Pith. sign in

REVIEW 2 major objections 6 minor 57 references

Phoneme- vs. Character-Level Targets and Selective State-Space Models for Intracortical Brain-to-Text

T0 review · 2 major / 6 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read On matched intracortical brain-to-text runs, a phonetic GRU still beats a Mamba hybrid, and phonemes beat direct characters on final word error.

desk verdict Honest negative-leaning 2×2 on Brain-to-Text ’25: phonetic GRU still beats a ConvMambaGRU hybrid, and the comparison is clean enough to trust within its single-participant scope. read the letter →

arxiv 2607.26751 v1 pith:IB34XJ7W submitted 2026-07-29 cs.CL cs.AIeess.SP

Lucas Zamora Vera , Jose A. Gonzalez-Lopez This is my paper · ORCID
classification cs.CLcs.AIeess.SP
keywords brain-computerinterfacespeechneuroprosthesisintracorticaldecodingstate-spacemodelsbrain-to-textCTCphonemevscharactertargetsMamba
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

This paper asks whether newer selective state-space decoders (Mamba hybrids) improve on ordinary recurrent GRUs for turning implanted brain signals into text, and whether decoding phonemes or characters is the better target. Under one fixed CTC training pipeline on the public Brain-to-Text ’25 validation set, the answer is no on both counts that matter for accuracy: the best phonetic GRU reaches 12.62% phoneme error and 21.19% word error, beating the strongest Mamba hybrid and also beating character-level decoding after language-model rescoring. Ablations show Mamba only becomes competitive once convolution and a final GRU are stacked on it, and error patterns split cleanly by target—articulatory phone swaps versus lexical and word-boundary mistakes. A sympathetic reader cares because clinical speech neuroprostheses still live in a data-limited, single-participant regime where architecture swaps are expensive; this study says keep the recurrent backbone and the phonetic two-stage path unless stronger evidence appears.

What carries the argument

A controlled 2×2 grid under one CTC pipeline: sequential backbone (unidirectional GRU vs ConvMambaGRU hybrid) crossed with output target (ARPAbet phonemes vs characters), with shared session adaptation, temporal patching, beam search, and optional LM rescoring so differences isolate backbone and target.

What would settle it

On the withheld public test set, or on a second participant with the same protocol, a ConvMambaGRU (or bare Mamba) would post lower validation-matched WER than the phonetic GRU, or a character decoder would beat the phonetic two-stage path after identical LM rescoring.

Watch

Extended reading notes

Core claim

Under one reproducible CTC protocol on Brain-to-Text ’25 validation, the recurrent phonetic GRU remains the strongest decoder: best-run 12.62% PER and 21.19% WER, significantly better than the best ConvMambaGRU hybrid (13.13% PER, 24.11% WER), while phonetic two-stage decoding also beats the best textual GRU after LM rescoring (13.39% CER, 26.28% WER). Selective state-space blocks alone do not help; only a Conv+Mamba+GRU stack approaches the baseline, and representation choice governs failure mode—articulatory confusions for phonemes versus short-word and boundary errors for characters.

Load-bearing premise

That rankings on one ALS participant’s validation split—with multi-seed runs only for phonetic models and single runs for textual ones—are enough to say which decoder architecture generalizes for intracortical brain-to-text.

Editorial extensions

If this is right

  • In this data-limited intracortical setting, replacing a GRU backbone with selective SSMs is unlikely to cut word error by itself.
  • Phonetic two-stage decoding remains the lower-WER path even when direct character decoding is viable after large LM rescoring.
  • Mamba-style cores need local convolution and recurrent refinement before they approach recurrent baselines on these signals.
  • Error audits should be target-specific: expect voicing and near-vowel phone swaps for phoneme systems, and function-word or space errors for character systems.
  • Gains will more likely come from training, ensembling, and rescoring than from swapping the sequential core.

Reading between the lines

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

  • If multi-participant or much larger day-count corpora reverse the GRU–Mamba ranking, the paper’s ‘data-limited regime’ explanation becomes the main scientific claim rather than a caveat.
  • A joint phoneme–character multi-task head could test whether articulatory structure still helps when the final output is orthographic.
  • The large non-linear jump from sub-point PER gaps to multi-point WER gaps implies future papers should report paired utterance-level LM ablation, not only best-α WER.
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

2 major / 6 minor

Summary. The paper reports a controlled 2×2 comparison on the public Brain-to-Text ’25 validation set (single ALS participant, 256 intracortical channels) crossing decoder backbone (GRU vs. a ConvMambaGRU hybrid) with output target (ARPAbet phonemes vs. characters), all under a shared CTC pipeline with session adaptation, temporal patching, and optional LM rescoring. Multi-seed phonetic results (Table 2) and paired Wilcoxon tests on the same 1,426 trials show the GRU baseline strongest (best 12.62% PER / 21.19% WER) over ConvMambaGRU (13.13% PER / 24.11% WER); bare Mamba and ConvMamba lag until conv front-end plus GRU refinement are added. Textual GRU after GPT-2 XL rescoring reaches 13.39% CER / 26.28% WER. Error analysis contrasts articulatory phoneme substitutions (Table 5) with lexical/function-word confusions (Table 6). The authors conclude that recurrent decoders remain hard to beat in this data-limited regime and that target choice shapes failure modes.

Significance. The contribution is a careful negative/comparative result rather than a new SOTA system. That is valuable for intracortical brain-to-text: the field has repeatedly found that training, ensembling, and LM rescoring dominate backbone swaps (Brain-to-Text ’24), and this paper tests the natural next candidate (selective SSMs) under a matched protocol with ablations, multi-seed phonetic statistics, paired tests, LM α sweeps, and representation-aware error tables. Explicit credit is due for the reproducible 2×2 design, the isolation of conv front-end vs. recurrent refinement in the SSM stack (Table 2), the Wilcoxon reporting on identical trials, and the concrete substitution inventories. If the rankings hold, they give practical architecture guidance and a clear error typology for phoneme vs. character targets. Scope is inherently limited to one participant and validation-only evaluation (test transcripts private), which the manuscript states openly.

major comments (2)
  1. [§4.2, Table 3, Table 4] §4.2 and Table 3: the claim that phonetic two-stage decoding outperforms textual decoding on final WER (21.19% vs. 26.28%) is not fully controlled on the linguistic stage. Phonetic WER uses the official 1-gram OpenWebText LM; textual WER uses GPT-2 XL with tuned α (Table 4). Because LM strength and role differ (phoneme-to-text conversion vs. character hypothesis rescoring), the cross-target WER gap cannot be attributed cleanly to the neural target. Please either (i) report a matched linguistic-stage comparison where feasible (e.g., same LM family / comparable effective strength, or neural-only CER/PER plus a clearly separated LM ablation), or (ii) explicitly qualify the phonetic-vs-textual WER claim as joint neural+LM performance under the benchmark’s standard phonetic pipeline vs. a stronger neural LM for text, and lead with CER/PER for the target comparison.
  2. [§3, Table 3] §3 and Table 3: phonetic backbones are evaluated over five seeds (Table 2) with mean±std and paired tests, but textual GRU and ConvMambaGRU are single-run. The 2×2 design and the abstract claim that “the recurrent baseline remains strongest” in both representations therefore rest on unequal statistical support. At minimum, add multi-seed textual runs (or bootstrap/utterance-level paired tests on the existing logits) for the two main textual systems so architecture rank under the character target has comparable uncertainty, or clearly demote textual architecture claims to exploratory.
minor comments (6)
  1. [Table 1] Table 1: ConvMambaGRU parameter counts differ slightly between phonetic (50.7M) and textual (54.2M) while GRU/Mamba/ConvMamba are listed as identical across targets; briefly explain the textual parameter increase (vocabulary / projection size).
  2. [Table 2] §2.4 / Table 2: MambaGRU shows one diverging run (22.41% PER) that inflates variance; a short note on training stability (learning-rate, gradient clipping, or seed exclusion policy) would help readers interpret that row.
  3. [Figure 1] Figure 1 is described but the caption alone does not specify patch size, number of sessions in the adaptation layer, or whether beam search is always on at inference; a one-line methods pointer in the caption would improve stand-alone readability.
  4. [§4.4, Table 5] §4.4 / Table 5: phoneme substitution counts are aggregated over five GRU runs without normalizing by opportunity (token frequency); adding rate-normalized confusions or a confusion-matrix figure would strengthen the articulatory-interpretation claim.
  5. [Introduction, §4.4] Typos / wording: “Abrain-to-textsystem” (Introduction); “V owel” with stray space (§4.4); arXiv id and venue formatting can be cleaned in camera-ready.
  6. [§3] Code/checkpoints are promised “with the final version”; for reproducibility of the multi-seed and α-sweep results, please confirm release of configs and validation prediction dumps as stated in §3.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical 2×2 architecture comparison on an external benchmark; rankings are measured, not forced by construction.

full rationale

The paper’s load-bearing claims are empirical outcomes of a matched CTC protocol on the public Brain-to-Text ’25 validation set (GRU vs. ConvMambaGRU × phonetic vs. character). PER/CER/WER are standard edit-distance metrics computed on held-out trials; they are not defined in terms of the architectures under test, nor recovered from fitted constants. Multi-seed phonetic PERs, paired Wilcoxon tests on the same 1,426 trials, and disclosed LM α sweeps (Table 4) are ordinary experimental practice—α is a rescoring hyperparameter chosen after the neural decoder is trained and does not construct the backbone PER/CER numbers that drive the architecture ranking. Citations to Brain-to-Text ’24 and prior neuroprosthesis work supply context and baselines, not uniqueness theorems or ansatzes that force the result. There is no self-definitional loop, no fitted-input-as-prediction, and no load-bearing self-citation chain. The derivation chain is simply: train decoders → decode → measure errors. Score 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The claim rests on standard sequence-modeling and BCI assumptions plus a few study-design choices, not on new physical entities. Load-bearing pieces are the public single-participant dataset, CTC as the right alignment-free objective, validation performance as a proxy for architecture quality, and the particular hybrid stacking (Conv1D+Mamba+GRU) as a fair SSM representative. Free parameters are ordinary ML knobs (LM α, depth, seeds) rather than physics-style fitted constants.

free parameters (4)
  • LM rescoring weight α = 1.5 (textual GRU + GPT-2 XL); 2.0 (textual ConvMambaGRU + GPT-2 XL); other pairs 0.8–7.5
    Chosen per decoder–LM pair on validation WER (Table 4); directly affects reported textual WER and thus the phonetic-vs-textual comparison.
  • Beam width and hypothesis pool for textual decoding = beam width 100, top 100
    Fixed to beam 100 and top-100 hypotheses for LM rescoring (§2.5); influences CER/WER after rescoring.
  • Network depth and hybrid stacking = 5 layers; ConvMambaGRU = Conv1D + 5×Mamba + GRU
    5× GRU or 5× Mamba blocks, optional Conv1D front-end and final GRU (Table 1); defines which model is called the 'SSM hybrid' representative.
  • Random seeds for phonetic multi-run table = 5 seeds (phonetic); single run (textual)
    Five seeds used for Table 2 means; best run carried into main WER table, which slightly favors reporting of peak rather than mean performance.
assumptions (5)
  • domain assumption CTC correctly handles unaligned neural-to-symbol sequences for both phoneme and character targets.
    Invoked in §2.3 as the sole training objective for all cells of the 2×2; standard in the cited brain-to-text literature.
  • domain assumption Ventral sensorimotor intracortical signals retain articulatory information usable for speech decoding in this ALS participant.
    Background premise in §1 citing [1,15–17]; without it the benchmark task is undefined.
  • ad hoc to paper Official validation split (1,426 trials) is an adequate surrogate for comparing architectures when test transcripts are private.
    Stated in §2.2 and §3; all reported PER/CER/WER and Wilcoxon tests use validation only.
  • ad hoc to paper ConvMambaGRU is a fair representative of selective SSM decoders for this comparison.
    Chosen after Table 2 ablations as the strongest SSM variant; conclusions about 'Mamba' not surpassing GRU depend on this choice.
  • standard math Wilcoxon signed-rank on paired utterance errors is an appropriate significance test for decoder comparison.
    Applied in §3 and §5 following [49,50].
invented entities (1)
  • ConvMambaGRU hybrid backbone
    purpose: Combine local Conv1D context, selective SSM long-range modeling, and GRU temporal refinement as the paper's proposed SSM-side decoder.
    Introduced in §2.4 as 'our proposed variant'; it is an engineering stack of existing blocks, not a new physical entity. Independent evidence outside this paper is limited to related ConvMamba ASR work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phoneme- vs. Character-Level Targets and Selective State-Space Models for Intracortical Brain-to-Text." pith.science (2026). https://pith.science/paper/IB34XJ7W

@misc{pith2026260726751,
  author       = {Pith},
  title        = {Pith review of: Phoneme- vs. Character-Level Targets and Selective State-Space Models for Intracortical Brain-to-Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IB34XJ7W}},
  note         = {Machine review of arXiv:2607.26751}
}
read the original abstract

State-of-the-art intracortical brain-to-text systems pair a neural-sequence phone decoder with an external language model. Two design axes remain underexplored: whether selective state-space models (Mamba) improve on recurrent decoders, and how the output target (phonetic vs.\ character) interacts with that choice. On the public Brain-to-Text '25 benchmark, we study a controlled 2x2 grid (GRU vs.\ hybrid Mamba decoder; phonetic vs.\ character targets) trained with a CTC objective under one reproducible protocol. The recurrent baseline remains strongest: the best phonetic GRU reaches 12.62\% PER and 21.19\% WER, while the best textual GRU after LM rescoring reaches 13.39\% CER and 26.28\% WER. The Mamba hybrid is competitive but does not surpass it. Ablations isolate architectural contributions, and error analysis shows representation-dependent failures: articulatory-like phoneme confusions vs.\ lexical and word-boundary errors.

Figures

Figures reproduced from arXiv: 2607.26751 by the authors.

Figure 1
Figure 1. Overview of the experimental pipeline. Intracortical neural activity is preprocessed and passed to a neural decoder. We study two factors: the sequential backbone (GRU, Mamba variants and ConvMambaGRU) and the target representation (phonetic vs. textual). The resulting sequences are decoded with CTC and optionally refined through beam search and language￾model rescoring to obtain the final sentence-level transcripti… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 6 linked inside Pith

  1. [1]

    Introduction People with amyotrophic lateral sclerosis (ALS), brainstem stroke or locked-in syndrome can lose intelligible speech while retaining cognition, and the recovery of a reliable communica- tion channel substantially improves autonomy and social par- ticipation [1, 2, 3]. Brain-computer interfaces (BCIs) aim to re- store this channel by translati...

    arXiv 2026

  2. [2]

    General procedure Fig

    Methods 2.1. General procedure Fig. 1 summarizes the pipeline shared by all configurations. A trial of intracortical neural features is first passed through a per-session adaptation layer that normalizes day-to-day distri- bution shifts, then through temporal patching that shortens the effective sequence. A sequential core (the recurrent baseline or one o...

  3. [3]

    Metrics.Phoneme Error Rate (PER) for the phonetic de- coder; Character Error Rate (CER) and Word Error Rate (WER) for the final text are used to evaluate the proposed systems

    Experimental setup The dataset, splits and signal processing are described in Sec- tion 2.2; all results are reported on the1,426-trial validation set. Metrics.Phoneme Error Rate (PER) for the phonetic de- coder; Character Error Rate (CER) and Word Error Rate (WER) for the final text are used to evaluate the proposed systems. Implementation.Models are imp...

  4. [4]

    Results 4.1. Backbone performance on the phonetic task We first explore how the sequential backbones behave on the phonetic task, using validation PER over five independent runs (Table 2) to identify the strongest SSM variant to carry for- ward. The GRU baseline obtains the lowest mean PER and is the most stable model. Among the SSM backbones, perfor- man...

  5. [5]

    textual) with the decoder architecture (recurrent vs

    Conclusion We presented a controlled study of intracortical brain-to-text de- coding crossing the target representation (phonetic vs. textual) with the decoder architecture (recurrent vs. a selective state- space hybrid) under one reproducible CTC pipeline. The re- current baseline remains the strongest decoder, and the pho- netic two-stage variant gives ...

  6. [6]

    Acknowledgments This work was supported by grants PID2022-141378OB- C22 and AIA2025-163317-C32 funded by MI- CIU/AEI/10.13039/501100011033 and ERDF/EU

    doi:10.13039/501100011033

  7. [7]

    These tools were not used to generate scientific con- tent, design the methodology, or interpret the results

    Generative AI Use Disclosure During the preparation of this work, the authors used genera- tive AI tools for language editing and to assist with code devel- opment. These tools were not used to generate scientific con- tent, design the methodology, or interpret the results. After us- ing these tools, the authors reviewed and edited the content as needed a...

  8. [8]

    The speech neuroprosthesis,

    A. B. Silva, K. T. Littlejohn, J. R. Liu, D. A. Moses, and E. F. Chang, “The speech neuroprosthesis,”Nature Reviews Neuro- science, vol. 25, no. 7, pp. 473–492, 2024

    2024

Show all 57 references
  1. [9]

    Brain-computer interfaces for communica- tion and control,

    J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G. Pfurtscheller, and T. M. Vaughan, “Brain-computer interfaces for communica- tion and control,”Clinical Neurophysiology, vol. 113, no. 6, pp. 767–791, 2002

    2002

  2. [10]

    Brain-computer interfaces for restoring communi- cation,

    E. F. Chang, “Brain-computer interfaces for restoring communi- cation,”New England Journal of Medicine, vol. 391, no. 7, pp. 654–657, 2024

    2024

  3. [11]

    Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,

    L. R. Hochberg, D. Bacher, B. Jarosiewicz, N. Y . Masse, J. D. Simeral, J. V ogel, S. Haddadin, J. Liu, S. S. Cash, P. van der Smagt, and J. P. Donoghue, “Reach and grasp by people with tetraplegia using a neurally controlled robotic arm,”Nature, vol. 485, no. 7398, pp. 372–375, 2012

    2012

  4. [12]

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

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

    2017

  5. [13]

    Fully implanted brain-computer interface in a locked-in patient with ALS,

    M. J. Vansteensel, E. G. M. Pels, M. G. Bleichner, M. P. Branco, T. Denison, Z. V . Freudenburg, P. Gosselaar, S. Leinders, T. H. Ottens, M. A. van den Boom, P. C. van Rijen, E. J. Aarnoutse, and N. F. Ramsey, “Fully implanted brain-computer interface in a locked-in patient wi...

  6. [14]

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

    J. Tang, A. LeBel, S. Jain, and A. G. Huth, “Semantic reconstruc- tion of continuous language from non-invasive brain recordings,” Nature Neuroscience, vol. 26, no. 5, pp. 858–866, 2023

    2023

  7. [15]

    Decoding speech perception from non-invasive brain record- ings,

    A. D ´efossez, C. Caucheteux, J. Rapin, O. Kabeli, and J.-R. King, “Decoding speech perception from non-invasive brain record- ings,”Nature Machine Intelligence, vol. 5, no. 10, pp. 1097–1107, 2023

    2023

  8. [16]

    Brain-to-text: Decoding spoken phrases from phone representations in the brain,

    C. Herff, D. Heger, A. de Pesters, D. Telaar, P. Brunner, G. Schalk, and T. Schultz, “Brain-to-text: Decoding spoken phrases from phone representations in the brain,”Frontiers in Neuroscience, vol. 9, p. 217, 2015

    2015

  9. [17]

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

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

    2020

  10. [18]

    Speech syn- thesis from neural decoding of spoken sentences,

    G. K. Anumanchipalli, J. Chartier, and E. F. Chang, “Speech syn- thesis from neural decoding of spoken sentences,”Nature, vol. 568, no. 7753, pp. 493–498, 2019

    2019

  11. [19]

    Speech synthesis from ECoG using densely connected 3D convolutional neural networks,

    M. Angrick, C. Herff, E. Mugler, M. C. Tate, M. W. Slutzky, D. J. Krusienski, and T. Schultz, “Speech synthesis from ECoG using densely connected 3D convolutional neural networks,”Journal of Neural Engineering, vol. 16, no. 3, p. 036019, 2019

    2019

  12. [20]

    Decoding spectrotemporal features of overt and covert speech from the hu- man cortex,

    S. Martin, P. Brunner, C. Holdgraf, H.-J. Heinze, N. E. Crone, J. Rieger, G. Schalk, R. T. Knight, and B. N. Pasley, “Decoding spectrotemporal features of overt and covert speech from the hu- man cortex,”Frontiers in Neuroengineering, vol. 7, p. 14, 2014

    2014

  13. [21]

    A high-performance speech neuroprosthesis,

    F. R. Willett, E. M. Kunz, C. Fan, D. T. Avansino, G. H. Wil- son, E. Y . Choi, F. Kamdar, L. R. Hochberg, J. M. Henderson, and K. V . Shenoy, “A high-performance speech neuroprosthesis,” Nature, vol. 620, no. 7976, pp. 1031–1036, 2023

    2023

  14. [22]

    An accurate and rapidly calibrating speech neuroprosthesis,

    N. S. Card, M. Wairagkar, C. Iacobacci, P. Bhatt, T. Singer-Clark, F. R. Willett, K. C. Ames, J. Liu, P. Rezaii, L. R. Hochberg, J. M. Henderson, K. V . Shenoy, and D. M. Brandman, “An accurate and rapidly calibrating speech neuroprosthesis,”New England Journal of Medicine, vo...

    2024

  15. [23]

    The cortical organization of speech processing,

    G. Hickok and D. Poeppel, “The cortical organization of speech processing,”Nature Reviews Neuroscience, vol. 8, no. 5, pp. 393– 402, 2007

    2007

  16. [24]

    Neural reuse of action perception circuits for language, concepts and communication,

    F. Pulverm ¨uller, “Neural reuse of action perception circuits for language, concepts and communication,”Progress in Neurobiol- ogy, vol. 160, pp. 1–44, 2018

    2018

  17. [25]

    Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,” inProc. Interna- tional Conference on Machine Learning (ICML), 2006, pp. 369– 376

    2006

  18. [26]

    Brain-computer interfaces for speech communica- tion,

    J. S. Brumberg, A. Nieto-Castanon, P. R. Kennedy, and F. H. Guenther, “Brain-computer interfaces for speech communica- tion,”Speech Communication, vol. 52, no. 4, pp. 367–379, 2010

    2010

  19. [27]

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

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

    2021

  20. [28]

    Neuropros- thesis for decoding speech in a paralyzed person with anarthria,

    D. A. Moses, S. L. Metzger, J. R. Liu, G. K. Anumanchipalli, J. G. Makin, P. F. Sun, J. Chartier, M. E. Dougherty, P. M. Liu, G. M. Abrams, A. Tu-Chan, K. Ganguly, and E. F. Chang, “Neuropros- thesis for decoding speech in a paralyzed person with anarthria,” New England Journa...

    2021

  21. [29]

    A high-performance neuroprosthesis for speech decoding and avatar control,

    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 c...

    2023

  22. [30]

    Generalizable spelling using a speech neuroprosthesis in an individual with severe limb and vocal paralysis,

    S. L. Metzger, J. R. Liu, D. A. Moses, M. E. Dougherty, M. P. Seaton, K. T. Littlejohn, J. Chartier, G. K. Anumanchipalli, A. Tu- Chan, K. Ganguly, and E. F. Chang, “Generalizable spelling using a speech neuroprosthesis in an individual with severe limb and vocal paralysis,”Na...

    2022

  23. [31]

    An instan- taneous voice-synthesis neuroprosthesis,

    M. Wairagkar, N. S. Card, T. Singer-Clark, X. Hou, C. Iacobacci, L. R. Hochberg, D. M. Brandman, and S. D. Stavisky, “An instan- taneous voice-synthesis neuroprosthesis,”Nature, vol. 644, pp. 145–152, 2025

    2025

  24. [32]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

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

    2020

  25. [33]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,”Proc. International Conference on Machine Learning (ICML), pp. 28 492–28 518, 2023

    2023

  26. [34]

    Brain-to-text benchmark ’24: Lessons learned,

    F. R. Willett, J. Li, T. Le, C. Fan, M. Chen, and E. Shlizerman, “Brain-to-text benchmark ’24: Lessons learned,”arXiv preprint arXiv:2412.17227, 2024

    2024 arXiv

  27. [35]

    Brain-to-text decoding with context-aware neural representations and large lan- guage models,

    J. Li, T. Le, C. Fan, M. Chen, and E. Shlizerman, “Brain-to-text decoding with context-aware neural representations and large lan- guage models,”arXiv preprint arXiv:2411.10657, 2024

    2024 arXiv

  28. [36]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    2023 arXiv

  29. [37]

    Exploring the ca- pability of Mamba in speech applications,

    K. Miyazaki, Y . Masuyama, and M. Murata, “Exploring the ca- pability of Mamba in speech applications,” inProc. Interspeech, 2024, pp. 176–180

    2024

  30. [38]

    Speech Slytherin: Examining the performance and efficiency of Mamba for speech separation, recognition, and synthesis,

    X. Jiang, Y . A. Li, A. N. Florea, C. Han, and N. Mesgarani, “Speech Slytherin: Examining the performance and efficiency of Mamba for speech separation, recognition, and synthesis,”arXiv preprint arXiv:2407.09732, 2024

    2024 arXiv

  31. [39]

    ConMamba: A convolution-augmented Mamba encoder model for efficient end-to-end ASR systems,

    H. Houet al., “ConMamba: A convolution-augmented Mamba encoder model for efficient end-to-end ASR systems,” inProc. IEEE International Conference on Signal, Information and Data Processing (ICSIDP), 2024

    2024

  32. [40]

    Assessing the performance and efficiency of Mamba ASR in low-resource scenarios,

    R. Zevallos, M. Cortada-Garcia, S. Solito, C. Mena, A. Peiro- Lilja, and J. Hernando, “Assessing the performance and efficiency of Mamba ASR in low-resource scenarios,” inProc. Interspeech, 2025, pp. 5198–5202

    2025

  33. [41]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  34. [42]

    Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,

    W. Chan, N. Jaitly, Q. Le, and O. Vinyals, “Listen, attend and spell: A neural network for large vocabulary conversational speech recognition,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2016, pp. 4960–4964

    2016

  35. [43]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,” in Proc. Interspeech, 2020, pp. 5036–5040

    2020

  36. [44]

    Efficient Conformer: Progressive downsampling and grouped attention for automatic speech recog- nition,

    M. Burchi and V . Vielzeuf, “Efficient Conformer: Progressive downsampling and grouped attention for automatic speech recog- nition,” inProc. IEEE Automatic Speech Recognition and Under- standing Workshop (ASRU), 2021, pp. 8–15

    2021

  37. [45]

    Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding,

    Y . Peng, S. Dalmia, I. Lane, and S. Watanabe, “Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding,” inProc. In- ternational Conference on Machine Learning (ICML), 2022, pp. 17 627–17 643

    2022

  38. [46]

    End-to-end intracortical speech de- coding from neural activity,

    O. M. Khanday, J. A. Gonzalez-Lopez, M. Ouellet, A. Galdon, and G. Olivares Granados, “End-to-end intracortical speech de- coding from neural activity,”arXiv preprint arXiv:2605.24313, 2026

    2026 arXiv

  39. [47]

    Neural control of cursor trajectory and click by a human with tetraplegia 1000 days after implant of an intracorti- cal microelectrode array,

    J. D. Simeral, S.-P. Kim, M. J. Black, J. P. Donoghue, and L. R. Hochberg, “Neural control of cursor trajectory and click by a human with tetraplegia 1000 days after implant of an intracorti- cal microelectrode array,”Journal of Neural Engineering, vol. 8, no. 2, p. 025027, 2011

    2011

  40. [48]

    Neural manifolds for the control of movement,

    J. A. Gallego, M. G. Perich, L. E. Miller, and S. A. Solla, “Neural manifolds for the control of movement,”Neuron, vol. 94, no. 5, pp. 978–984, 2017

    2017

  41. [49]

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

    B. Jarosiewicz, A. A. Sarma, D. Bacher, N. Y . Masse, J. D. Simeral, B. Sorice, E. M. Oakley, C. Blabe, C. Pandarinath, V . Gilja, S. S. Cash, E. N. Eskandar, G. Friehs, J. M. Hender- son, K. V . Shenoy, J. P. Donoghue, and L. R. Hochberg, “Virtual typing by people with tetrap...

    2015

  42. [50]

    Stabilization of a brain- computer interface via the alignment of low-dimensional spaces of neural activity,

    A. D. Degenhart, W. E. Bishop, E. R. Oby, E. C. Tyler-Kabara, S. M. Chase, A. P. Batista, and B. M. Yu, “Stabilization of a brain- computer interface via the alignment of low-dimensional spaces of neural activity,”Nature Biomedical Engineering, vol. 4, no. 7, pp. 672–685, 2020

    2020

  43. [51]

    SpecAugment: A simple data augmen- tation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “SpecAugment: A simple data augmen- tation method for automatic speech recognition,” inProc. Inter- speech, 2019, pp. 2613–2617

    2019

  44. [52]

    Learning phrase rep- resentations using RNN encoder–decoder for statistical machine translation,

    K. Cho, B. van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase rep- resentations using RNN encoder–decoder for statistical machine translation,” inProc. Conference on Empirical Methods in Natu- ral Language Processing (EMNLP),...

    2014

  45. [53]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inProc. International Conference on Lea...

    2021

  46. [54]

    Language models are unsupervised multitask learn- ers,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language models are unsupervised multitask learn- ers,” OpenAI, Tech. Rep., 2019

    2019

  47. [55]

    KenLM: Faster and smaller language model queries,

    K. Heafield, “KenLM: Faster and smaller language model queries,” inProc. Sixth Workshop on Statistical Machine Trans- lation (WMT), 2011, pp. 187–197

    2011

  48. [56]

    Individual comparisons by ranking methods,

    F. Wilcoxon, “Individual comparisons by ranking methods,”Bio- metrics Bulletin, vol. 1, no. 6, pp. 80–83, 1945

    1945

  49. [57]

    Dealing with the evaluation of supervised classification algorithms,

    G. Santaf ´e, I. Inza, and J. A. Lozano, “Dealing with the evaluation of supervised classification algorithms,”Artificial Intelligence Re- view, vol. 44, no. 4, pp. 467–508, 2015

    2015

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.