Pith. sign in

REVIEW 3 major objections 7 minor 27 references

Brain2Char: A Deep Architecture for Decoding Text from Brain Recordings

T0 review · 3 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Brain2Char decodes ECoG brain recordings directly into text, reaching 7.0–10.6% word error rates across three participants and setting a new state of the art in brain-to-text decoding.

desk verdict The architecture and ablations are solid, but the headline WERs come from repeated test sentences and a task-specific LM, so the SOTA/BCI claim isn't supported yet. read the letter →

arxiv 1909.01401 v1 pith:HK35ZRU4 submitted 2019-09-03 cs.LG cs.CLq-bio.NCstat.ML

Pengfei Sun , Gopala K. Anumanchipalli , Edward F. Chang This is my paper · ORCID
classification cs.LGcs.CLq-bio.NCstat.ML
keywords brain-to-textdecodingelectrocorticographybrain-computerinterfaceneuralspeechrecognitionconnectionisttemporalclassification3Dconvolutionalnetworkarticulatorykinematicssilent
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

Brain2Char is an end-to-end deep network that maps electrocorticography (ECoG) recordings straight to written text, bypassing an explicit speech-reconstruction stage. The paper argues that jointly optimizing neural feature extraction, latent-space regularization, and character decoding with a connectionist temporal classification loss makes better use of limited clinical data than previous two-stage pipelines. In three participants who read sentences aloud, it reports word error rates of 10.6%, 8.5%, and 7.0% on vocabularies of 1200–1900 words, and it also decodes silently mimed sentences in two participants. If these numbers hold for novel utterances, the architecture is a practical step toward a high-bandwidth communication brain-computer interface.

What carries the argument

The load-bearing mechanism is the jointly trained encoder–regularizer–decoder stack. The encoder uses 3D Inception-style convolutional blocks with kernels of different sizes in space, time, and frequency, followed by bidirectional LSTM layers, to turn high-gamma and low-frequency ECoG features into a latent sequence. The decoder applies residual dilated convolutions over that latent sequence and trains with a connectionist temporal classification (CTC) loss—a loss that allows variable-length input to be aligned to a character sequence without per-frame labels—then uses a task-specific 4-gram language model to weight beam search at inference. At training time, auxiliary regression losses pull the latent features toward articulatory kinematic trajectories, acoustic (MFCC) features, and learned session embeddings, so the encoder is forced to preserve speech-related structure and discard session-specific artifacts. The CTC objective makes it possible to train on unsegmented ECoG-to-character pairs, while the auxiliary losses keep the latent space anchored to physiological and session-specific structure.

What would settle it

Hold out sentences that never occur in training and decode them with a language model trained only on a general English corpus; if the word error rate jumps to the 60–80% range, the claimed generalizable brain-to-text decoding is not supported.

Watch

Extended reading notes

Core claim

The discovery is that text can be decoded directly from cortical surface recordings by an end-to-end network whose intermediate representations are constrained by known physiology. Brain2Char takes high-gamma and low-frequency ECoG features, extracts multiscale spatiotemporal patterns with 3D Inception-style convolutions and bidirectional LSTMs, then converts the latent sequence to characters with dilated convolutions and a CTC loss, using a 4-gram language model in beam search at inference. Auxiliary losses push the latent features to predict articulatory kinematic trajectories and acoustic (MFCC) features and to identify recording session, regularizing the network for data collected across clinical sessions. In three participants, this reaches 10.6%, 8.5%, and 7.0% WER; the paper reports a roughly 30% WER gain over its joint synthesis-plus-ASR baseline and roughly another 30% over a two-stage speech-then-text baseline, and shows incremental decoding and 40% and 67% WER on silently mimed sentences.

Load-bearing premise

The reported word-error rates come from sentences whose text appeared in training, decoded with a language model built on the same task corpus; the paper's central bet is that such accuracy also holds for entirely new sentences, which is what a communication aid has to handle.

Editorial extensions

If this is right

  • A two-stage pipeline that first synthesizes speech and then runs an off-the-shelf speech recognizer is not necessary: a single network trained end-to-end on ECoG-to-character objectives achieves materially lower word error rates on the same participants.
  • Incremental decoding works: feeding neural data in 0.2-second windows produces a progressively completed sentence, so the architecture is compatible with a real-time BCI rather than requiring the full utterance.
  • The decoder is not just reading out auditory feedback of the participant's own voice; silently mimed speech is decodable, which matters for users who cannot vocalize.
  • Articulatory kinematic targets are a stronger regularizer than acoustic targets, consistent with the view that sensorimotor cortex encodes movement plans rather than the sound alone.
  • Task-specific language models provide large gains, but the network still performs at a respectable level without any language model, suggesting the neural signal itself carries real text information.

Reading between the lines

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

  • A natural extension would measure word error rate on sentences never shown during training under a language model trained only on an external general-English corpus; this would separate how much of the performance is neural decoding versus memorization of repeated sentences.
  • The session-embedding regularizer is modality-agnostic in spirit, so the same calibration idea could transfer to other non-stationary neural recording modalities such as intracortical spikes recorded across days.
  • An even stronger test of the motor-plan hypothesis would decode attempted speech from participants who cannot produce any sound, where no acoustic reference and no overt vocalization exists.
  • The 0.2-second incremental result suggests reporting prefix-level character error as a function of lookahead, which would define the information rate of the BCI more precisely than full-sentence word error rate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes Brain2Char, a deep architecture that decodes text (character sequences) directly from ECoG recordings. The model combines 3D Inception layers for multi-scale spatiotemporal feature extraction, BiLSTMs and dilated CNNs, CTC loss, and language-model-weighted beam search, with auxiliary losses that regularize latent representations toward articulatory kinematics (AKT), acoustic features (MFCC), and session embeddings. On three participants who spoke prompted sentences, the reported word error rates are 10.6%, 8.5%, and 7.0%, claimed as a new state-of-the-art in brain-to-text decoding and as evidence for a high-performance communication BCI. The paper also reports results for two participants during silently mimed speech and an incremental decoding demonstration.

Significance. If the reported WERs reflected generalizable decoding of novel sentences, this would be a substantial advance: direct character-level decoding from ECoG with 7–10% WER would indeed set a new benchmark and strengthen the case for communication BCIs. The architecture itself is thoughtfully designed: the 3D Inception encoder, the latent-regularization scheme with physiological targets, and the CTC/beam-search decoder are all sensible and internally well-motivated. The internal comparisons (DS0 vs. DS1 vs. B2C) show consistent gains that support the value of the proposed components, and the ablation studies in Fig. 3 are a useful contribution. However, the headline claim of a generalizable brain-to-text system is not supported by the evaluation protocol, as detailed in the major comments. The significance of the paper therefore hinges on whether the evaluation can be redone to separate neural decoding from memorization of repeated prompts and task-specific language-model priors.

major comments (3)
  1. [Section 2, Fig. 2(b)] The paper's own analysis in Fig. 2(b) shows that WER improves with the number of prior repetitions of the test sentence in training data, yet the headline WERs (10.6%, 8.5%, 7.0%) are reported on a test set whose sentences have appeared repeatedly in training. The data description in Section 2 states that P1 and P2 read from the MOCHA-TIMIT corpus of only 460 sentences, and P3/P4 used a limited picture-description set. The manuscript does not report WER for sentences with zero prior repeats, so the reported figures may reflect decoding of memorized prompts rather than generalization to novel utterances. This is load-bearing because the abstract's central claim is a high-performance communication BCI, which must handle sentences never seen in training.
  2. [Section 1.4 and Fig. 2(c)] The best language model used at inference (_L2) is a 4-gram KenLM trained on all training data from the same task corpus. For P1 and P2 this corpus is the closed set of 460 MOCHA-TIMIT sentences, so the language model can encode entire prompted sentences and may compensate for weak neural decoding; the LM weight of 1.5 strengthens this effect. The paper reports no WER on held-out sentences using only a general-purpose language model (such as _L1 or LibriSpeech), so the relative contributions of neural decoding and task-specific linguistic prior are not disentangled. This is necessary to support the claim that the architecture itself achieves near-human text decoding from brain.
  3. [Section 2, Fig. 2(d) and Table 1] The offset/onset cutoff experiments in Fig. 2(d) demonstrate sensitivity to input duration, but they do not test sensitivity to sentence novelty. Table 1's incremental decoding example also appears to be a prompted sentence from the task corpus, not a novel utterance. Consequently, the claim that Brain2Char is suitable for real-time, synchronous, incremental decoding for a communication BCI is not supported by evidence that the decoder generalizes to new sentence content; the incremental behavior could in principle be driven by the task-specific LM completing a memorized prompt after only a short neural prefix.
minor comments (7)
  1. [Section 2, Fig. 2(b) caption] The caption contains a typo: "numer of trials" should be "number of trials."
  2. [Section 1.4, Implementation paragraph] The text says "convlutional layer" and "dialted CNN"; these should be "convolutional" and "dilated."
  3. [Section 1.2, Eq. (1)] The notation F{j,p|κ} is used before it is introduced, and the subscript/superscript convention for F is confusing; please clarify whether the index is a feature index or a basis index.
  4. [Section 1.3, Eq. (2)] The two displayed equations for the regularization cost appear to be missing closing braces and are visually incomplete; please fix the formatting so the two variants are clearly separated.
  5. [Section 2, first paragraph] The data description says participants "P 3 andP 4" with inconsistent spacing; please standardize participant labels throughout the manuscript.
  6. [Section 2, quantitative results] The paper reports single-number WERs without confidence intervals or significance tests across the three participants; given the small number of participants and limited sentences, some measure of variability would strengthen the comparison.
  7. [Section 2, mimed speech results] The mimed-speech results are based on only 20 trials per participant and report 40% and 67% WER; the claim that this shows "potential as a silent communication BCI" should be qualified by the small sample size and the absence of statistical analysis.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported WERs rely on a task-corpus language model and repeated-sentence test trials, making the SOTA/BCI claim partially circular because no held-out-sentence evaluation is reported.

  1. fitted input called prediction [Section 2, Quantitative results, Fig. 2(c)]
    "a task specific language model created using all training data from the task( _L2)."

    The 4-gram language model is trained on all training data from the task. For participants P1 and P2 the task is MOCHA-TIMIT, a closed corpus of 460 independent sentences, so every test sentence's text is present in the LM training text. Decoding uses this LM with weight 1.5 during beam search, allowing the decoder to recover the exact target sentence from LM n-gram statistics rather than from neural evidence. The reported WERs are therefore not a measure of decoding novel text; they are partly a replay of the LM's memorized training text.

  2. fitted input called prediction [Section 2, Quantitative results, Fig. 2(b)]
    "Fig2(b) shows the WER as a trend against the number of prior repeats of testing sentences in training data (in other trials than those in test). It is clear that having several prior example trials of the sentence helps in future decoding of the same sentence at inference time."

    The evaluation computes WER on test trials whose sentences occurred multiple times in training. The paper's own figure shows WER improves as the number of prior repeats increases, and no zero-repeat condition is reported. The headline WERs of 10.6%, 8.5% and 7.0% are therefore measured on repeated/memorized prompts rather than on held-out sentences. Since the central claim is a communication BCI, which must handle novel sentences, the reported prediction is not independent of the training data.

full rationale

The network architecture itself is not derived from its output; it uses standard CTC loss, dilated convolutions, and regularizers, and the cited prior work is not the mechanism forcing the headline numbers. However, the evaluation protocol creates a partial circularity: the task-specific 4-gram LM is fit on the task corpus containing the exact test sentences, and the test sentences are repeated in training. Fig. 2(d)'s offset-cutoff control only demonstrates sensitivity to input duration, not to sentence novelty, so it does not break the circularity. Thus the SOTA/BCI claim is partially circular because the measured WERs are for memorized prompts aided by a task-specific LM, not for novel text. Score 6 reflects that one or more 'predictions' reduce by construction in the evaluation, while the underlying architecture retains independent content.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several hand-tuned hyperparameters and domain assumptions, most critically the evaluation assumption that repeated-sentence training plus a task-specific LM approximates real BCI use.

free parameters (7)
  • 3D inception kernel sizes kt, kw, kh = selected from {1,3,5,7}
    Chosen by hand to capture multi-scale spatiotemporal features; no search or principled selection described.
  • Dilated CNN filter size and dilation ratios = filter 11; ratios [1,2,4,8,16]
    Standard ASR hyperparameters; fixed without reported tuning.
  • Language model weight in beam search = 1.5
    Weight applied to the 4-gram LM score during beam search; value reported but no sensitivity analysis.
  • Auxiliary loss weights alpha_i and decay = linear decay, values not given
    Weights for MFCC, AKT, and session-embedding losses; linear decay coefficients are mentioned but not specified.
  • Dropout rates = 0.5 BiLSTM, 0.15 dilated CNN
    Chosen by hand to control overfitting.
  • Cyclic learning rate = max 0.005, min 0.0001
    Optimizer schedule; not justified.
  • Session embedding vectors Q_Ti = learned continuous vectors
    Learned skip-gram-style time embeddings used as regression targets for session calibration; their dimensionality is not reported.
assumptions (5)
  • domain assumption Analytic amplitude of high gamma (70-150 Hz) and low frequency (0-40 Hz) bands is a sufficient neural feature basis for text decoding.
    The paper preprocesses ECoG into these bands and uses them as inputs without alternative feature evaluation. Section 2 Data.
  • domain assumption Acoustic-to-articulatory inversion (AKT) provides a valid estimate of articulatory kinematics for regularizing the latent representation.
    AKT is estimated from acoustics, not measured directly; if the inversion is inaccurate, the regularization target is wrong. Section 2 Data and Section 1.3.
  • ad hoc to paper The train/test split with repeated sentences of the same text is an acceptable proxy for BCI generalization.
    Fig 2(b) varies the number of training repeats of test sentences and shows WER improves with repeats; the main evaluation does not clearly use held-out novel sentences. This is the load-bearing evaluation assumption.
  • ad hoc to paper A 4-gram language model trained on the task corpus is an appropriate inference-time prior.
    The LM (L2) is trained on all task training data, which can memorize the limited corpus and inflate WER. Section 2, Fig 2(c).
  • domain assumption CTC loss combined with additive auxiliary MSE losses is an effective training objective.
    Standard multitask learning assumption; no analysis of objective conflicts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Brain2Char: A Deep Architecture for Decoding Text from Brain Recordings." pith.science (2026). https://pith.science/paper/HK35ZRU4

@misc{pith2026190901401,
  author       = {Pith},
  title        = {Pith review of: Brain2Char: A Deep Architecture for Decoding Text from Brain Recordings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HK35ZRU4}},
  note         = {Machine review of arXiv:1909.01401}
}
read the original abstract

Decoding language representations directly from the brain can enable new Brain-Computer Interfaces (BCI) for high bandwidth human-human and human-machine communication. Clinically, such technologies can restore communication in people with neurological conditions affecting their ability to speak. In this study, we propose a novel deep network architecture Brain2Char, for directly decoding text (specifically character sequences) from direct brain recordings (called Electrocorticography, ECoG). Brain2Char framework combines state-of-the-art deep learning modules --- 3D Inception layers for multiband spatiotemporal feature extraction from neural data and bidirectional recurrent layers, dilated convolution layers followed by language model weighted beam search to decode character sequences, optimizing a connectionist temporal classification (CTC) loss. Additionally, given the highly non-linear transformations that underlie the conversion of cortical function to character sequences, we perform regularizations on the network's latent representations motivated by insights into cortical encoding of speech production and artifactual aspects specific to ECoG data acquisition. To do this, we impose auxiliary losses on latent representations for articulatory movements, speech acoustics and session specific non-linearities. In 3 participants tested here, Brain2Char achieves 10.6\%, 8.5\% and 7.0\% Word Error Rates (WER) respectively on vocabulary sizes ranging from 1200 to 1900 words. Brain2Char also performs well when 2 participants silently mimed sentences. These results set a new state-of-the-art on decoding text from brain and demonstrate the potential of Brain2Char as a high-performance communication BCI.

Figures

Figures reproduced from arXiv: 1909.01401 by the authors.

Figure 1
Figure 1. Architecture of Brain2Char: Neural data is recorded as participants produce speech. Different filters spanning multiple space, time and frequency dimensions convert recorded potentials into appropriate feature representations. The intermediate features are fed into regularization network and the decoder network. The regularization networks impose MSE losses on regressed speech representations and session embeddings.… view at source ↗
Figure 2
Figure 2. Performance evaluation of Brain2Char (a) WER as a function of increasing amount of training data for baselines and Brain2Char. (b) The effect on WER of sentences against the numer of trials for a given sentence provided in training. (c) Contribution of various language models. (d) Performance on partial neural data with cutoff on either onset or offset. The offset cutoff condition in Fig2(d) also shows that Brain2Ch… view at source ↗
Figure 3
Figure 3. Importance of Regularization factors in Brain2Char. (a) Effect of session calibration on two subjects. (b) WER gains by imposing physiological features targets (e.g., MFCC and AKT) One confound in neural speech recognition is the possibility of decoding from the auditory regions during the participants’ self-perception of their own speech. While, Fig.3(b) confirms that articulation is the primary driver of the perfo… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 25 canonical work pages

  1. [1]

    M., Shinkareva, S

    Mitchell, T. M., Shinkareva, S. V ., Carlson, A., Chang, K. M., Malave, V . L., Mason, R. A., & Just, M. A. (2008). Predicting human brain activity associated with the meanings of nouns. science, 320(5880): 1191-1195

    work page 2008

  2. [2]

    Wehbe, L., Murphy, B., Talukdar, P., Fyshe, A., Ramdas, A., & Mitchell, T. (2014). Simultaneously uncovering the patterns of brain regions involved in different story reading subprocesses. PloS one, 9(11): e112575

    work page 2014

  3. [3]

    G., de Heer, W

    Huth, A. G., de Heer, W. A., Griffiths, T. L., Theunissen, F. E., & Gallant, J. L. (2016). Natural speech reveals the semantic maps that tile human cerebral cortex. Nature, 532(7600): 453

    work page 2016

  4. [4]

    J., Kanwisher, N., Botvinick, M

    Pereira, F., Lou, B., Pritchett, B., Ritter, S., Gershman, S. J., Kanwisher, N., Botvinick, M. & Fedorenko, E. (2018). Toward a universal decoder of linguistic meaning from brain activation. Nature communications, 9(1): 963

    work page 2018

  5. [5]

    J., Churchland, P

    Sejnowski, T. J., Churchland, P. S., & Movshon, J. A. (2014). Putting big data to good use in neuroscience. Nature neuroscience, 17(11): 1440

    work page 2014

  6. [6]

    F., Rieger, J

    Chang, E. F., Rieger, J. W., Johnson, K., Berger, M. S., Barbaro, N. M., & Knight, R. T. (2010). Categorical speech representation in human superior temporal gyrus. Nature neuroscience, 13(11): 1428

    work page 2010

  7. [7]

    Chang, E.F. (2015). Towards large-Scale, human-Based, mesoscopic neurotechnologies. Neuron, 86: 68-78

    work page 2015

  8. [8]

    N., David, S

    Pasley, B. N., David, S. V ., Mesgarani, N., Flinker, A., Shamma, S. A., Crone, N. E., Knight, R. T., & Chang, E. F. (2012). Reconstructing speech from human auditory cortex. PLoS biology, 10(1): e1001251

    work page 2012

Show all 27 references
  1. [9]

    L., Mehta, A

    Akbari, H., Khalighinejad, B., Herrero, J. L., Mehta, A. D., & Mesgarani, N. (2019). Towards reconstructing intelligible speech from the human auditory cortex. Scientific reports, 9(1): 874

    2019

  2. [10]

    Martin, S., Brunner, P., Iturrate, I., Millán, J. D. R., Schalk, G., Knight, R. T., & Pasley, B. N. (2016). Word pair classification during imagined speech using direct brain recordings. Scientific reports, 6: 25803

    2016

  3. [11]

    A., Mesgarani, N., Leonard, M

    Moses, D. A., Mesgarani, N., Leonard, M. K., & Chang, E. F. (2016). Neural speech recognition: continuous phoneme decoding using spatiotemporal representations of human cortical activity. Journal of neural engineering, 13(5): 056004

    2016

  4. [12]

    M., Patton, J

    Mugler, E. M., Patton, J. L., Flint, R. D., Wright, Z. A., Schuele, S. U., Rosenow, J., Shih, J. J., Krusienski, D. J., & Slutzky, M. W. (2014). Direct classification of all American English phonemes using signals from functional speech motor cortex. Journal of neural engineeri...

    2014

  5. [13]

    Herff, C., Heger, D., De Pesters, A., Telaar, D., Brunner, P., Schalk, G., & Schultz, T. (2015). Brain-to-text: decoding spoken phrases from phone representations in the brain. Frontiers in neuroscience, 9: 217

    2015

  6. [14]

    Mugler, E., Tate, M

    Angrick, M., Herff, C. Mugler, E., Tate, M. C., Slutzky, M. W., Krusienski, D. J. & Schultz, T. (2019). Speech synthesis from ECoG using densely connected 3D convolutional neural networks. Journal of neural engineering, 16, 036019

    2019

  7. [15]

    K., Chartier, J., & Chang, E

    Anumanchipalli, G. K., Chartier, J., & Chang, E. F. (2019). Speech synthesis from neural decoding of spoken sentences. Nature, 568: 493-498

    2019

  8. [16]

    Hannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., Prenger, R., Satheesh, S., Sengupta, S., Coates, A., & Ng, A. Y . (2014). Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567

    2014 arXiv

  9. [17]

    Collobert, R., Puhrsch, C., & Synnaeve, G. (2016). Wav2letter: an end-to-end convnet-based speech recognition system. arXiv preprint arXiv:1609.03193

    2016 arXiv

  10. [18]

    Y ., Berger, M

    Edwards, E., Soltani, M., Deouell, L. Y ., Berger, M. S., & Knight, R. T. (2005). High gamma activity in response to deviant auditory stimuli recorded directly from human cortex. Journal of neurophysiology, 94(6): 4269-4280

    2005

  11. [19]

    E., Korzeniewska, A., & Franaszczuk, P

    Crone, N. E., Korzeniewska, A., & Franaszczuk, P. J. (2011). Cortical gamma responses: searching high and low. International Journal of Psychophysiology, 79(1): 9-15

    2011

  12. [20]

    Y ., Franaszczuk, P

    Flinker, A., Korzeniewska, A., Shestyuk, A. Y ., Franaszczuk, P. J., Dronkers, N. F., Knight, R. T., & Crone, N. E. (2015). Redefining the role of Broca’s area in speech. Proceedings of the National Academy of Sciences, 112(9): 2871-2875. 9

    2015

  13. [21]

    K., Johnson, K., & Chang, E

    Chartier, J., Anumanchipalli, G. K., Johnson, K., & Chang, E. F. (2018). Encoding of articulatory kinematic trajectories in human speech sensorimotor cortex. Neuron, 98(5): 1042-1054

    2018

  14. [22]

    Mesgarani, N., Cheung, C., Johnson, K., & Chang, E. F. (2014). Phonetic feature encoding in human superior temporal gyrus. Science, 343(6174): 1006-1010

    2014

  15. [23]

    M., Byron, M

    Churchland, M. M., Byron, M. Y ., Cunningham, J. P., Sugrue, L. P., Cohen, M. R., Corrado, G. S., Newsome, W. T., Clark, A. M., Hosseini, P., Scott, B. B., Bradley, D. C., Smith M. A., Kohn, A., Movshon, J. A., Armstrong, K. M., Moore, T., Chang, S. W., Snyder, L. H., Lisberge...

    2010

  16. [24]

    A., & Chang, E

    Sun, P., Moses, D. A., & Chang, E. F. (2019). Modeling neural dynamics during speech production using a state space variational autoencoder. arXiv:1901.04024

    2019 arXiv

  17. [25]

    Szegedy, C., Ioffe, S., Vanhoucke, V ., & Alemi, A. A. (2017). Inception-v4, inception-resnet and the impact of residual connections on learning. In Thirty-First AAAI Conference on Artificial Intelligence

    2017

  18. [26]

    S., & Dean, J

    Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, 3111-3119

    2013

  19. [27]

    H., & Koehn, P

    Heafield, K., Pouzyrevsky, I., Clark, J. H., & Koehn, P. (2013). Scalable modified Kneser-Ney language model estimation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics 2: 690-696. 10

    2013

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.