Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Reconstructing continuous speech from scalp electroencephalography (EEG) remains fundamentally challenging. EEG provides a weak, spatially diffuse, and highly variable measurement of distributed cortical activity, whereas speech is organized as a coherent acoustic trajectory with strong harmonic and temporal structure. The resulting mismatch makes waveform regression unstable and causes stochastic multi-step generation to be sensitive to artifact-dependent conditioning and subject variability. We introduce \textbf{NeuroSonic}, a conditional flow-matching framework for EEG-to-speech reconstruction. Instead of predicting waveforms directly or refining them through stochastic denoising, NeuroSonic learns a deterministic probability-flow velocity field that transports a noise-corrupted acoustic state toward clean speech under EEG conditioning. EEG and audio are embedded into a shared token space and processed by a time-conditioned gated Transformer that parameterizes the transport ordinary differential equation. This formulation models trajectory evolution explicitly while avoiding iterative stochastic sampling. We evaluate NeuroSonic on the CineBrain and EAV benchmarks under cross-subject evaluation. Across both datasets, the proposed method improves distributional realism, spectral fidelity, and perceptual quality over representative GAN-, diffusion-, and mean-flow baselines, with up to a 26.3\% gain in overall perceptual quality. The performance gap is most evident in artifact-heavy segments, where conditioning variability is strongest. These findings indicate that deterministic conditional transport provides a stable and effective formulation for EEG-driven speech reconstruction.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4295_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/4295_supp.zip

Link to the Code Repository

https://github.com/Y-Research-SBU/NeuroSonic

Link to the Dataset(s)

N/A

BibTex

@InProceedings{GaoWen_NeuroSonic_MICCAI2026,
        author = { Gao, Wenhao AND Wang, Yifan AND Ma, Yijia AND Yang, Carl AND Li, Wen AND You, Chenyu},
        title = { { NeuroSonic: Conditional Flow Matching for EEG-to-Speech Reconstruction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16894},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This work presents a novel technique for EEG-based speech reconstruction, which is based on conditional flow matching. The authors claim this reduces artifact driven noise and instabilities compared to direct waveform rejection. The results show that this method outperforms other generative techniques such as GANs or Diffusion models.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The idea of using conditional flow matching for this problem has been justified well and is a significant novel contribution. They evaluate the model in a cross subject task, which is vital to ensure robustness. The authors provide spectrogram, waveform and PSD plots for qualitative comparisons in addition to quantitative evaluations. They even provide supplementary files with reconstructed audio.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1) The authors claim this in the introduction “speech evolves along a highly organized acoustic trajectory characterized by harmonic structure and temporal coherence” and “ This perspective is well suited to speech reconstruction, where temporal coherence is intrinsic to the signal structure. “. However, I feel the authors have confused the acoustic temporal dynamics and the time variable used in the flow matching process, the latter having nothing to do with actual EEG/speech time. As I understand it, the t variable (called continuous time by the authors) is just a parameter in the conditional flow matching algorithm used to map the noisy sample (t=0) to the clean sample (t=1). It has no relation to actual temporal characteristics of the waveform as far as I know. Even the time-conditioned transformer works with this t variable and not actual speech/EEG time. The authors should carefully explain how the flow method leads to better temporal coherence, or else remove such claims in the introduction.

    2) A related issue is that the explanation of conditional flow matching is not provided in enough detail in this paper. The explanations could be better structured with clear inputs and outputs for each component of the model and corresponding objectives. For example, it is not clear exactly where zt enters the model. It is not clear how the output Z of equation 10 is used in the next line Xpred = net(zt, t, E). If E is used for Xpred, where is Z used? The audio decoder that is used is also not specifically mentioned.

    3) The baselines for comparison seem to be constructed and implemented by the authors. Perhaps it would have been better to also compare with existing models that were specifically built for EEG-speech reconstruction. This would allow direct comparison with published work.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
    • It would be better to add some numerical results to the abstract
    • In Fig 1, what is zt, epsilon? Since these terms are defined much later on, it would be better to give a short description in the caption. Also, the first part of the caption for (c) actually describes (b) when talking about the decoder. Instead make it specific to the figure (c)
    • There is an unnecessary repetition of the definitions of E and X.
    • Define AdaLN in eq 7 and 8
    • Fig 3 seems to show that DM is almost as good or even better than NeuroSonic. The authors should note this and discuss it.
    • The authors should cite recent work on flow-based EEG-speech reconstruction. For example: Xiang, Shu, Haobo Ling, and Meihong Wu. “Cross-Modal Alignment and Rectified Flow-Based Latent Representation Synthesis for Enhanced Speech-Driven Alzheimer’s Disease Detection. “ Bioengineering 13.3 (2026): 370.
  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I think the paper provides a novel contribution in terms of applying conditional flow matching to this application. However, it could benefit from clearer explanations and structure especially in the methods.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors propose NeuroSonic, a conditional flow matching framework that reconstructs continuous speech from scalp EEG. The method reformulates EEG-to-speech reconstruction as conditional acoustic transport: a time-conditioned gated Transformer parameterizes a probability-flow ODE that maps a noise-corrupted acoustic state toward clean speech under EEG conditioning. EEG and audio are patchified and embedded into a shared token space, and the learned ODE is integrated with a fixed-step Heun solver at inference. The authors evaluate NeuroSonic on two public EEG-audio benchmarks, CineBrain and EAV, against GAN, diffusion, and mean-flow baselines under cross-subject splits.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper offers a coherent formulation of EEG-conditioned speech reconstruction as deterministic acoustic transport. Compared with stochastic diffusion, a single-pass ODE integration avoids artifact-dependent noise schedules that can degrade reconstruction across sampling steps. The motivation is well aligned with the statistical properties of scalp EEG described in the introduction.
    • The use of RMS-normalized attention on queries and keys is a well-motivated design choice. Flow matching with time-dependent interpolation is known to induce feature-scale drift across interpolation times, and the authors explicitly address this source of instability.
    • The evaluation compares NeuroSonic against GAN, diffusion, and mean-flow baselines on two public datasets under consistent cross-subject splits. The inference-time advantage of flow matching over stochastic diffusion is clearly substantiated by the reported numbers.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    -The evaluation does not validate cross-stimulus generalization. All CineBrain participants are exposed to the same audiovisual stimuli, so a cross-subject split leaves the test audio already present in the training set, paired with other subjects’ EEG. Under this setup, the model can learn to associate an EEG pattern with an already-seen audio segment, rather than decode speech from neural activity. A held-out-episode protocol or cross-dataset generalization from CineBrain to EAV is necessary to support the central claim of EEG-to-speech reconstruction.

    • The reported metrics measure distributional realism and reference-free perceptual quality, but none assess sample-wise correspondence between the generated waveform and the specific stimulus the subject was listening to. Correspondence-level metrics such as ASR-based WER or stimulus identification accuracy are missing. In particular, NeuroSonic exceeding the ground-truth recording on EAV OVRL is more consistent with a prior-dominant generator sampling from a narrow acoustic manifold than with faithful decoding.
    • The ablation study is confined to the internal loss formulation, namely clean-state velocity versus endpoint x-loss regression. There is no ablation that isolates the causal contribution of EEG conditioning itself. A shuffled-EEG baseline with mismatched pairs and a null-EEG baseline with zero or Gaussian input are the most direct way to establish that the reported improvements arise from neural decoding rather than from flow matching fitting a narrow audio prior.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    EAV consists of spontaneous conversational recordings, which may implicitly provide cross-stimulus structure since different subjects likely produced different utterances. If the authors clarify this and report, for instance, that there is no overlap between train and test subjects’ speech content, the main concern on stimulus leakage becomes substantially weaker on EAV. This would be a low-cost but high-impact rebuttal addition.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a technically sound application of conditional flow matching to EEG-to-speech reconstruction, with stable training, cross-subject generalization, and consistent improvements over GAN, diffusion, and mean-flow baselines. However, the current evaluation cannot distinguish EEG-driven decoding from class-conditional speech generation over a narrow audio manifold, since CineBrain is shared-stimulus across subjects and all reported metrics measure marginal distribution quality rather than sample-wise correspondence with the target stimulus. I recommend weak accept conditional on the rebuttal addressing stimulus overlap between train and test splits, providing at least one correspondence-level metric such as ASR-based WER, and reporting a null or shuffled EEG baseline.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal partially addresses my main concerns. The added ASR-WER results are useful because they assess content correspondence beyond FAD/LSD/SC and DNSMOS, and the clarification on EAV reduces the stimulus-overlap concern for that dataset.

    However, I do not think the Null-EEG result fully establishes EEG-driven decoding. Since the control uses Gaussian EEG, the degradation can mainly reflect an out-of-distribution conditioning input. A stronger control would be to shuffle real EEG segments within the split and test whether the reconstruction follows the mismatched EEG-paired speech or remains driven by the acoustic prior/corrupted audio state.

    Overall, the paper remains technically sound and the rebuttal improves the evidence, but the central EEG-to-speech claim is still not fully separated from speech-prior-driven reconstruction. I therefore keep a marginal positive recommendation.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a method for reconstructing speech from EEG signals using a generative modeling framework based on conditional flow matching. The approach uses a Transformer-based architecture to model a deterministic process that transforms noise into speech while being conditioned on EEG. However, the formulation and overall objective are not clearly articulated, making it difficult to understand whether the method performs true EEG-to-speech reconstruction or a form of EEG-guided speech generation.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • EEG-to-speech reconstruction, which is an important direction for brain-computer interfaces and assistive communication.
    • The use of modern generative modeling techniques, particularly conditional flow matching, is timely and aligns with recent advances in speech generation and multimodal learning.
    • The overall idea of leveraging generative models for neural decoding is interesting and could be promising if supported by stronger evaluation and analysis.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The overall objective is unclear: it is difficult to determine whether the method targets speech reconstruction, generation, or denoising.
    • The paper does not clearly justify why EEG should be used within a generative denoising/transport framework, nor what specific information (e.g., phonetic, temporal, or semantic cues) is expected to guide reconstruction. Prior work in neural speech decoding (e.g., Willett, 2023; Anumanchipalli, 2019) explicitly models linguistic or articulatory representations, which is not clearly defined here.
    • The methodological novelty is unclear, and the paper does not demonstrate why this formulation is particularly suited to EEG-to-speech reconstruction compared to existing approaches.
    • The evaluation relies on audio-quality metrics (FAD, LSD, SC, DNSMOS), which do not assess whether the reconstructed speech preserves correct linguistic content. In contrast, prior work evaluates intelligibility using ASR-based metrics (WER/CER), phoneme accuracy, or human transcription.
    • The paper lacks standard control experiments (e.g., EEG shuffling, temporal misalignment, or no-EEG baselines), making it unclear to what extent the model actually leverages EEG rather than learned audio priors.
    • The ablation study is limited and does not isolate key components; comparisons with GAN and diffusion models are not controlled for architecture or capacity, making it difficult to attribute improvements to the proposed method.
    • No comparison is provided against simpler baselines (e.g., regression-based decoding, CCA, or speech denoising/enhancement), leaving unclear whether the proposed approach offers a meaningful advantage.
    • Qualitative results (e.g., spectrograms) suggest possible hallucinated harmonic structures, raising concerns about faithful reconstruction.
    • The datasets include complex audio (speech, music, environmental sounds), but no analysis is provided to separate these components or assess their impact.
    • Key implementation details (preprocessing, EEG–audio alignment, segmentation, inference) are insufficiently described, limiting reproducibility.
  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses an interesting problem, but the current submission does not provide sufficient evidence to support its claims. In particular, the lack of clear motivation, limited ablation analysis, and absence of appropriate evaluation for speech reconstruction (e.g., intelligibility metrics and EEG-specific controls) make it difficult to assess the actual contribution of the proposed method. Additionally, the experimental comparisons are not sufficiently controlled to isolate the effect of the proposed approach. As a result, the work does not yet meet the standard expected for acceptance.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    Thank you for the response. The added WER results and Null-EEG control are useful. The WER values remain a bit high, suggesting limited linguistic-content preservation. Please clarify how WER was computed, which ASR system was used, and where the method fails. Regarding harmonic fidelity, my concern was not that harmonics are unexpected in speech. Rather, it is unclear whether the observed harmonic/formant structures are faithfully reconstructed from the paired reference, driven by EEG conditioning, or produced by the synthesis/transport prior. This concern is consistent with the high WER values, which suggest substantial reconstruction or synthesis errors, even if WER does not directly measure harmonic fidelity.

    Finally, the lack of ablations remains a major concern. The Null-EEG control is helpful but insufficient to isolate the roles of EEG conditioning, the acoustic prior, flow matching, model capacity, and learned synthesis priors. Therefore, the main concerns remain unresolved.



Author Feedback

[R1.1/R1.2] Flow notation: t is a flow-path coordinate, not physical EEG/speech time; we will revise the Intro/Sec. 2.2.In training, the audio encoder takes z_t, a corrupted full speech segment, not clean X. EEG/corrupted-audio tokens form Z, which is processed by the t-conditioned Transformer; acoustic tokens are then MLP-decoded to X_pred. Temporal coherence arises indirectly from EEG-conditioned deterministic transport of the spectro-temporal structure of z_t toward X.

[R1.3/R3.6/R3.7] Baselines: All learned baselines use the same preprocessing and comparable parameter budgets. We will add FESDE (2406.08644), and a denoising baseline. NeuroSonic improves FESDE on Cine: FAD/LSD/SC/DNSMOS(SIG/BAK/OVRL)/WER 62.78/29.69/1.07/1.61-1.62-1.32/1.01 → 39.06/14.24/0.64/1.95-1.64-1.44/0.49, and on EAV: LSD/SC/DNSMOS/WER 24.07/1.08/2.94-2.77-2.24/0.99 → 12.98/0.28/3.31-3.07-2.59/0.46.It also outperforms denoising: Cine FAD/LSD/SC 231.73/120.27/1.13 vs. 39.06/14.24/0.64; EAV 103.66/52.31/0.98 vs. 11.64/12.98/0.28. [R1.4–R1.7] Text fixes: We will add representative numbers to the abstract; define z_t, ε, and t in Fig. 1; correct panels (b)(c); define E and X only once; and add the AdaLN definition near Eqs. 7–8. [R1.8] Spectral trade-off: Fig. 3 shows a trade-off: DM is closer to GT in some high-frequency regions, while NeuroSonic better matches speech-relevant low/mid bands. Metrics support this: WER on Cine/EAV drops 0.97/0.97 → 0.49/0.46. [R1.9] Related work: We will cite Xiang et al. as recent flow-based EEG-speech work.

[R2.1] Generalization: EAV is conversational rather than a shared fixed-stimulus set; with our cross-subject split, both test subjects and utterances are unseen. For CineBrain, stimuli are shared, but the split is subject-disjoint.

[R2.2/R3.4] Content correspondence: We add ASR-WER as a sample-correspondence metric. NeuroSonic reduces WER vs. DM from 0.97 to 0.49 on Cine and 0.97 to 0.46 on EAV, showing better paired-content alignment. DNSMOS is reference-free: EAV OVRL above GT reflects cleanness/noise suppression, not content fidelity; WER checks content.

[R2.3/R3.5] EEG control: We add Null-EEG: trained NeuroSonic with fixed weights, but Gaussian-noise EEG; WER only on dialogue-speech. Performance drops: Cine FAD/LSD/SC/WER 39.06/14.24/0.64/0.49 → 155.64/92.28/0.99/1.02; EAV 11.64/12.98/0.28/0.46 → 17.84/19.44/1.00/0.97.With the acoustic prior fixed, valid EEG adds reconstruction constraints.

[R3.1] Objective: NeuroSonic performs EEG-conditioned acoustic reconstruction, given aligned EEG, it reconstructs paired speech using EEG conditioning and an acoustic transport prior.

[R3.2] Motivation: EEG contains speech-evoked responses containing timing, rhythmic/prosodic, and coarse phonetic/semantic cues. Its non-invasiveness makes EEG-guided reconstruction relevant to assistive speech generation for severe speech impairment, while transport supplies coherent acoustics. Unlike Willett/Anumanchipalli, which use invasive recordings and explicit phoneme representations, we use scalp EEG as coarse transport conditioning.

[R3.3] Method novelty: We formulate EEG-to-speech as conditional flow matching, enabling weak neural cues to guide acoustic reconstruction without stochastic denoising chains. This makes the approach well suited to low-SNR, cross-subject EEG.

[R3.8] Harmonic fidelity: Spectrograms do not indicate hallucinated harmonics: harmonic/formant patterns are intrinsic to speech, and reconstructions follow reference temporal modulation, not arbitrary templates.

[R3.9] Audio composition: Cine composition analysis shows speech-dominant vs. non-speech/background-dominant FAD/LSD/SC = 37.31/11.19/0.62 vs. 40.72/16.34/0.65 and OVRL = 1.47 vs. 1.40. [R3.10] Reproducibility: Secs. 3.1/3.2 already describe preprocessing/inference. Original datasets synchronize EEG-audio pairs; each EEG segment is recorded during its paired audio stimulus. We use 4-s CineBrain and 5-s EAV windows.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Please provide justifications and clarifications for the concerns raised by the reviewers.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper proposes NeuroSonic, a conditional flow matching framework for EEG-to-speech reconstruction. Reviewer scores are 5/4/2, with R1 accepting unconditionally, R2 accepting marginally post-rebuttal, and R4 rejecting.

    Reviewers broadly agree that the targeted application is clinically relevant, the formulation is well-motivated, and the method consistently outperforms baselines across two datasets.

    The shared concerns fall in the evaluation protocols, such as how the train/test sets are split, the absence of correspondence-level metrics, lack of EEG-specific controls, and limited ablation. The rebuttal partially addresses these by adding WER results, a Null-EEG control showing substantial degradation without valid EEG, and clarifying that test subject EEG is always unseen. The shuffled real-EEG control remains unresolved after the rebuttal.

    Considering that the proposed method is genuinely interesting and the evaluations, though not ideal, sufficiently support the claims, the strength outweighs the weaknesses and I recommend acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    R4’s reject rests on concerns (unclear EEG usage, hallucinated harmonics) that despite worth taking into account, but that R1 and R2 -both more senior reviewers- do not share. In order to understand this discrepancy I think it is worth noting that R4 categorized this as an “application study” while R1 and R2 both categorized it as a methodological contribution.

    In their rebuttal, the authors credibly attribute R4 concerns to a misunderstanding of scalp-EEG constraints and basic speech acoustics. The null-EEG control demonstrates meaningful EEG conditioning, and R2’s post-rebuttal framing as “EEG-conditioned acoustic reconstruction” I believe appropriately scopes the contribution - the framing mismatch likely contributed to R4’s harsh assessment. Moreover, R2 upgraded his score to accept post-rebuttal after the authors provided ASR-WER metrics and clarified the EAV stimulus-overlap concern.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The rebuttal address the reviewers’ concerns but not in full. The work is better packaged as EEG-conditioned acoustic reconstruction, and it should be clarified what the source of the performance gain is.



back to top