Pith. sign in

REVIEW 3 major objections 3 minor 37 references

A state-guided gate lets EEG attention decoding switch faster without sacrificing stability.

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

2026-08-05 00:05 UTC pith:UN24T2YS

load-bearing objection A solid applied contribution: the SGAD decision module and the six-protocol evaluation scheme are genuinely new, but the switch-event metrics lean heavily on generous keypress-derived labels and lack statistical support. the 3 major comments →

arxiv 2608.01618 v1 pith:UN24T2YS submitted 2026-08-03 eess.SP cs.SD

SGAD: A State-Guided Adaptive Decision Framework for Robust EEG-Based Auditory Attention Switch Decoding

Yuting Ding , Xuefei Wang , Ximin Chen , Chunlin Li , Fei Chen This is my paper
classification eess.SP cs.SD
keywords EEGauditory attention switch decodingstate-guided adaptive decisioncausal state detectionadaptive gatingtemporal smoothingcross-condition evaluationhearing aids
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 tries to solve a practical bottleneck in EEG-based auditory attention switch decoding: when a listener shifts attention between two speakers, the decoder must react quickly but also stay stable between switches. Simple post-processing that smooths the decoder output over time reduces flicker but slows down detection of real switches. The authors propose a decision-level framework, SGAD, that estimates whether an attention transition is happening and adjusts the amount of temporal smoothing accordingly. They report that this state-guided approach raises average switch detection F1 from 53.7% to 67.3% while keeping response latency lower than fixed smoothing, and they add six evaluation protocols showing how much performance drops when audio, speaker, and subject are held out. If correct, the contribution is a lightweight, encoder-agnostic way to improve dynamic attention tracking for neuro-steered hearing aids.

Core claim

The paper claims that adding a causal state detector and an adaptive gating mechanism to an existing EEG-speech matching decoder substantially improves the trade-off among decoding accuracy, switch detection, and response latency. The causal state detector uses causal multi-head attention over a sliding window of EEG embeddings to output a transition probability at each time step. That probability is mapped through a learnable monotonic gate to a smoothing factor: high transition probability reduces smoothing so the system can follow the switch quickly, low probability increases smoothing to suppress noise in stable periods. Trained with a triangular soft label around each keypress-reported

What carries the argument

The central mechanism is the state-guided adaptive gate: a causal state detector (CSD) with a key-value cache and multi-head attention produces a transition probability s-hat_t, and a learnable sigmoidal mapping converts s-hat_t into a smoothing factor g_t between 0.05 and 0.95. This factor feeds a recursive smoother on the decision margin delta-r-hat_t = g_t * delta-r-hat_{t-1} + (1-g_t) * delta-r_t, where delta-r_t is the current EEG-speech correlation margin. The gate decouples the fixed stability-versus-latency trade-off of conventional EMA smoothing, because the smoothing strength adapts to the estimated state.

Load-bearing premise

The framework's training and evaluation assume that the participant keypresses mark the true attention-switch times closely enough, so the soft triangular labels teach a genuine transition signal and the reported switch metrics are unbiased.

What would settle it

Shift every keypress-reported switch time by random offsets of 1-3 seconds in both the training labels and the evaluation ground truth, and rerun LOTO and LOSSO. If SGAD's Sw-F1 and SDL degrade substantially more than the fixed-smoothing EBD baseline, then the state detector has learned the labeling noise rather than actual attention transitions.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Because SGAD is a decision-level module that operates on the output margin of any EEG-speech matching encoder, it can be attached to existing AASD systems without retraining the encoder, offering a direct performance upgrade in deployment.
  • The reported jump in Sw-F1 from 53.7% to 67.3% at similar or lower latency suggests that sequential decision strategy matters as much as encoder quality for dynamic attention tracking.
  • Under the six protocols, accuracy falls from about 85% in LOTO to about 77% in LOSSO, so models evaluated only on random trial splits will overstate real-world performance when speaker and audio content are unseen.
  • The low SDL of 1.18 s across protocols indicates that adaptive smoothing can keep hearing-aid response times near those of no smoothing while giving the stability of fixed smoothing.
  • The ablation results show that both the causal temporal context and the learnable gate contribute to the improvement, implying that neither simple thresholding of transition probability nor a fixed mapping reproduces the full gain.

Where Pith is reading between the lines

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

  • The same state-guided gating idea could be transferred to other non-stationary sequential decoding problems, such as speech-activity detection, EMG gesture switching, or brain-computer interface command changes, wherever a trade-off between stability and responsiveness exists.
  • The triangular soft label with R = 5 s deserves scrutiny: since it makes roughly half of all windows non-zero, the state detector may be trained to predict a broad transition zone rather than the exact switch time, which would partly explain the low latency but might also limit precise switch localization.
  • If the keypress reports carry systematic delay, the reported SDL of 1.18 s may be partly absorbing that delay; a separate behavioral ground truth such as eye tracking or speech-pause analysis could reveal whether SGAD actually reacts to neural transitions or to correlates of the keypress.
  • The six-protocol ranking could be turned into a standardized AASD benchmark that reports the full protocol matrix, making it harder for future methods to cherry-pick an easy split.

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 / 3 minor

Summary. The paper proposes SGAD, a state-guided adaptive decision framework for EEG-based auditory attention switch decoding. SGAD adds a causal state detector (CSD) that estimates attention-switch probability from a causal transformer over cached EEG embeddings, and an adaptive gating mechanism that modulates the smoothing factor of a recursive decision-margin filter (Eqs. 1-5). The framework is trained with a multi-task objective containing decoding, state-supervision, and gate-smoothness terms (Eq. 6). The authors evaluate SGAD against window-wise decoding, persistence-based decision, and exponential moving average baselines across two EEG encoders and six evaluation protocols that vary audio, speaker, and subject partitions (Tables 1-2). They report that SGAD improves mean decoding accuracy to 79.8% and switch-F1 to 67.3% while keeping switch detection latency at 1.18 s, compared with 79.0% / 53.7% / 1.38 s for EBD. An ablation study (Table 3) attributes the gains to the CSD, the learnable gate, the auxiliary state loss, and the smoothness regularization. The paper also uses the protocol-to-protocol performance degradation to argue that data-partition biases should be evaluated hierarchically.

Significance. If the reported results hold, the contribution is meaningful: SGAD is an encoder-agnostic decision-level module that addresses a real limitation of simple temporal smoothing in dynamic AASD, and the six-protocol evaluation framework is a useful step toward more honest generalization assessment in this area. The paper explicitly targets confounding factors such as audio content and speaker identity, and the observed monotonic performance decline under stricter protocols is a plausible and valuable result. The ablation study is well structured and supports the claim that each component contributes. However, the manuscript currently lacks the statistical evidence needed to establish that the improvements are not within run-to-run or cross-subject variability, and the switch-detection metric is calibrated in a way that could systematically favor the proposed method. The dataset is publicly referenced, which is a strength, but no code is released to verify the event-matching pipeline.

major comments (3)
  1. [Sec. 3.3, Eq. (8)] The 5.0 s tolerance for a true positive switch is defined as the 99th percentile of validation-set detection latency. Because this tolerance is computed from the models being evaluated (and SGAD is explicitly trained to reduce latency), Sw-F1 is not a fixed, pre-specified metric; part of SGAD's Sw-F1 advantage may be an artifact of calibrating the tolerance to its own detection behavior. Please report Sw-F1 under a fixed set of tolerances (e.g., 1, 2, 3, 5 s) or derive the tolerance from independent behavioral data, and show that the ranking in Table 2 is robust to this choice.
  2. [Table 2 and Table 3] No error bars, confidence intervals, or statistical tests are reported for any protocol-encoder condition. The central claim (Sec. 4.1) that SGAD achieves a better trade-off and that this is 'consistently observed across all experimental settings' rests entirely on point estimates. With only 13 subjects and several hyperparameters (K, alpha, W, lambda_1, lambda_2) selected on validation trials, the observed differences, especially the mean Sw-F1 gain from 53.7% to 67.3%, could be within run-to-run or cross-subject variability. Please report per-fold or per-subject variability, and paired significance tests across protocols (e.g., permutation tests or Wilcoxon signed-rank tests) for Acc, Sw-F1, and SDL.
  3. [Sec. 2.2.3 and Sec. 3.1] Both the CSD training labels (triangular s_t with R=5 s) and the evaluation ground truth for Sw-F1/SDL are derived from self-reported keypress times. The paper acknowledges keypress timing variability through dominant-label windowing, but it does not quantify the lag or noise in the keypresses. If keypresses systematically lag the true neural attention switch, the CSD will learn a delayed state signal, and SGAD's lower SDL relative to EBD (1.18 s vs. 1.38 s) may reflect faster reaction to keypress-correlated EEG features rather than to true neural switches. Please provide an analysis of keypress timing (e.g., distribution of inter-switch intervals or agreement with a neural marker), and/or a sensitivity analysis training with different R values and evaluating with fixed temporal offsets.
minor comments (3)
  1. [Sec. 2.2.3] The values of lambda_1 and lambda_2 are stated as 'based on validation experiments' but the search range or selection criterion is not given. Please add a sentence or a supplementary table.
  2. [Sec. 2.2.1, Eq. (2)] The notation k_{t-W+1:t} and v_{t-W+1:t} is not defined precisely; specify that the cache stores the last W keys/values. Also, the 'Softmax' operator should be typeset consistently.
  3. [Table 1] The row 'LOTO / LOSO' with 'Seen / Unseen' is ambiguous because the subject column has two entries. Clarify which cell corresponds to LOTO (subject-dependent) and which to LOSO (subject-independent).

Circularity Check

0 steps flagged

No significant circularity: SGAD's gains are evaluated on held-out protocol folds and the decision module is trained with standard supervised losses, not by construction.

full rationale

The derivation chain is self-contained: EEG–speech matching produces correlation scores, the CSD estimates a transition probability from causal attention, the AGM maps that probability to a smoothing factor, and the recursive smoother updates the decision margin. Training uses L_dec, L_state, and L_smooth, where the state labels are triangular functions of keypress-reported switch times; this is ordinary supervised training, not a circular definition. The reported Acc, Sw-F1, and SDL numbers come from held-out folds of six protocols, with hyperparameters selected on a reserved validation set, so the central improvement is not a fitted re-statement of the training input. The only mildly self-referential element is the Sw-F1 tolerance (5.0 s) being set to the 99th percentile of validation-set detection latency; however, it is a fixed threshold applied identically to all decision frameworks and does not by construction force SGAD's higher Sw-F1. Self-citations to the MS-AASD dataset and to the authors' earlier encoder architectures are data and backbone inputs, not load-bearing circular justification; the dataset is externally deposited. The concern about keypress timing noise is a construct-validity issue, not a circularity, because the model can fail to recover the labels. No equation reduces a prediction to an input by definition.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on supervised switch labels from keypresses, several validation-tuned hyperparameters, and a hand-designed gate; no new physical entities or external theoretical objects are introduced.

free parameters (6)
  • gate slope a and bias b = learned
    Learnable parameters in Eq. (4) mapping transition probability to smoothing factor.
  • loss weights lambda_1 and lambda_2 = 0.5, 0.05
    Chosen via validation experiments (Sec. 2.2.3).
  • gate bounds g_min, g_max = 0.05, 0.95
    Hand-set bounds of the smoothing factor in Eq. (4).
  • triangular label width R = 5 s
    Sets the soft-label support for the switch state and yields non-zero labels for about half the windows (Sec. 2.2.3).
  • causal cache window W = 15
    Selected by validation grid search over W in {5, 10, ..., 25}; controls the temporal context of the state detector.
  • baseline decision hyperparameters K and alpha = K = 5, alpha = 0.8
    Selected via validation grid search; they affect the fairness of the PBD and EBD comparisons.
axioms (5)
  • domain assumption Keypress timestamps approximate true attention switch times and can serve as ground truth for training and evaluation.
    Used to build the triangular state labels in Sec. 2.2.3 and the Sw-F1/SDL metrics in Sec. 3.3.
  • domain assumption Triangular soft labels with width R = 5 s provide useful supervision for the transition state s_t.
    Invoked in Sec. 2.2.3 to define L_state and to justify the smoothing gate's behavior.
  • domain assumption The three male-female speaker pairs and audio stimuli in MS-AASD provide enough content diversity for leave-one-speaker and leave-one-audio protocols to expose speaker/audio confounds.
    Underpins the six-protocol evaluation claims in Sec. 3.2 and the bias discussion in Sec. 4.2.
  • ad hoc to paper The monotonic sigmoid gate in Eq. (4), with chosen bounds, is an appropriate functional form for state-guided smoothing.
    This design choice is not derived from theory or from external benchmarks; it is introduced for the proposed framework.
  • standard math Causal multi-head attention with a key-value cache and layer normalization behaves as a valid sequence model for the state detector.
    Assumed from the transformer literature in Sec. 2.2.1; no formal proof is given.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SGAD: A State-Guided Adaptive Decision Framework for Robust EEG-Based Auditory Attention Switch Decoding." pith.science (2026). https://pith.science/paper/UN24T2YS

@misc{pith2026260801618,
  author       = {Pith},
  title        = {Pith review of: SGAD: A State-Guided Adaptive Decision Framework for Robust EEG-Based Auditory Attention Switch Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UN24T2YS}},
  note         = {Machine review of arXiv:2608.01618}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Achieving robust EEG-based auditory attention switch decoding (AASD) is crucial for intelligent hearing aids. However, its application is limited as EEG non-stationarity complicates sequential decision-making, and insufficient control of potential confounding factors may overestimate performance. Therefore, we propose a state-guided adaptive decision (SGAD) framework that infers attention transition states via causal state detection and dynamically modulates temporal smoothing through state-guided adaptive gating. We further introduce six hierarchical evaluation protocols to assess generalization across audio, speaker, and subject dimensions. Experimental results show that SGAD improves decoding accuracy and stability while maintaining low response latency across evaluation scenarios. Performance variations across protocols further suggest data partition-related biases. Together, these findings advance robust AASD for neuro-steered hearing applications.

Figures

Figures reproduced from arXiv: 2608.01618 by Chunlin Li, Fei Chen, Ximin Chen, Xuefei Wang, Yuting Ding.

Figure 1
Figure 1. Figure 1: Architecture of the proposed SGAD-based sequential decision framework with three stages: (i) sequential input partitioning via sliding windows, (ii) EEG-speech matching via frozen encoders, and (iii) state-guided adaptive decision making [PITH_FULL_IMAGE:figures/full_fig_p003_1.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

37 extracted references · 35 canonical work pages · 1 internal anchor

  1. [1]

    SGAD: A State-Guided Adaptive Decision Framework for Robust EEG-Based Auditory Attention Switch Decoding

    Introduction Individuals with normal hearing can selectively attend to a tar- get speaker in noisy environments, a phenomenon known as the “cocktail party effect” [1, 2]. However, people with hearing loss often struggle with this ability, posing a core challenge for hearing-assistive devices [3]. With advances in neuroscience, auditory attention decoding ...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    EEG-speech matching To provide attention evidence for sequential decision-making, we first formulate AASD as an EEG-speech matching problem

    Methods 2.1. EEG-speech matching To provide attention evidence for sequential decision-making, we first formulate AASD as an EEG-speech matching problem. For each timet, the multi-band EEG segmentE t ∈R N×C×T , whereN,C, andTdenote the number of sub-bands, chan- nels, and time steps, respectively, is mapped by an EEG encoder fEEG(·)toe t ∈R d. Similarly, ...

    work page

  3. [3]

    Experiments 3.1. Dataset and preprocessing Experiments were conducted on the MS-AASD dataset [26, 27] with EEG recordings from 13 normal-hearing adults (aged 21–28 years) performing self-initiated attention switching in mixed-speech environments. EEG was recorded at 500 Hz using 64 channels. In each 60-s trial, two root-mean-square (RMS)-normalized speech...

    work page

  4. [4]

    Performance and ablation analysis Table 2 compares SGAD with three baseline decision frame- works across two encoders and six evaluation protocols

    Results and Discussion 4.1. Performance and ablation analysis Table 2 compares SGAD with three baseline decision frame- works across two encoders and six evaluation protocols. WD achieves the lowest SDL of 0.89 s by operating independently on each window, but its Sw-F1 remains low at 29.0%, indicat- ing limited robustness in tracking attention transitions...

    work page

  5. [5]

    By adaptively modulating temporal smooth- ing according to inferred transition states, SGAD mitigates the trade-off between stability and responsiveness

    Conclusion This study addresses decision instability and evaluation bias in dynamic AASD by introducing a state-guided adaptive decision (SGAD) framework together with structured hierarchical eval- uation protocols. By adaptively modulating temporal smooth- ing according to inferred transition states, SGAD mitigates the trade-off between stability and res...

    work page

  6. [6]

    Acknowledgments This work was supported by the National Key Research and Development Program of China (2025YFF0518003, 2023YFF1203502), the National Natural Science Foundation of China (62371217), and the Center for Computational Science and Engineering at Southern University of Science and Tech- nology

  7. [7]

    Gener- ative AI tools were used for editing and polishing manuscripts, but were not used for producing a significant part of the manuscript

    Generative AI Use Disclosure All (co-)authors are responsible and accountable for the work and content of the paper, and consent to its submission. Gener- ative AI tools were used for editing and polishing manuscripts, but were not used for producing a significant part of the manuscript

  8. [8]

    Selective audi- tory attention decoding in naturalistic conversations using EEG- based speech envelope tracking in multi-speaker environments,

    G. Ivucic, S. Pahuja, D. Da Silva, and T. Schultz, “Selective audi- tory attention decoding in naturalistic conversations using EEG- based speech envelope tracking in multi-speaker environments,” inProceedings of the INTERSPEECH, 2025, pp. 2925–2929

    work page 2025

  9. [9]

    Some experiments on the recognition of speech, with one and with two ears,

    E. C. Cherry, “Some experiments on the recognition of speech, with one and with two ears,”Journal of the Acoustical Society of America, vol. 25, no. 5, pp. 975–979, 1953

    work page 1953

  10. [10]

    Neural coding of continuous speech in auditory cortex during monaural and dichotic listening,

    N. Ding and J. Z. Simon, “Neural coding of continuous speech in auditory cortex during monaural and dichotic listening,”Journal of Neurophysiology, vol. 107, no. 1, pp. 78–89, 2012

    work page 2012

  11. [11]

    The cocktail party problem,

    S. Haykin and Z. Chen, “The cocktail party problem,”Neural Computation, vol. 17, no. 9, pp. 1875–1902, 2005

    work page 1902

  12. [12]

    Attentional selection in a cocktail party environment can be decoded from single-trial EEG,

    J. A. O’sullivan, A. J. Power, N. Mesgarani, S. Rajaram, J. J. Foxe, B. G. Shinn-Cunningham, M. Slaney, S. A. Shamma, and E. C. Lalor, “Attentional selection in a cocktail party environment can be decoded from single-trial EEG,”Cerebral Cortex, vol. 25, no. 7, pp. 1697–1706, 2015

    work page 2015

  13. [13]

    Relating EEG to continuous speech using deep neural networks: A review,

    C. Puffay, B. Accou, L. Bollens, M. J. Monesi, J. Vanthornhout, H. Van Hamme, and T. Francart, “Relating EEG to continuous speech using deep neural networks: A review,”Journal of Neural Engineering, vol. 20, p. 041003, 2023

    work page 2023

  14. [14]

    The cocktail-party problem revisited: Early processing and selection of multi-talker speech,

    A. W. Bronkhorst, “The cocktail-party problem revisited: Early processing and selection of multi-talker speech,”Attention, Per- ception, and Psychophysics, vol. 77, no. 5, pp. 1465–1487, 2015

    work page 2015

  15. [15]

    EEG decoding of the target speaker in a cocktail party scenario: Considerations regarding dynamic switching of talker location,

    E. S. Teoh and E. C. Lalor, “EEG decoding of the target speaker in a cocktail party scenario: Considerations regarding dynamic switching of talker location,”Journal of Neural Engineering, vol. 16, no. 3, p. 036017, 2019

    work page 2019

  16. [16]

    What are we really decoding? Unveiling biases in EEG-based decoding of the spatial focus of auditory attention,

    I. Rotaru, S. Geirnaert, N. Heintz, I. Van de Ryck, A. Bertrand, and T. Francart, “What are we really decoding? Unveiling biases in EEG-based decoding of the spatial focus of auditory attention,” Journal of Neural Engineering, vol. 21, no. 1, p. 016017, 2024

    work page 2024

  17. [17]

    EEG-based decoding of auditory attention to conversations with turn-taking speakers,

    I. Van de Ryck, N. Heintz, I. Rotaru, S. Geirnaert, A. Bertrand, and T. Francart, “EEG-based decoding of auditory attention to conversations with turn-taking speakers,”Hearing Research, vol. 471, p. 109539, 2026

    work page 2026

  18. [18]

    Electroencephalography-based auditory attention decoding: To- ward neurosteered hearing devices,

    S. Geirnaert, S. Vandecappelle, E. Alickovic, A. De Cheveigne, E. Lalor, B. T. Meyer, S. Miran, T. Francart, and A. Bertrand, “Electroencephalography-based auditory attention decoding: To- ward neurosteered hearing devices,”IEEE Signal Processing Magazine, vol. 38, no. 4, pp. 89–102, 2021

    work page 2021

  19. [19]

    Impact of auditory attention decoding accuracy on noise reduction systems for hearing aids,

    L. S. Schiavon, J. C. M. Bermudez, and M. H. Costa, “Impact of auditory attention decoding accuracy on noise reduction systems for hearing aids,”Biomedical Signal Processing and Control, vol. 113, p. 109162, 2026

    work page 2026

  20. [20]

    Seeing helps hearing: A multi-modal dataset and a mamba-based dual branch parallel network for auditory attention decoding,

    C. Fan, H. Zhang, Q. Ni, J. Zhang, J. Tao, J. Zhou, J. Yi, Z. Lv, and X. Wu, “Seeing helps hearing: A multi-modal dataset and a mamba-based dual branch parallel network for auditory attention decoding,”Information Fusion, p. 102946, 2025

    work page 2025

  21. [21]

    ListenNet: A lightweight spatio-temporal enhancement nested network for auditory attention detection,

    C. Fan, X. Yang, H. Zhang, Y . Chen, L. Li, J. Zhou, and Z. Lv, “ListenNet: A lightweight spatio-temporal enhancement nested network for auditory attention detection,” inProceedings of the 2025 International Joint Conference on Artificial Intelligence (IJ- CAI), 2025, pp. 4137–4145

    work page 2025

  22. [22]

    EEG-based auditory attention decoding for speaker identification under mixed-speech hearing-assistive conditions,

    Y . Ding, L. Wang, J. Lu, Z. Lin, and F. Chen, “EEG-based auditory attention decoding for speaker identification under mixed-speech hearing-assistive conditions,”IEEE Transactions on Biomedical Engineering, 2025

    work page 2025

  23. [23]

    Buzs ´aki,Rhythms of the Brain

    G. Buzs ´aki,Rhythms of the Brain. Oxford University Press, 2006

    work page 2006

  24. [24]

    Alignment is quantified by the Pearson correlationr (i) t = corr(et,a (i) t )

    with a1×1convolutional projection layer. Alignment is quantified by the Pearson correlationr (i) t = corr(et,a (i) t ). The encoders are pretrained using a cross-entropy lossL match = CE([r(1) t , r(2) t ], yt), wherey t ∈ {1,2}denotes the ground- truth attended speaker. 2.2. Proposed SGAD framework Based on the raw correlation scoresr (i) t , a straightf...

    work page

  25. [25]

    Real-time control of a hearing instrument with EEG-based attention decoding,

    J. Hjortkjær, D. D. Wong, A. Catania, J. M ¨archer-Rørsted, E. Ce- olini, S. A. Fuglsang, I. Kiselev, G. Di Liberto, S.-C. Liu, T. Dau et al., “Real-time control of a hearing instrument with EEG-based attention decoding,”Journal of Neural Engineering, vol. 22, no. 1, p. 016027, 2025

    work page 2025

  26. [26]

    Enhancing spatial au- ditory attention decoding with wavelet-based prototype training,

    Z. Qiu, J. Gu, D. Yao, J. Li, and Y . Yan, “Enhancing spatial au- ditory attention decoding with wavelet-based prototype training,” Biomedical Signal Processing and Control, vol. 100, p. 107130, 2025

    work page 2025

  27. [27]

    Developing lightweight models with data optimization for attending speaker identity from EEG without spatial information,

    Y . Ding, L. Wang, X. Wang, and F. Chen, “Developing lightweight models with data optimization for attending speaker identity from EEG without spatial information,” inProceedings of the 2025 IEEE International Conference on Engineering in Medicine and Biology Society (EMBC), 2025, pp. 1–4

    work page 2025

  28. [28]

    Robust decoding of the audi- tory attention from EEG recordings through graph convolutional networks,

    S. Cai, R. Zhang, and H. Li, “Robust decoding of the audi- tory attention from EEG recordings through graph convolutional networks,” inProceedings of the 2024 IEEE International Con- ference on Acoustics, Speech, and Signal Processing (ICASSP), 2024, pp. 2320–2324

    work page 2024

  29. [29]

    SWIM: Short-window CNN integrated with mamba for EEG- based auditory spatial attention decoding,

    Z. Zhang, A. Thwaites, A. Woolgar, B. Moore, and C. Zhang, “SWIM: Short-window CNN integrated with mamba for EEG- based auditory spatial attention decoding,” inProceedings of the 2024 IEEE International Conference on Spoken Language Tech- nology Workshop (SLT), 2024, pp. 1031–1038

    work page 2024

  30. [30]

    Multi-class decoding of attended speaker direction using electroencephalo- gram and audio spatial spectrum,

    Y . Zhang, J. Lu, F. Chen, H. Du, X. Gao, and Z. Lin, “Multi-class decoding of attended speaker direction using electroencephalo- gram and audio spatial spectrum,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 33, pp. 2892–2903, 2025

    work page 2025

  31. [31]

    Con- trastive representation learning with Transformers for robust au- ditory EEG decoding,

    L. Bollens, B. Accou, H. Van Hamme, and T. Francart, “Con- trastive representation learning with Transformers for robust au- ditory EEG decoding,”Scientific Reports, vol. 15, no. 1, p. 28744, 2025

    work page 2025

  32. [32]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 12 449–12 460, 2020

    work page 2020

  33. [33]

    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,” Advances in Neural Information Processing Systems (NeurIPS), vol. 30, p. 5998–6008, 2017

    work page 2017

  34. [34]

    SAASD- Net: An EEG-based streaming auditory attention switch decoding network for self-initiated attention switching in mixed speech,

    Y . Ding, S. Yu, X. Chen, X. Wang, Y . Ban, and F. Chen, “SAASD- Net: An EEG-based streaming auditory attention switch decoding network for self-initiated attention switching in mixed speech,” inProceedings of the 2026 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2026, pp. 7016–7020

    work page 2026

  35. [35]

    MS-AASD: An EEG dataset for self-initiated auditory attention switch decoding in mixed speech,

    Y . Ding, S. Yu, X. Wang, Y . Ban, and F. Chen, “MS-AASD: An EEG dataset for self-initiated auditory attention switch decoding in mixed speech,”Zenodo, Sep. 2025. [Online]. Available: https://doi.org/10.5281/zenodo.17149387

    work page doi:10.5281/zenodo.17149387 2025

  36. [36]

    Sleep-spindle detection: Crowdsourcing and evaluat- ing performance of experts, non-experts and automated methods,

    S. C. Warby, S. L. Wendt, P. Welinder, E. G. S. Munk, O. Car- rillo, H. B. D. Sorensen, P. Jennum, P. E. Peppard, P. Perona, and E. Mignot, “Sleep-spindle detection: Crowdsourcing and evaluat- ing performance of experts, non-experts and automated methods,” Nature Methods, vol. 11, no. 4, pp. 385–392, 2014

    work page 2014

  37. [37]

    Objective evaluation metrics for automatic classification of EEG events,

    V . Shah, M. Golmohammadi, I. Obeid, and J. Picone, “Objective evaluation metrics for automatic classification of EEG events,” inBiomedical signal processing: innovation and applications. Springer, 2021, pp. 223–255

    work page 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.