Pith. sign in

REVIEW 3 major objections 6 minor 45 references

The 2025 PNPL Competition: Speech Detection and Phoneme Classification in the LibriBrain Dataset

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This competition paper presents the largest within-subject MEG dataset to date, over 50 hours, paired with two standardized speech-decoding tasks to spark an 'ImageNet moment' for non-invasive brain-computer interfaces.

desk verdict First standardized benchmark for non-invasive speech decoding on a 50+ hour MEG dataset, with a real infrastructure contribution; but the validity of the labels is unverified and the baselines lack detail—worth refereeing conditionally. read the letter →

arxiv 2506.10165 v1 pith:NSOSXHTK submitted 2025-06-11 cs.LG cs.SDeess.AS

Gilad Landau , Miran Özdogan , Gereon Elvers , Francesco Mantegna , Pratik Somaiya , Dulhan Jayalath , Luisa Kurth , Teyun Kwon
show 10 more authors
Brendan Shillingford Greg Farquhar Minqi Jiang Karim Jerbi Hamza Abdelhedi Yorguin Mantilla Ramos Caglar Gulcehre Mark Woolrich Natalie Voets Oiwi Parker Jones
This is my paper · ORCID
classification cs.LGcs.SDeess.AS
keywords NeuralDecodingNon-InvasiveMEGSpeechComprehensionLibriBrainDetectionPhonemeClassificationBrain-ComputerInterface
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 proposes the first competition and benchmark for language decoding from non-invasive brain recordings, built around LibriBrain, a magnetoencephalography (MEG) dataset of over 50 hours recorded from a single person listening to audiobooks. It defines two supervised tasks, speech detection and phoneme classification, with fixed data splits and an F1-macro evaluation metric, and supplies a Python library, tutorials, baseline models, and leaderboards so different groups can compare results on identical data. The authors' stated aim is to create the conditions for an 'ImageNet moment' in non-invasive neural decoding: enough deep, standardized data that machine learning methods can make non-invasive speech decoding competitive with surgical brain-computer interfaces. A sympathetic reader would take the paper's claim to be that this combination of scale and standardization is the missing ingredient for progress toward restoring communication to paralysed individuals without brain surgery.

What carries the argument

The load-bearing object is LibriBrain itself: more than 50 hours of 306-channel magnetoencephalography from a single subject listening to audiobooks, sampled at 250 Hz and aligned to word- and phoneme-level transcriptions inherited from the source speech corpus. Around this dataset, the paper's machinery is standardization: a Python library that downloads and batches the data into deep-learning-ready tensors, public train, validation, test, and competition holdout splits, the F1-macro metric, and reference models that establish baseline scores of 68.04% and 60.39%. The speech-detection task labels every temporal sample as speech or non-speech, making efficient use of each recording at 250 labels per second; the phoneme task uses more than 1.5 million examples across 39 classes, an order of magnitude more examples per class than word classification. Together these components turn a raw data release into a competition with comparable endpoints.

What would settle it

Permute the phoneme labels across training windows and retrain the reference model: if the F1-macro stays materially above chance, the model is exploiting a non-phoneme artifact; likewise, if shifting the speech labels by one 4 ms sample on a held-out subset does not drop the reference model's speech-detection score, the labels are too coarse or mistimed to pin the task to the neural signal.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that LibriBrain is the largest within-subject MEG dataset recorded to date, over 50 hours of 306-sensor data from one healthy participant listening to audiobooks, downsampled to 250 Hz, and that it is large enough to support a new kind of benchmark. The paper pairs the dataset with two foundational tasks that have direct analogues in automatic speech recognition: per-sample binary speech detection and fixed-window phoneme classification over the 39 standard phoneme classes. It reports reference models reaching F1-macro scores of 68.04% for speech detection and 60.39% for phoneme classification, compared with naive majority-class baselines of 45.30% and 0.47%, and argues that these tasks are far more label-efficient than word classification. The contribution is not a new model or algorithm; it is the assertion that a deep single-subject dataset plus standardized evaluation can convert a field where end-to-end non-invasive decoders have hovered at chance into a measurable, comparable research program.

Load-bearing premise

The whole benchmark rests on the assumption that the word and phoneme alignments inherited by LibriBrain are correct and that the audiobook audio stayed synchronized with the MEG clock; if those labels are wrong or drift, the reference scores and every leaderboard entry measure alignment error rather than neural decoding.

Editorial extensions

If this is right

  • Any research group can now evaluate a speech decoder on the same held-out MEG data and report an F1-macro score directly comparable with the published 68.04% and 60.39% baselines.
  • If the reference models are reproducible, the two tasks give the field sanity baselines: speech-detection systems should score well above 50%, and phoneme systems well above the 2.56% chance level, replacing near-chance end-to-end results reported for MEG-to-text.
  • The Standard track, restricted to LibriBrain data, tests whether algorithmic innovation alone can push decoding performance on a fixed corpus.
  • The Extended track, which permits any external data, provides a controlled test of whether pooling additional brain recordings improves decoding under the same evaluation metric.
  • A successful competition would make future non-invasive speech-decoding papers report comparable benchmark numbers, the same way speech recognition moved to shared evaluation corpora.

Reading between the lines

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

  • Because LibriBrain is a single participant, the benchmark's conclusions about decoding accuracy may not transfer to other brains; a natural next step would be a multi-speaker extension, which the paper's planned multi-year curriculum already gestures toward.
  • The label-correctness assumption could be tested cheaply: if a one-hour subset is re-aligned with an independent forced aligner and the reference model's scores change materially, the leaderboard would be measuring alignment quality as much as neural decoding.
  • The competition structure could also serve as a probe for scaling laws: if Extended-track entries with more compute consistently outscore Standard-track entries on the same holdout, that would be direct evidence that non-invasive speech decoding benefits from data and compute scaling rather than only architecture.
  • If the two foundational tasks prove solvable at the reported levels, it would suggest that the information needed for speech decoding is present in MEG and that the bottleneck for full text decoding lies in sequence-level decoding, not in the absence of a decodable signal.
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

3 major / 6 minor

Summary. The paper describes the 2025 PNPL competition, the first competition for language decoding from non-invasive brain data. It introduces LibriBrain, a single-subject MEG corpus of over 50 hours of audiobook-listening data, and the pnpl Python library for data access and integration with PyTorch. The competition defines two supervised tasks, Speech Detection (binary per-sample classification) and Phoneme Classification (39 CMU/ARPAbet classes), with fixed train/validation/test/competition-holdout splits and an F1-macro evaluation metric. Reference models achieve F1-macro scores of 68.04% for Speech Detection and 60.39% for Phoneme Classification, compared with naive baselines of 45.30% and 0.47%. The paper also details the competition tracks, rules, prizes, schedule, and community infrastructure.

Significance. If the label synchronization and baselines are trustworthy, this paper provides a valuable community resource: the largest within-subject MEG dataset for speech decoding, standardized tasks and splits, an accessible data-loading library, tutorials, a leaderboard, and a clear evaluation protocol. These infrastructure contributions are concrete and will likely accelerate comparable work across research groups. The central claim is not a mathematical derivation but an empirical benchmark, so the paper's value rests on the correctness of the dataset labels and the reproducibility of the reference scores.

major comments (3)
  1. [1.3 (Data) and Table 2] The benchmark's validity rests on the inherited LibriSpeech alignments being correctly synchronized with the MEG samples, but Section 1.3 provides no trigger description, no clock-drift measurement, and no alignment validation. A misalignment of even tens of milliseconds would systematically mislabel phoneme boundaries and short non-speech pauses, and then the reference scores in Table 2 and all leaderboard rankings could reflect temporal autocorrelation of labels rather than neural decoding. Please add synchronization evidence, for example: trigger timing reports, estimated audio-to-MEG clock drift, an audio-envelope-to-MEG cross-correlation or temporal-response-function analysis on held-out segments, or manual QA on a sample of phoneme boundaries. This is load-bearing because every supervised label and both reference scores depend on it.
  2. [1.6 (Baselines) and Table 2] The reference model scores are reported as single F1-macro numbers with no architecture description, training hyperparameters, or uncertainty estimates. Since Section 2.2 makes beating 68.04% for Speech Detection and 60.39% for Phoneme Classification a condition for prizes, these numbers must be reproducible and statistically credible. Please report the model architectures, training details, number of independent runs, and confidence intervals or standard deviations. In addition, clarify what "averaged over sets of 100 samples" means for the phoneme classification model: is the input a fixed 100-sample window, or are predictions averaged over 100 raw sample labels, or both?
  3. [Table 1 and Section 1.6] The public validation and test sets are very small relative to the training set (0.36 and 0.38 hours out of 51.57 hours). For sample-level Speech Detection this yields on the order of 3e5 labeled samples, but for Phoneme Classification, when labels are averaged over 100-sample windows, this reduces to roughly 3e3 windows across 39 classes, which can produce wide error bars on F1-macro. Since rankings and prize eligibility are based on these splits, please report per-class counts, bootstrap confidence intervals for the reference scores, and a brief analysis of whether the fixed small holdout is sufficient for reliable discrimination among submissions.
minor comments (6)
  1. [1.5 (Metrics)] Typo: "prepose" should be "propose" in the sentence "...which we prepose to do...".
  2. [1.4 (Tasks and application scenarios)] Typo: "attemps" should be "attempts" in the sentence "early attemps in EEG and MEG have yielded weak results".
  3. [2.3 (Schedule and readiness)] The date line "10 May 26 May, 2025" appears garbled; it should likely read "10 May 2025" and "26 May 2025" as separate events.
  4. [Organizing team] Typo: "researchs" should be "researchers" in the sentence "a strategic collaboration between researchs in Oxford, Montreal, Switzerland, Google, and Meta".
  5. [1.2 (Novelty) and References] The claim of the "largest within-subject MEG dataset recorded to date" is cited only to the companion preprint [42], which shares several authors with this paper. Please state the overlap explicitly and, if possible, provide independent or concurrently reviewed evidence for the scale comparison shown in Figure 1B.
  6. [1.3 and 1.4] Consider listing the 39 phoneme classes or citing the exact ARPAbet inventory, since the phoneme label set is a core element of the task. Also, define the "sets of 100 samples" windowing exactly and state whether windows are overlapping or disjoint.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper reports a new dataset, defines two supervised tasks, and gives reference-model baselines; no claim reduces to its own inputs by construction.

full rationale

This paper is a competition/benchmark proposal rather than a derivation. The central claims are (i) the existence and scale of LibriBrain, (ii) the definition of two supervised tasks with fixed splits, and (iii) reference baseline scores. None of these is derived from an assumed input in a way that would constitute circularity. The 68.04% and 60.39% reference F1 scores (Table 2, Section 1.6) are presented as trained reference models; they are empirical benchmark results, not fitted parameters renamed as predictions. The dataset size claim is supported internally by Table 1 (51.57 train hours plus validation/test splits) and by the companion dataset paper [42]; citing the companion paper is a normal self-citation, and although the 'largest within-subject MEG dataset recorded to date' superlative relies partly on the authors' own reporting, it is an empirical assertion backed by the paper's own duration table and comparison to prior datasets, not a mathematical consequence that reduces to the citation. The Extended-track motivation cites the organizers' own prior work [17, 18, 31] on pooling MEG data, but that citation is used to justify a competition design choice (permitting external training data), not to establish a predicted result. No uniqueness theorem, ansatz, or fitted input is imported from self-citations. The unvalidated LibriSpeech-alignment/synchronization assumption flagged in review is a correctness/validity risk for the benchmark, but it is not circularity: the labels are treated as ground-truth inputs to the supervised tasks, not as quantities derived from MEG predictions or from the competition design. Overall, no step in the paper's reasoning chain reduces by construction to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical or scientific entities are postulated; the two-track competition structure is organizational scaffolding, not an entity requiring independent evidence. The load-bearing inputs are the dataset from companion paper [42] and the cited scaling assumptions, both of which carry self-citation weight. The reference baseline scores are the main fitted numbers the competition's prize rule depends on.

free parameters (2)
  • Reference Speech Detection F1-macro = 68.04%
    Reported performance of the provided reference model on the competition evaluation, used as the threshold teams must beat to win prizes. No error bars, variance, or training details are given, so it functions as a single fitted number.
  • Reference Phoneme Classification F1-macro = 60.39%
    Same role for the phoneme task, computed by averaging predictions over sets of 100 samples; the evaluation procedure is described in one sentence and cannot be reproduced from the paper.
assumptions (4)
  • domain assumption Deep, within-subject data yields the largest gains in decoding performance
    Stated in Section 1.1 and cited to reference [10]. The entire rationale for building a 50+ hour single-subject dataset and the competition's scaling bet rests on this premise.
  • domain assumption Word- and phoneme-level alignments from LibriSpeech apply to the MEG recordings
    Section 1.3 states LibriBrain 'comes with word- and phoneme-level alignments, as in the LibriSpeech corpus'. Both supervised tasks depend on these labels being correct and synchronized with the neural signal.
  • domain assumption MEG can approach or exceed invasive precision for speech decoding under ideal conditions
    Section 1.1: 'MEG recordings can therefore approach, or, under ideal conditions, even exceed the precision of invasive modalities'. Stated as a working hypothesis, not demonstrated; it motivates the competition's ambition.
  • domain assumption F1-macro with balanced-class chance assumptions is the right evaluation
    Section 1.5 defines the metric and the chance-level interpretations (50% for balanced speech detection, 2.56% for balanced phoneme classification). The metric choice shapes the leaderboard but is not argued against alternatives.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The 2025 PNPL Competition: Speech Detection and Phoneme Classification in the LibriBrain Dataset." pith.science (2026). https://pith.science/paper/NSOSXHTK

@misc{pith2026250610165,
  author       = {Pith},
  title        = {Pith review of: The 2025 PNPL Competition: Speech Detection and Phoneme Classification in the LibriBrain Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NSOSXHTK}},
  note         = {Machine review of arXiv:2506.10165}
}
read the original abstract

The advance of speech decoding from non-invasive brain data holds the potential for profound societal impact. Among its most promising applications is the restoration of communication to paralysed individuals affected by speech deficits such as dysarthria, without the need for high-risk surgical interventions. The ultimate aim of the 2025 PNPL competition is to produce the conditions for an "ImageNet moment" or breakthrough in non-invasive neural decoding, by harnessing the collective power of the machine learning community. To facilitate this vision we present the largest within-subject MEG dataset recorded to date (LibriBrain) together with a user-friendly Python library (pnpl) for easy data access and integration with deep learning frameworks. For the competition we define two foundational tasks (i.e. Speech Detection and Phoneme Classification from brain data), complete with standardised data splits and evaluation metrics, illustrative benchmark models, online tutorial code, a community discussion board, and public leaderboard for submissions. To promote accessibility and participation the competition features a Standard track that emphasises algorithmic innovation, as well as an Extended track that is expected to reward larger-scale computing, accelerating progress toward a non-invasive brain-computer interface for speech.

Figures

Figures reproduced from arXiv: 2506.10165 by the authors.

Figure 1
Figure 1. Overview of data and labels. (A) Illustration of a non-invasive MEG scanner and corre [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 32 canonical work pages

  1. [1]

    Accou et al

    B. Accou et al. SparrKULee: A speech-evoked auditory response repository of the KU Leuven, containing EEG of 85 participants. bioRxiv, 2023. URL https://doi.org/10.1101/2023. 07.24.550310

    doi:10.1101/2023 2023

  2. [2]

    Armeni, U

    K. Armeni, U. Güçlü, M. van Gerven, and J.-M. Schoffelen. A 10-hour within-participant magnetoencephalography narrative dataset to test models of language comprehension.Scientific Data, 9:278, 2022

    work page 2022

  3. [3]

    E. L. Barratt, S. T. Francis, P. G. Morris, and M. J. Brookes. Mapping the topological or- ganisation of beta oscillations in motor cortex using MEG. NeuroImage, 181:831–844, 2018. doi: 10.1016/j.neuroimage.2018.06.041. URL https://doi.org/10.1016/j.neuroimage. 2018.06.041

    work page doi:10.1016/j.neuroimage.2018.06.041 2018

  4. [4]

    Benesty, M

    J. Benesty, M. Sondhi, and Y . Huang. Speech enhancement. Springer, 2007

    work page 2007

  5. [5]

    T. M. Blakely, K. J. Miller, R. P. N. Rao, M. D. Holmes, and J. G. Ojemann. Localization and classification of phonemes using high spatial resolution electrocorticography (ECoG) grids. In 2008 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pages 4964–4967. IEEE, 2008. doi: 10.1109/IEMBS.2008.4650328. URL https://d...

    work page arXiv 2008

  6. [6]

    M. P. Broderick, A. J. Anderson, G. M. Di Liberto, M. J. Crosse, and E. C. Lalor. Electrophys- iological correlates of semantic dissimilarity reflect the comprehension of natural, narrative speech. Current Biology, 28:803–809, 2018

    work page 2018

  7. [7]

    J. S. Brumberg, E. J. Wright, D. S. Andreasen, F. H. Guenther, and P. R. Kennedy. Classification of intended phoneme production from chronic intracortical microelectrode recordings in speech- motor cortex. Frontiers in Neuroscience, 5:65, May 2011. doi: 10.3389/fnins.2011.00065. URL https://doi.org/10.3389/fnins.2011.00065

    work page Pith review arXiv 2011

  8. [8]

    N. S. Card, M. Wairagkar, C. Iacobacci, X. Hou, T. Singer-Clark, F. R. Willett, E. M. Kunz, and D. M. Brandman. An accurate and rapidly calibrating speech neuroprosthesis. New England Journal of Medicine, 391(7):609–618, 2024

    work page 2024

Show all 45 references
  1. [9]

    The CMU pronouncing dictionary, 1993

    Carnegie Mellon University. The CMU pronouncing dictionary, 1993. http://www.speech. cs.cmu.edu/cgi-bin/cmudict

    1993

  2. [10]

    d’Ascoli, C

    S. d’Ascoli, C. Bel, J. Rapin, H. Banville, Y . Benchetrit, C. Pallier, and J.-R. King. Decoding individual words from non-invasive brain recordings across 723 participants. arXiv preprint,

  3. [11]

    Défossez, C

    A. Défossez, C. Caucheteux, J. Rapin, O. Kabeli, and J.-R. King. Decoding speech per- ception from non-invasive brain recordings. Nature Machine Intelligence, 5(10):1097–1107,

  4. [12]

    S. Furui. Speech recognition: Technology and applications. Academic Press, 2000

    2000

  5. [13]

    J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. TIMIT acoustic-phonetic continuous speech corpus. Linguistic Data Consortium, 1993. https: //catalog.ldc.upenn.edu/LDC93S1

    1993

  6. [14]

    J. J. Godfrey, E. C. Holliman, and J. McDaniel. Switchboard: Telephone speech corpus for research and development. In Proc. IEEE ICASSP, pages 517–520, 1992

    1992

  7. [15]

    F. H. Guenther, J. S. Brumberg, E. J. Wright, A. Nieto-Castanon, J. A. Tourville, M. Panko, R. Law, S. A. Siebert, J. L. Bartels, D. S. Andreasen, P. Ehirim, H. Mao, and P. R. Kennedy. 10 A wireless brain-machine interface for real-time speech synthesis. PLoS ONE, 4(12):e8218,...

    2009 doi

  8. [16]

    Gwilliams, G

    L. Gwilliams, G. Flick, A. Marantz, L. Pylkkanen, D. Poeppel, and J.-R. King. Introducing MEG-MASC: A high-quality magneto-encephalography dataset for evaluating natural speech processing. Scientific Data, 10(1):862, 2023. doi: 10.1038/s41597-023-02752-5. URL https://www.natur...

    2023 doi

  9. [17]

    Jayalath, G

    D. Jayalath, G. Landau, and O. Parker Jones. Unlocking non-invasive brain-to-text. arXiv preprint arXiv:2505.13446, 2025

    2025 arXiv

  10. [18]

    Jayalath, G

    D. Jayalath, G. Landau, B. Shillingford, M. Woolrich, and O. Parker Jones. The Brain’s Bitter Lesson: Scaling speech decoding with self-supervised learning. Forty-second International Conference on Machine Learning, ICML, 2025

    2025

  11. [19]

    H. Jo, Y . Yang, J. Han, Y . Duan, H. Xiong, and W. H. Lee. Are EEG-to-text models working? arXiv preprint, 2024. URL https://arxiv.org/abs/2405.06459

    2024 arXiv

  12. [20]

    Krizhevsky

    A. Krizhevsky. Learning multiple layers of features from tiny images. Technical Report TR-2009, University of Toronto, 2009

    2009

  13. [21]

    Lee and H.-W

    K.-F. Lee and H.-W. Hon. Speaker-independent phone recognition using hidden markov models. IEEE Transactions on Acoustics, Speech, and Signal Processing, 37(11):1641–1648, 1989

    1989

  14. [22]

    E. C. Leuthardt, C. Gaona, M. Sharma, N. Szrama, J. Roland, Z. Freudenberg, J. Solis, J. Bres- hears, and G. Schalk. Using the electrocorticographic speech network to control a brain- computer interface in humans. Journal of Neural Engineering, 8(3):036004, Jun 2011. doi: 10. ...

    2011 doi

  15. [23]

    Librivox: Free public domain audiobooks, 2024.https://librivox

    LibriV ox V olunteers. Librivox: Free public domain audiobooks, 2024.https://librivox. org

    2024

  16. [24]

    S. L. Metzger, K. T. Littlejohn, A. B. Silva, D. A. Moses, M. P. Seaton, R. Wang, M. E. Dougherty, J. R. Liu, P. Wu, M. A. Berger, I. Zhuravleva, A. Tu-Chan, K. Ganguly, G. K. Anumanchipalli, and E. F. Chang. A high-performance neuroprosthesis for speech decoding and avatar co...

    2023

  17. [25]

    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, A. Tu-Chan, K. Ganguly, and E. F. Chang. Neuroprosthesis for decoding speech in a paralyzed person with anarthria. New England Journal of...

    2021 doi

  18. [26]

    M. S. Nieuwland et al. Large-scale replication study reveals a limit on probabilistic prediction in language comprehension. eLife, Apr. 2018

    2018

  19. [27]

    Panayotov, G

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an ASR corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206–5210. IEEE, 2015

    2015

  20. [28]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Te- jani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An imperative style, high-p...

    2019

  21. [29]

    X. Pei, D. L. Barbour, E. C. Leuthardt, and G. Schalk. Decoding vowels and consonants in spoken and imagined words using electrocorticographic signals in humans. Journal of Neural Engineering, 8:046028, 2011

    2011

  22. [30]

    L. R. Rabiner and B.-H. Juang. Fundamentals of Speech Recognition. Prentice-Hall, 1993

    1993

  23. [31]

    Ridge and O

    J. Ridge and O. Parker Jones. Resolving domain shift for representations of speech in non- invasive brain recordings. arXiv preprint, 2024. URL https://arxiv.org/abs/2410. 19986

    2024

  24. [32]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211–252, 2015. 11

    2015

  25. [33]

    Schoffelen et al

    J.-M. Schoffelen et al. A 204-subject multimodal neuroimaging dataset to study language processing. Scientific Data, 6:1–13, Apr. 2019

    2019

  26. [34]

    M. W. J. van Es, C. Gohil, A. J. Quinn, and M. W. Woolrich. osl-ephys: a python toolbox for the analysis of electrophysiology data. Frontiers in Neuroscience, 19:1522675, Feb. 2025. doi: 10.3389/fnins.2025.1522675

    2025

  27. [35]

    F. R. Willett, E. M. Kunz, C. Fan, D. T. Avansino, G. H. Wilson, E. Y . Choi, F. Kam- dar, M. F. Glasser, L. R. Hochberg, S. Druckmann, K. V . Shenoy, and J. M. Hender- son. A high-performance speech neuroprosthesis. Nature, 620:1031–1036, 2023. doi: 10.1038/s41586-023-06377-x

    2023 doi

  28. [36]

    F. R. Willett, J. Li, T. Le, C. Fan, M. Chen, E. Shlizerman, Y . Chen, X. Zheng, T. S. Okubo, T. Benster, H. D. Lee, M. Kounga, E. K. Buchanan, D. Zoltowski, S. W. Linderman, and J. M. Henderson. Brain-to-Text Benchmark ’24: Lessons learned. arXiv preprint arXiv:2412.17227,

    arXiv

  29. [37]

    G. H. Wilson, S. D. Stavisky, F. R. Willett, D. T. Avansino, J. N. Kelemen, L. R. Hochberg, J. M. Henderson, S. Druckmann, and K. V . Shenoy. Decoding spoken English from intracortical electrode arrays in dorsal precentral gyrus. Journal of Neural Engineering, 17(6):066007, 20...

    2020 doi

  30. [38]

    Xiong, J

    W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Stolcke, D. Yu, and G. Zweig. Achieving human parity in conversational speech recognition. arXiv preprint arXiv:1610.05256, 2016. URL https://arxiv.org/abs/1610.05256

    2016 arXiv

  31. [39]

    URL https://doi.org/10.48550/arXiv.2412.17227

    doi arXiv

  32. [40]

    Y . Yang, Y . Duan, Q. Zhang, R. Xu, and H. Xiong. NeuSpeech: Decode neural signal as speech. arXiv preprint, 2024. URL https://arxiv.org/abs/2403.01748

    2024 arXiv

  33. [41]

    Y . Yang, H. Jo, Y . Duan, Q. Zhang, J. Zhou, W. H. Lee, R. Xu, and H. Xiong. MAD: Multi- alignment MEG-to-text decoding. arXiv preprint, 2024. URL https://arxiv.org/abs/ 2406.01512

    2024

  34. [42]

    Y . Yang, Y . Duan, H. Jo, Q. Zhang, R. Xu, O. Parker Jones, X. Hu, C.-T. Lin, and H. Xiong. NeuGPT: Unified multi-modal neural GPT. arXiv preprint, 2024. URL https://arxiv.org/ abs/2410.20916

    2024 arXiv

  35. [45]

    Özdogan, G

    M. Özdogan, G. Landau, G. Elvers, D. Jayalath, P. Somaiya, F. Mantegna, M. Woolrich, and O. Parker Jones. LibriBrain: Over 50 hours of within-subject MEG to improve speech decoding methods at scale. arXiv preprint, 2025. URL https://arxiv.org/abs/2506.02098. 12

    2025 arXiv

  36. [2023]

    URL https://www.nature.com/articles/ s42256-023-00714-5

    doi: 10.1038/s42256-023-00714-5. URL https://www.nature.com/articles/ s42256-023-00714-5

    doi

  37. [2024]

    URL https://arxiv.org/abs/2412.17829

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.