Pith. sign in

REVIEW 5 major objections 6 minor 55 references

Bridging Brain Signals and Language: A Deep Learning Approach to EEG-to-Text Decoding

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a two-stage EEG-to-text decoding framework that learns subject-specific representations from raw EEG, generates sentences with BART, and refines them with GPT-4, aiming for open-vocabulary brain-to-text communication.

desk verdict A candid replication report that undercuts its own state-of-the-art claim: every comparable metric is below the baseline it set out to beat. read the letter →

arxiv 2502.17465 v1 pith:TTFQVM2O submitted 2025-02-11 eess.SP cs.CLcs.LG

Mostafa El Gedawy , Omnia Nabil , Omar Mamdouh , Mahmoud Nady , Nour Alhuda Adel , Ahmed Fares This is my paper · ORCID
classification eess.SPcs.CLcs.LG
keywords brain-to-textsystemsEEGdecodingEEG-to-textrepresentationlearningassistivetechnologiesbrain-computerinterfaceopen-vocabularypersonalizedbrain
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that EEG recordings taken while a person reads English sentences can be decoded into open-vocabulary text, not just a fixed set of words. It proposes a three-part pipeline: a brain module that learns subject-specific representations from raw EEG, a BART language model that turns those representations into sentences, and a GPT-4 refinement step that improves fluency and semantic fidelity. The authors test the framework on the ZuCo dataset, which pairs EEG and eye-tracking with sentence reading, and report BLEU, ROUGE, and BERTScore numbers. Their abstract claims these numbers exceed current methods; the tables they show put the BART variant slightly below the quoted comparison baseline on most metrics. The intended contribution is a template for personalized, open-vocabulary brain-to-text decoding that could eventually support assistive communication.

What carries the argument

The load-bearing object is the subject-specific Learnable Features Module followed by a frozen-then-fine-tuned BART decoder. The module stacks a bidirectional GRU, a linear projection, pointwise convolution, a learned subject row vector, a transformer encoder, and a residual MLP to turn raw word-level EEG into a latent representation $Z$. Stage-one training ties $Z$ to BART's token embeddings by MSE; stage-two training fine-tunes BART as a sequence generator from $Z$; GPT-4 then refines the sentence. This two-stage alignment is what is supposed to carry open-vocabulary decoding.

What would settle it

A decisive check is to rerun the pipeline under the exact split and preprocessing of [33] and recompute BLEU-1 and BERTScore-F with the official scripts; if BART then scores at or below 42.75 BLEU-1 and 53.86 BERTScore-F, the abstract's claim of higher performance is not supported.

Watch

Extended reading notes

Core claim

The central claim is that aligning EEG-derived word-level features to the token embedding space of a frozen BART encoder, then fine-tuning BART to decode sentences, lets a model generate words outside any closed vocabulary while adapting to individual subjects through a learned per-subject vector. The paper presents this as a two-stage training procedure: first minimize MSE between the brain module's output and BART token embeddings; then fine-tune BART with cross-entropy on the target sentences. At inference, GPT-4 revises the BART output. On ZuCo, the authors report BLEU-1 of 42.34, ROUGE-1-F of 32.66, and BERTScore-F of 53.53 for BART, with BART+GPT-4 reaching BERTScore-F of 54.91. They describe these results as close to, and in the abstract as higher than, the comparison system; the tabulated comparison shows slightly lower numbers on most metrics, and the authors themselves list limitations including only 30 subjects, English-only reading, the need for accurate temporal alignment, costly computation, and extensive preprocessing.

Load-bearing premise

The load-bearing premise is that the evaluation protocol matches the cited baseline, meaning the same training/validation/test split, no subject overlap between training and test, and the same word-level EEG segmentation, yet the paper does not state these details.

Editorial extensions

If this is right

  • If the representation alignment works, EEG-to-text systems can generate any word in the language rather than choosing from a preset list, which is the difference between laboratory spellers and natural communication.
  • The per-subject learnable vector gives a concrete route to personalization: the same architecture can adapt to a new user by learning one vector, without redesigning the decoder.
  • The modular split between brain and language modules means improvements in pretrained language models can be dropped in without retraining the EEG encoder from scratch.
  • The reported numbers imply that GPT-4 refinement trades a little exact n-gram overlap for higher semantic similarity, which matters when the goal is meaning rather than verbatim transcription.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper leaves implicit is an ablation that removes the subject-specific vector; if per-user adaptation is the true source of the gains, decoding accuracy should drop measurably when a single shared vector replaces it.
  • The close scores between this pipeline and the cited baseline suggest that much of the output may be carried by BART's language prior rather than by EEG-specific signal; permuting the EEG-to-sentence pairing during training would reveal how much neural information the brain module actually encodes.
  • Because the examples show named entities sometimes decoded correctly and sometimes confused, a focused evaluation on rare proper nouns would map where the open-vocabulary claim breaks down.
  • A fair reading of the paper's own tables is that the contribution is a near-reproduction of the comparison architecture across additional language models, T5 and ProphetNet, rather than a demonstrated improvement over it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper presents a two-stage deep learning framework for EEG-to-text decoding on the ZuCo dataset. A brain module (Bi-GRU, linear projection, pointwise convolution, subject-specific layer, transformer encoder, residual MLP) maps word-level EEG to features aligned to BART token embeddings via MSE loss; BART is then fine-tuned with cross-entropy to generate sentences, and GPT-4 optionally refines the outputs. The authors report BLEU, ROUGE, and BERTScore for BART, BART+GPT-4, T5, and ProphetNet, and claim the framework achieves higher performance than current methods. They also describe a web platform and text-to-speech delivery. The central claim of state-of-the-art performance is contradicted by the paper's own comparison tables: every overlapping metric is lower than the cited baseline from Amrani et al. [33].

Significance. If the reported performance were supported, the work would be an incremental but useful contribution to open-vocabulary EEG-to-text decoding: it demonstrates a subject-specific feature extractor, two-stage training with a pretrained language model, and a publicly described web platform. The two-stage training objective is a legitimate modeling choice, and the evaluation compares generated text to ground truth rather than reusing test labels, so there is no basic circularity in the main numbers. However, the reported results do not support the headline claim of superiority, and the evaluation is too incomplete to establish even competitive performance: only one external baseline is used, no error bars or statistical tests are provided, no per-subject results appear, the data split is unspecified, and the claimed ablation study is absent. The detailed web-platform appendix and the qualitative decoding examples in Table VI are the strongest positive elements.

major comments (5)
  1. [Abstract; Section IV, Tables IV and V] The abstract claims 'higher BLEU, ROUGE and BERTScore performance when compared to current methods,' but the only external baseline in the paper, Amrani et al. [33], outperforms the proposed system on every overlapping metric. For BART, BLEU-1 is 42.34 vs 42.75, BLEU-4 is 8.88 vs 9.56, ROUGE-1-F is 32.66 vs 33.28, and BERTScore-F is 53.53 vs 53.86. For BART+GPT-4, BLEU-1 is 40.72 vs 40.87, ROUGE-1-F is 32.10 vs 32.61, and BERTScore-F is 54.91 vs 55.34. The claim of superiority is therefore directly contradicted by the paper's own evidence, and no other state-of-the-art comparison is provided.
  2. [Section III-A and Section IV] The evaluation protocol is not specified. The paper never states how the 30 subjects and the sentences were divided into training, validation, and test sets, whether subjects overlap across splits, or which sentences were held out. Because the brain module includes a subject-specific row vector r_s (Section III-G4), same-subject overlap between training and test would allow the model to memorize subjects and inflate scores. Without a described split, the numbers in Table IV cannot be properly interpreted, and the comparison to the quoted numbers from [33] is meaningful only if the same split and preprocessing are used.
  3. [Section IV and Section VI] No error bars, confidence intervals, statistical tests, or per-subject results are reported. Differences such as BLEU-1 42.34 vs 42.81 between BART and T5 are small and could easily arise from random seed or fold choices. The statement in Section IV that 'BART model is almost the best' and the conclusion's claim that BART is 'the preeminent model for this task' are therefore unsupported.
  4. [Section VI (Conclusion)] The conclusion lists 'an ablation study that examines each module's contributions' as a contribution of the paper, but no ablation experiment appears anywhere in Section IV or elsewhere in the manuscript. This is an unsupported claim and should either be carried out or removed.
  5. [Section III-G4 and Table IV] The benefit of the subject-specific layer is not demonstrated. The paper never compares results with and without the subject-specific row vector, nor does it report per-subject performance, so the central design choice for addressing inter-subject variability is unvalidated. This is particularly important because the authors state that their architecture follows the baseline [33].
minor comments (6)
  1. [Section V, Eq. (13)] Equation (13) is labeled 'Recall' but contains a harmonic-mean expression for the F1 score and is missing the factor 2; the F1 score should be 2PR/(P+R), not 1/(1/P + 1/R).
  2. [Section IV and Table IV] The text reports a ROUGE-2-F score of 9.60% and a BLEU-4 score of 8.89%, but Table IV contains only ROUGE-1 and reports BLEU-4 as 8.88; these inconsistencies should be reconciled.
  3. [Throughout] The manuscript contains numerous typographical and grammatical errors that impede readability, including 'mortal reading', 'pickel', 'architure', and 'Table IV and Table IV' in Section V; a thorough language edit is needed.
  4. [References] Reference [30] is dated 2029 but is the 2019 Nature paper by Anumanchipalli et al.; the bibliography needs correction.
  5. [Section IV] The paper does not report training hyperparameters or random seeds (e.g., number of layers L, attention heads H, hidden dimension d_h, learning rate, batch size), and there is no code availability statement, which limits reproducibility.
  6. [Appendix A] Appendix A states that passwords are stored in plaintext; the authors acknowledge the need for hashing, but this should be fixed before any deployment claim is made.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the EEG-to-text pipeline is trained and evaluated against ground-truth text, and the comparison baseline is external rather than self-cited.

full rationale

The claimed derivation chain is a supervised two-stage training pipeline: a Brain module is trained to regress word-level EEG features onto BART token embeddings (Eq. 7), and then BART is fine-tuned with cross-entropy against ground-truth sentences (Eq. 8). The reported BLEU/ROUGE/BERTScore numbers compare generated text to reference sentences, so the evaluation metric is not an input to the model. The architecture is explicitly stated to follow the external state-of-the-art paper by Hamza Amrani et al. [33], and no load-bearing premise is justified by a self-citation; the only cited prior implementation is by different authors. Although the abstract's claim of 'higher' performance is not supported by the paper's own Tables IV and V (every overlapping metric is lower than [33]), that is an internal-consistency/correctness problem, not circularity. The paper does not define its predictions in terms of its inputs or fit a parameter and then rename it as a prediction. Accordingly, no circular step meeting the quoted-evidence standard is present.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The central empirical claims depend on two unstated premises: (1) that the data are split and preprocessed identically to Amrani et al. [33] so that quoted baseline numbers are comparable, and (2) that word-level EEG segments, windowed by gaze duration, contain enough signal to align with BART token embeddings. The paper's own text provides no hyperparameters, no per-subject test results, and no code, so these premises cannot be checked.

free parameters (2)
  • subject-specific row vector r_s = learned per subject, not reported
    The Subject-Specific Layer learns a vector per subject to account for inter-subject variability; its values and regularization are unspecified, and the paper does not report per-subject performance.
  • model hyperparameters (layers L, heads H, hidden dim d_h, learning rates, batch sizes) = not reported
    The Brain Transformer Encoder and training procedures depend on these hand-chosen values, but they are not listed anywhere in the paper.
assumptions (2)
  • domain assumption Word-level EEG segments windowed by gaze duration contain sufficient signal to align with BART token embeddings
    Section III-A describes segmenting EEG based on eye fixation periods; the entire training signal for the brain module depends on this alignment being learnable.
  • domain assumption The test data and evaluation protocol are equivalent to Hamza Amrani et al. [33], so quoted comparison numbers are valid baselines
    Section IV compares against values quoted from [33] without re-running the baseline under the same splits, seeds, or preprocessing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Brain Signals and Language: A Deep Learning Approach to EEG-to-Text Decoding." pith.science (2026). https://pith.science/paper/TTFQVM2O

@misc{pith2026250217465,
  author       = {Pith},
  title        = {Pith review of: Bridging Brain Signals and Language: A Deep Learning Approach to EEG-to-Text Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TTFQVM2O}},
  note         = {Machine review of arXiv:2502.17465}
}
read the original abstract

Brain activity translation into human language delivers the capability to revolutionize machine-human interaction while providing communication support to people with speech disability. Electronic decoding reaches a certain level of achievement yet current EEG-to-text decoding methods fail to reach open vocabularies and depth of meaning and individual brain-specific variables. We introduce a special framework which changes conventional closed-vocabulary EEG-to-text decoding approaches by integrating subject-specific learning models with natural language processing methods to resolve detection obstacles. This method applies a deep representation learning approach to extract important EEG features which allow training of neural networks to create elaborate sentences that extend beyond original data content. The ZuCo dataset analysis demonstrates that research findings achieve higher BLEU, ROUGE and BERTScore performance when compared to current methods. The research proves how this framework functions as an effective approach to generate meaningful and correct texts while understanding individual brain variations. The proposed research aims to create a connection between open-vocabulary Text generation systems and human brain signal interpretation for developing efficacious brain-to-text systems. The research produces interdisciplinary effects through innovative assistive technology development and personalized communication systems which extend possibilities for human-computer interaction in various settings.

Figures

Figures reproduced from arXiv: 2502.17465 by the authors.

Figure 1
Figure 1. Input: The subject reads sentences while EEG signals are recorded. Brain Module: The EEG signals corresponding to individual words are processed to extract subject-dependent features. Language Module: The features are input to a BART-based language model, which generates a preliminary sentence. Refinement Module: A GPT-4-based model refines the preliminary sentence to improve its fluency and accuracy. Output: The fi… view at source ↗
Figure 2
Figure 2. shows recording reference was at Cz. it is located at the central electrode location on the scalp. This central placement minimizes signal biases and ensures balanced referencing across scalp regions. Other electrodes measure signals relative to Cz, an electrode near your forehead might measure +50 µV another electrode at the back of your head might measure -30 µV these values are relative to Cz, which is 0. Referri… view at source ↗
Figure 3
Figure 3. shows the content of each dataset file (its extension is pickle). Each subject represents the key in the big dictionary ‘. . . .’ (dataset) and the value is list ‘[. . . .]’ of dictionaries, each dictionary in the list contains a sentence and its EEG signals and EEG signals during (gaze duration (GD), total reading time (TRT), first fixation duration (FFD) ). Data was collected from a total of 30 subjects - 12 subje… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustrates a comprehensive view of the proposed system for translating EEG signals into open vocabulary text. Initially, raw EEG signals at the word level are input into the Brain module, which is responsible for extracting deep and meaningful representations tailored…
Figure 5
Figure 5. Figure 5: The Learnable features module consists of (1) Bi-Directional Gated Recurrent Unit (GRU), (2) Linear Projection Layer, (3) 1D Pointwise Convolution Transformer Encoder), (4) Subject-Specific Layer , (5) Brain Transformer Encoder (BTE) and (6) Residual Multi-Layer Percep…
Figure 6
Figure 6. Figure 6: BLEU-1,2,3,4 Score Comparison [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: ROUGE-1 Score Comparison [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: BERTScore Score Comparison [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Signin and Signup pages [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Submit page that used to upload the pickel file [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Predicted and Refined string page and contain the audio files [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 43 canonical work pages

  1. [33]

    Deep Representation Learning for Open V ocabulary Electroencephalography-to-Text Decoding,

    H. Amrani, D. Micucci, and P. Napoletano, “Deep Representation Learning for Open V ocabulary Electroencephalography-to-Text Decoding,” IEEE Journal of Biomedical and Health Informatics , 2024

    work page 2024

  2. [1]

    Aphasia Statistics - The National Aphasia Association,

    Unknown Author, “Aphasia Statistics - The National Aphasia Association,” Mar. 2022. [Online]. Available: https://aphasia.org/ aphasia-resources/aphasia-statistics/

    work page 2022

  3. [2]

    Prevalence of voice disorders in the general population, based on the Stockholm Public Health cohort,

    V . Lyberg- ˚Ahlander, R. Rydell, P. Fredlund, C. Magnusson, and S. Wil´en, “Prevalence of voice disorders in the general population, based on the Stockholm Public Health cohort,” Journal of Voice, vol. 33, no. 6, pp. 900–905, Aug. 2018. [Online]. Available: https: //doi.org/10.1016/j.jvoice.2018.07.007

    work page doi:10.1016/j.jvoice.2018.07.007 2018

  4. [3]

    H. Le, F. Lui, and M. Y . Lui, Aphasia. StatPearls Publishing,

    work page

  5. [4]

    Public Figures and Celebrities with Aphasia - The National Aphasia Association,

    Unknown Author, “Public Figures and Celebrities with Aphasia - The National Aphasia Association,” Mar. 2022. [Online]. Available: https://aphasia.org/aphasia-resources/ public-figures-and-celebrities-with-aphasia/

    work page 2022

  6. [5]

    Brain signals converted into words ‘speak’ for person with paralysis,

    K. S. Servick, “Brain signals converted into words ‘speak’ for person with paralysis,” Jul. 2021. [On- line]. Available: https://www.science.org/content/article/ brain-signals-converted-words-speak-person-paralysis

    work page 2021

  7. [6]

    Neuroprosthesis for decoding speech in a paralyzed person with anarthria,

    D. A. Moses, S. L. Metzger, J. R. Liu, G. K. Anumanchipalli, J. G. Makin, P. F. Sun, J. Chartier, M. E. Dougherty, P. M. Liu, G. M. Abrams et al., “Neuroprosthesis for decoding speech in a paralyzed person with anarthria,” New England Journal of Medicine , vol. 385, no. 3, pp. 217–227, 2021

    work page 2021

  8. [7]

    A high-performance speech neuroprosthesis,

    F. R. Willett, E. M. Kunz, C. Fan, D. T. Avansino, G. H. Wilson, E. Y . Choi, F. Kamdar, L. R. Hochberg, S. Druckmann, K. V . Shenoy et al., “A high-performance speech neuroprosthesis,” bioRxiv, 2023

    work page 2023

Show all 55 references
  1. [8]

    ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading,

    N. Hollenstein, J. Rotsztejn, M. Troendle, A. Pedroni, C. Zhang, and N. Langer, “ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading,” Scientific Data, vol. 5, no. 1, pp. 1–13, 2018

    2018

  2. [9]

    Decoding covert speech from EEG—A comprehensive review,

    J. T. Panachakel and A. G. Ramakrishnan, “Decoding covert speech from EEG—A comprehensive review,” Frontiers in Neu- roscience, vol. 15, p. 392, 2021

    2021

  3. [10]

    De- Wave: Discrete EEG waves encoding for brain dynamics to text translation,

    Y . Duan, J. Zhou, Z. Wang, Y .-K. Wang, and C.-T. Lin, “De- Wave: Discrete EEG waves encoding for brain dynamics to text translation,” arXiv preprint arXiv:2309.14030 , 2023

    2023 arXiv

  4. [11]

    Semantic-aware contrastive learn- ing for electroencephalography-to-text generation with curriculum learning,

    X. Feng, X. Feng, and B. Qin, “Semantic-aware contrastive learn- ing for electroencephalography-to-text generation with curriculum learning,” arXiv preprint arXiv:2301.09237 , 2023

    2023 arXiv

  5. [12]

    Open vocabulary electroencephalography- to-text decoding and zero-shot sentiment classification,

    Z. Wang and H. Ji, “Open vocabulary electroencephalography- to-text decoding and zero-shot sentiment classification,” in Pro- ceedings of the AAAI Conference on Artificial Intelligence , vol. 36, 2022, pp. 5350–5358

    2022

  6. [13]

    High performance communication by people with paralysis using an intracortical brain-computer interface,

    C. Pandarinath, P. Nuyujukian, C. H. Blabe, B. L. Sorice, J. Saab, F. R. Willett, L. R. Hochberg, K. V . Shenoy, and J. M. Henderson, “High performance communication by people with paralysis using an intracortical brain-computer interface,” eLife, vol. 6, p. e18554, 2017

    2017

  7. [14]

    Low- dimensional subject representation-based transfer learning in EEG decoding,

    P.-Y . Jeng, C.-S. Wei, T.-P. Jung, and L.-C. Wang, “Low- dimensional subject representation-based transfer learning in EEG decoding,” IEEE Journal of Biomedical and Health Informatics , vol. 25, no. 6, pp. 1915–1925, 2020

    1915

  8. [15]

    BERT: Pre-training of deep bidirectional transformers for language un- derstanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language un- derstanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  9. [16]

    BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension,

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer, “BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension,”arXiv preprint arXiv:1910.13461, 2019

    1910 arXiv

  10. [17]

    Brains and algorithms partially converge in natural language processing,

    C. Caucheteux and J.-R. King, “Brains and algorithms partially converge in natural language processing,” Communications Biol- ogy, vol. 5, no. 1, p. 134, 2022

    2022

  11. [18]

    Semantic reconstruc- tion of continuous language from non-invasive brain recordings,

    J. Tang, A. LeBel, S. Jain, and A. G. Huth, “Semantic reconstruc- tion of continuous language from non-invasive brain recordings,” Nature Neuroscience, pp. 1–9, 2023

    2023

  12. [20]

    Does the brain represent words? An evaluation of brain decoding studies of language understanding,

    J. Gauthier and A. Ivanova, “Does the brain represent words? An evaluation of brain decoding studies of language understanding,” arXiv preprint arXiv:1806.00591 , 2018

    2018 arXiv

  13. [21]

    Thinking out loud, an open-access EEG-based BCI dataset for inner speech recognition,

    N. Nieto, V . Peterson, H. L. Rufiner, J. E. Kamienkowski, and R. Spies, “Thinking out loud, an open-access EEG-based BCI dataset for inner speech recognition,” Scientific Data, vol. 9, no. 1, p. 52, 2022

    2022

  14. [22]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , 2017, pp. 5998–6008

    2017

  15. [23]

    ROUGE: A package for automatic evaluation of summaries,

    C.-Y . Lin, “ROUGE: A package for automatic evaluation of summaries,” in Text Summarization Branches Out , 2004, pp. 74–81

    2004

  16. [24]

    BERTScore: Evaluating text generation with BERT,

    T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi, “BERTScore: Evaluating text generation with BERT,” arXiv preprint arXiv:1904.09675, 2019

    1904 arXiv

  17. [25]

    On the properties of neural machine translation: Encoder-decoder approaches,

    K. Cho, B. Van Merrienboer, D. Bahdanau, and Y . Bengio, “On the properties of neural machine translation: Encoder-decoder approaches,” arXiv preprint arXiv:1409.1259 , 2014

    2014 arXiv

  18. [26]

    Virtual typing by people with tetraplegia using a self-calibrating intracortical brain-computer interface,

    B. Jarosiewicz, A. A. Sarma, D. Bacher, N. Y . Masse, J. D. Simeral, B. Sorice, E. M. Oakley, C. Blabe, C. Pandarinath, V . Gilja et al., “Virtual typing by people with tetraplegia using a self-calibrating intracortical brain-computer interface,” Science Translational Medicine...

    2015

  19. [27]

    High-performance brain-to-text communica- tion via handwriting,

    F. R. Willett, D. T. Avansino, L. R. Hochberg, J. M. Henderson, and K. V . Shenoy, “High-performance brain-to-text communica- tion via handwriting,” Nature, vol. 593, no. 7858, pp. 249–254, 2021

    2021

  20. [28]

    A high performance spelling system based on EEG-EOG signals with visual feedback,

    M.-H. Lee, J. Williamson, D.-O. Won, S. Fazli, and S.-W. Lee, “A high performance spelling system based on EEG-EOG signals with visual feedback,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 26, no. 7, pp. 1443–1459, 2018

    2018

  21. [29]

    Machine translation of cortical activity to text with an encoder-decoder framework,

    J. G. Makin, D. A. Moses, and E. F. Chang, “Machine translation of cortical activity to text with an encoder-decoder framework,” Nature Neuroscience, vol. 23, no. 4, pp. 575–582, 2020

    2020

  22. [30]

    Speech synthesis from neural decoding of spoken sentences,

    G. K. Anumanchipalli, J. Chartier, and E. F. Chang, “Speech synthesis from neural decoding of spoken sentences,” Nature, vol. 568, no. 7753, pp. 493–498, 2029

  23. [31]

    Imagined speech classification with EEG signals for silent communication: A preliminary in- vestigation into synthetic telepathy,

    K. Brigham and B. V . Kumar, “Imagined speech classification with EEG signals for silent communication: A preliminary in- vestigation into synthetic telepathy,” in 2010 4th International Conference on Bioinformatics and Biomedical Engineering , 2010, pp. 1–4

    2010

  24. [32]

    Decoding speech perception from non-invasive brain recordings,

    A. Defossez, C. Caucheteux, J. Rapin, O. Kabeli, and J.-R. King, “Decoding speech perception from non-invasive brain recordings,” Nature Machine Intelligence , pp. 1–11, 2023

    2023

  25. [34]

    Speech motor cortex enables BCI cursor control and click,

    T. Singer-Clark et al., “Speech motor cortex enables BCI cursor control and click,” bioRxiv (Cold Spring Harbor Laboratory), Nov. 2024, doi: 10.1101/2024.11.12.623096

    2024 doi

  26. [35]

    METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judg- ments,

    S. Banerjee and A. Lavie, “METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judg- ments,” ACL Anthology , Jun. 01, 2005. https://aclanthology.org/ W05-0909/

    2005

  27. [36]

    The brain’s bitter lesson: Scaling speech decoding with Self-Supervised Learning,

    D. Jayalath, G. Landau, B. Shillingford, M. Woolrich, and O. P. Jones, “The brain’s bitter lesson: Scaling speech decoding with Self-Supervised Learning,” arXiv.org, Jun. 06, 2024. https://arxiv. org/abs/2406.04328. 19 APPENDIX A WEBSITE A full description of the web-based sys...

    2024 arXiv

  28. [39]

    The system checks for unique email addresses before permitting new account creation to prevent duplicate accounts

    To create an account, a user must enter their name and email and set their desired password during registration. The system checks for unique email addresses before permitting new account creation to prevent duplicate accounts

  29. [40]

    The data upload section becomes available to users after they suc- cessfully log into the system

    Registered users can log into the platform using their email address and password. The data upload section becomes available to users after they suc- cessfully log into the system

  30. [41]

    JMECT applies the GPT-4 and BART models to data processed at this stage

    A pickle file containing EEG data is available for upload by platform users when submitting their data to the system. JMECT applies the GPT-4 and BART models to data processed at this stage

  31. [42]

    BART generates the initial paragraph, while the GPT-4 processing algorithm refines a new text output

    The results are displayed through two subsections that separate BART Generated Text from GPT-4 Refined Text. BART generates the initial paragraph, while the GPT-4 processing algorithm refines a new text output

  32. [43]

    Through its integration of the gTTS (Google Text-to-Speech) software, the platform converts text outputs into spoken audio

    A text-to-speech function integrated within the UI enables users to hear the provided texts. Through its integration of the gTTS (Google Text-to-Speech) software, the platform converts text outputs into spoken audio. Users can directly access the audio through the website afte...

  33. [44]

    A logout option appears within the system, allowing users to return to the home page of the platform following session termination. C. Frontend Structure Users create their interface elements through HTML templates and generate dynamic outputs using the Jinja2 engine. Importan...

  34. [45]

    base.html: Functions as the fundamental design template, generating both header and footer ele- ments while providing CSS styling and the primary content section for the entire website

  35. [46]

    signup.html: Allows users to become members by requesting a name, password, and email while displaying field occupation warnings for duplicate email entries

  36. [47]

    signin.html: Implements user authentication through an email-address and password form while validating the authentication credentials

  37. [48]

    Users need to use the logout button for exit as shown in Figure 10

    index.html: Enables users to upload pickle file data. Users need to use the logout button for exit as shown in Figure 10

  38. [49]

    results.html: Displays decoded text, text-to- speech controls, and BART and GPT-4 responses

  39. [50]

    index2.html: Provides access to the text entry page featuring the text-to-speech functionality as shown in Figure 11. D. Backend Architecture The backend implements its structure using FastAPI as its Python framework. The application handles user verification, data input, EEG ...

  40. [51]

    The demonstration storage of passwords in plain text requires migration to hashing methods for production security

    A SQLite database stores all information regard- ing user management, including username, email address, and password. The demonstration storage of passwords in plain text requires migration to hashing methods for production security

  41. [52]

    During registration, the system checks for duplicate emails while validating user-provided login creden- tials

  42. [53]

    The generated text appears as an output that gets saved within the system

    GPT-4 and BART models function within the data management system for pickle file uploading and processing. The generated text appears as an output that gets saved within the system

  43. [54]

    Through the gTTS library, the application converts text output into audio files, which get saved in the static directory before being forwarded to the frontend

  44. [55]

    The session period ends at logout time

    The FastAPI session management tracks all users who have active logins. The session period ends at logout time. 20 Fig. 9. Signin and Signup pages Fig. 10. Submit page that used to upload the pickel file 21 Fig. 11. Predicted and Refined string page and contain the audio files

  45. [56]

    The system handles upload errors, authentication issues, and speech synthesis errors by delivering essential updates to users. E. Deployment Strategy The FastAPI application receives its public URL from ngrok when conducting platform launches. Ngrok cre- ates a channel that co...

  46. [2023]

    Available: https://arxiv.org/abs/2303.08774

    arXiv

  47. [2025]

    Available: https://www.ncbi.nlm.nih.gov/books/ NBK559315/

    [Online]. Available: https://www.ncbi.nlm.nih.gov/books/ NBK559315/

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.