2025

    2025

  18. [18]

    Scaling law in neural data: Non-invasive speech decoding with 175 hours of eeg data, 2024

    Motoshige Sato, Kenichi Tomeoka, Ilya Horiguchi, Kai Arulkumaran, Ryota Kanai, and Shuntaro Sasai. Scaling law in neural data: Non-invasive speech decoding with 175 hours of eeg data, 2024

    2024

  19. [19]

    An eeg dataset for multimodal semantic alignment and neural decoding during reading and listening.Scientific Data, 2025

    Sitong Chen, Beiqianyi Li, Cuilin He, Dongyang Li, Mingyang Wu, Xinke Shen, Song Wang, Xuetao Wei, Xindi Wang, Haiyan Wu, et al. An eeg dataset for multimodal semantic alignment and neural decoding during reading and listening.Scientific Data, 2025

    2025

  20. [20]

    Decoding speech perception from non-invasive brain recordings.Nature Machine Intelligence, 5 (10):1097–1107, 2023

    Alexandre Défossez, Charlotte Caucheteux, Jérémy Rapin, Ori Kabeli, and Jean-Rémi King. Decoding speech perception from non-invasive brain recordings.Nature Machine Intelligence, 5 (10):1097–1107, 2023

    2023

  21. [21]

    Eeg-clip: A transformer-based framework for eeg-guided image generation.Neural Networks, page 108167, 2025

    Xuhao Cao, Peiliang Gong, Liying Zhang, and Daoqiang Zhang. Eeg-clip: A transformer-based framework for eeg-guided image generation.Neural Networks, page 108167, 2025

    2025

  22. [22]

    Self-supervised speech representation and contextual text embedding for match-mismatch classification with eeg recording

    Bo Wang, Xiran Xu, Zechen Zhang, Haolin Zhu, YuJie Yan, Xihong Wu, and Jing Chen. Self-supervised speech representation and contextual text embedding for match-mismatch classification with eeg recording. In2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pages 111–112. IEEE, 2024

    2024

  23. [23]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  24. [24]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    2022

  25. [25]

    Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks

    Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. InProceedings of the 23rd international conference on Machine learning, pages 369–376, 2006

    2006

  26. [26]

    Bge m3- embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation.arXiv preprint arXiv:2402.03216, 4(5), 2024

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3- embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation.arXiv preprint arXiv:2402.03216, 4(5), 2024

    Pith/arXiv arXiv 2024

  27. [27]

    wav2vec 2.0: A framework for self-supervised learning of speech representations.Advances in neural information processing systems, 33:12449–12460, 2020

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations.Advances in neural information processing systems, 33:12449–12460, 2020

    2020

  28. [28]

    Conformer: Convolution-augmented transformer for speech recognition.arXiv preprint arXiv:2005.08100, 2020

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition.arXiv preprint arXiv:2005.08100, 2020

    Pith/arXiv arXiv 2005

  29. [29]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. InInternational conference on machine learning, pages 2790–2799. PMLR, 2019

    2019

  30. [30]

    Chineseeeg: A chinese linguistic corpora eeg dataset for semantic alignment and neural decoding.Scientific Data, 11(1):550, 2024

    Xinyu Mou, Cuilin He, Liwei Tan, Junjie Yu, Huadong Liang, Jianyu Zhang, Yan Tian, Yu-Fang Yang, Ting Xu, Qing Wang, et al. Chineseeeg: A chinese linguistic corpora eeg dataset for semantic alignment and neural decoding.Scientific Data, 11(1):550, 2024

    2024

  31. [31]

    Similarity of neural network representations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. InInternational conference on machine learning, pages 3519–3529. PMlR, 2019

    2019

  32. [32]

    RA encoder vs. PL encoder

    Jacky Tai-Yu Lu, Jung Chiang, Chi-Sheng Chen, Anna Nai-Yun Tung, Hsiang Wei Hu, and Yuan Chiao Cheng. Eeg2text-cn: An exploratory study of open-vocabulary chinese text-eeg alignment via large language model and contrastive learning on chineseeeg.arXiv preprint arXiv:2506.00854, 2025. 12 A Comparison with representative neural decoding meth- ods Table 3 po...

    Pith/arXiv arXiv 2025