Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; the references it cites are not checked yet.

High-resolution neural recordings improve the accuracy of speech decoding

Open access

Published version of a preprint

Abstract

Patients suffering from debilitating neurodegenerative diseases often lose the ability to communicate, detrimentally affecting their quality of life. One solution to restore communication is to decode signals directly from the brain to enable neural speech prostheses. However, decoding has been limited by coarse neural recordings which inadequately capture the rich spatio-temporal structure of human brain signals. To resolve this limitation, we performed high-resolution, micro-electrocorticographic (µECoG) neural recordings during intra-operative speech production. We obtained neural signals with 57× higher spatial resolution and 48% higher signal-to-noise ratio compared to macro-ECoG and SEEG. This increased signal quality improved decoding by 35% compared to standard intracranial signals. Accurate decoding was dependent on the high-spatial resolution of the neural interface. Non-linear decoding models designed to utilize enhanced spatio-temporal neural information produced better results than linear techniques. We show that high-density µECoG can enable high-quality speech decoding for future neural speech prostheses.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-023-42555-1

Cites

Statistics

Figures

Fig. 1 Recording from high-density micro-electrocorticographic (µECoG) electrodes on the human brain during a speech production task. a Flexible liquid crystal polymer (LCP) electrode arrays with 128 (top) and 256 (bottom) recording electrodes (200-µm diameter). b Visual comparison of high-density micro-electrodes (black square markers) with macro-ECoG (red circles) and high-density standard ECoG (blue circles). c µECoG electrodes exhibit higher spatial density (34 – 57 electrodes/cm2) compared to the existing macro-ECoG (1 electrode/cm2) and high-density ECoG arrays (6 electrodes/cm2). d Electrode arrays were implanted over speech motor cortex (SMC) in four awake patients (projected onto an average MNI brain). Electrode arrays had either 128 channels with 1.33 mm center to center spacing (pitch) for S1 (violet), S2 (green), and S3 (blue), or 256 channels with 1.72 mm pitch for S4 (red). e A schematic of the intra-operative speech production task. Color bars indicate the duration of the auditory stimulus (blue), time-to-response (orange), and spoken duration (green). f Subjects performed the speech production task with behavioral results above 95% accuracy for correctly repeated non-words. Source data are provided as a Source Data file.
Fig. 1 Recording from high-density micro-electrocorticographic (µECoG) electrodes on the human brain during a speech production task. a Flexible liquid crystal polymer (LCP) electrode arrays with 128 (top) and 256 (bottom) recording electrodes (200-µm diameter). b Visual comparison of high-density micro-electrodes (black square markers) with macro-ECoG (red circles) and high-density standard ECoG (blue circles). c µECoG electrodes exhibit higher spatial density (34 – 57 electrodes/cm2) compared to the existing macro-ECoG (1 electrode/cm2) and high-density ECoG arrays (6 electrodes/cm2). d Electrode arrays were implanted over speech motor cortex (SMC) in four awake patients (projected onto an average MNI brain). Electrode arrays had either 128 channels with 1.33 mm center to center spacing (pitch) for S1 (violet), S2 (green), and S3 (blue), or 256 channels with 1.72 mm pitch for S4 (red). e A schematic of the intra-operative speech production task. Color bars indicate the duration of the auditory stimulus (blue), time-to-response (orange), and spoken duration (green). f Subjects performed the speech production task with behavioral results above 95% accuracy for correctly repeated non-words. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 2 Highly resolved neural features as demonstrated by high-density µECoG during speech production. a Spectrograms for each electrode from each array for two example subjects (averaged across all spoken trials): S1, 128 electrodes (1.33 mm spacing) & S4, 256 electrodes (1.72 mm spacing). Data for all subjects are available in the supplemental materials. Electrodes with significant high-gamma modulation (HG: 70 – 150 Hz, FDR-corrected non-parametric permutation tests) during the speech utterance are indicated by black borders. Shaded grey circles represent superimposed simulated macro-ECoG electrode resolution with 10 mm interelectrode distance, illustrating the increased density of µECoG as compared to standard IEEG. b A spectrogram of an example electrode from the 128-channel array (red border) demonstrated increase power in HG band which was time-locked to the speech utterance. c µECoG arrays captured HG activity at significantly higher power than standard IEEG (ECoG and SEEG), *p < 0.05, ***p < 0.001, one-sided Mann-Whitney U test, showing the increased ability of micro-electrodes to measure neural signals with a higher signal-to-noise ratio as compared to standard methods (IEEG vs. S1: p = 1.4e-14, IEEG vs. S2: p = 1.7e-5, IEEG vs. S3: p = 1e-3, IEEG vs. S4: p = 3.3e-5). The red lines and the blue boxes indicate the median and 25th/75th percentile and solid lines represent the full range of the distribution (total number of electrodes in each grouping, standard IEEG: 60, S1: 111, S2: 111, S3: 63, S4: 149). d Inter-electrode HG correlation decreased with increased electrode spacing for all subjects. The correlation values at each electrode spacing are represented by mean and standard error (n = all possible electrode pairs at fixed electrode distance in mm). The correlation values at spatial resolutions less than 2 mm (r = 0.6), reveals evidence for spatially discriminative neural signals at the micro-scale during a speech utterance. Source data are provided as a Source Data file.
Fig. 2 Highly resolved neural features as demonstrated by high-density µECoG during speech production. a Spectrograms for each electrode from each array for two example subjects (averaged across all spoken trials): S1, 128 electrodes (1.33 mm spacing) & S4, 256 electrodes (1.72 mm spacing). Data for all subjects are available in the supplemental materials. Electrodes with significant high-gamma modulation (HG: 70 – 150 Hz, FDR-corrected non-parametric permutation tests) during the speech utterance are indicated by black borders. Shaded grey circles represent superimposed simulated macro-ECoG electrode resolution with 10 mm interelectrode distance, illustrating the increased density of µECoG as compared to standard IEEG. b A spectrogram of an example electrode from the 128-channel array (red border) demonstrated increase power in HG band which was time-locked to the speech utterance. c µECoG arrays captured HG activity at significantly higher power than standard IEEG (ECoG and SEEG), *p < 0.05, ***p < 0.001, one-sided Mann-Whitney U test, showing the increased ability of micro-electrodes to measure neural signals with a higher signal-to-noise ratio as compared to standard methods (IEEG vs. S1: p = 1.4e-14, IEEG vs. S2: p = 1.7e-5, IEEG vs. S3: p = 1e-3, IEEG vs. S4: p = 3.3e-5). The red lines and the blue boxes indicate the median and 25th/75th percentile and solid lines represent the full range of the distribution (total number of electrodes in each grouping, standard IEEG: 60, S1: 111, S2: 111, S3: 63, S4: 149). d Inter-electrode HG correlation decreased with increased electrode spacing for all subjects. The correlation values at each electrode spacing are represented by mean and standard error (n = all possible electrode pairs at fixed electrode distance in mm). The correlation values at spatial resolutions less than 2 mm (r = 0.6), reveals evidence for spatially discriminative neural signals at the micro-scale during a speech utterance. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 3 µECoG captures fine-scale spatio-temporal activity during speech production. a HG activations showed fine-scale spatio-temporal patterns in SMC, shown here for an example subject (S1) during speech production. Spatial maps of HG responses showed discriminative activation patterns for each phoneme. b State-space analysis using a signal value decomposition t-distributed stochastic neighbor embedding method (SVD-tSNE) depicted clear hierarchical separation of trials with respect to articulators (dorsal - /g, k/, labial (/b, p, v/), high (/i, u/), and low (/ae, a/) as well as individual phonemes. Phonemes produced using similar articulators were grouped together in two-dimensional state-space suggesting similarity in motor activation patterns. c Electrode sub-sampling to simulate lower resolution recordings reduced the separation of clusters for both phoneme (top) and articulator (bottom) groupings in SVD-tSNE state-space (same color palette as in (b)). This reduction of grouping demonstrates the utility of high-resolution spatial sampling. d To quantify this reduction, we performed a silhouette analysis that measures the relative closeness of a trial to articulator/phoneme group with respect to other groups. To show this significant decrease in cluster separation/groupings empirically, we performed a one-way ANOVA (articulator: F3, 196 = 114, p = 9.6e-43, phoneme: F3, 196 = 47.8, p = 3e-23) and post hoc t-tests (see main text). Each box plot depicts the distribution of silhouette scores obtained from electrode sub-samplings (50 samples using Poisson disc sampling). The red lines and the blue boxes indicate the median and 25th/75th percentile and dashed lines represent the full-range of the distribution. The dotted horizonal line indicate chance silhouette metric obtained by shuffling the articulator/phoneme labels (1000 iterations). These analyses demonstrate the improved ability of high-resolution neural sampling to distinguish between articulator/phoneme motor features. Source data are provided as a Source Data file.
Fig. 3 µECoG captures fine-scale spatio-temporal activity during speech production. a HG activations showed fine-scale spatio-temporal patterns in SMC, shown here for an example subject (S1) during speech production. Spatial maps of HG responses showed discriminative activation patterns for each phoneme. b State-space analysis using a signal value decomposition t-distributed stochastic neighbor embedding method (SVD-tSNE) depicted clear hierarchical separation of trials with respect to articulators (dorsal - /g, k/, labial (/b, p, v/), high (/i, u/), and low (/ae, a/) as well as individual phonemes. Phonemes produced using similar articulators were grouped together in two-dimensional state-space suggesting similarity in motor activation patterns. c Electrode sub-sampling to simulate lower resolution recordings reduced the separation of clusters for both phoneme (top) and articulator (bottom) groupings in SVD-tSNE state-space (same color palette as in (b)). This reduction of grouping demonstrates the utility of high-resolution spatial sampling. d To quantify this reduction, we performed a silhouette analysis that measures the relative closeness of a trial to articulator/phoneme group with respect to other groups. To show this significant decrease in cluster separation/groupings empirically, we performed a one-way ANOVA (articulator: F3, 196 = 114, p = 9.6e-43, phoneme: F3, 196 = 47.8, p = 3e-23) and post hoc t-tests (see main text). Each box plot depicts the distribution of silhouette scores obtained from electrode sub-samplings (50 samples using Poisson disc sampling). The red lines and the blue boxes indicate the median and 25th/75th percentile and dashed lines represent the full-range of the distribution. The dotted horizonal line indicate chance silhouette metric obtained by shuffling the articulator/phoneme labels (1000 iterations). These analyses demonstrate the improved ability of high-resolution neural sampling to distinguish between articulator/phoneme motor features. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 4 High accuracy phoneme decoding from µECoG. a Confusion matrix for an example subject (S1) for phoneme position one (P1) using singular value decomposition linear discriminant analysis (SVD-LDA) decoding demonstrated high accuracy for each phoneme. b Decoding was also highly accurate for each position within the non-word (P1, P2, P3), as well as for all phonemes combined across position (mean P1 = 49.9%, P2 = 45%, P3 = 38.8%, All = 36.6%, and chance = 11.11%). c This accurate decoding outperformed standard IEEG in a 4-way vowel decoder (P1: µECoG (n = 4): 71%, vs. ECoG (n = 4): 36.7%, p = 0.015, vs. SEEG (n = 7): 36.5%, p = 0.011; chance = 25%, one-sided permutation test with 10,000 iterations). This nearly doubled performance increase further demonstrates the benefit of higher resolution neural recordings for speech decoding. The red lines and the blue boxes indicate the median and 25th/75th percentile and dashed lines represent the full-range of the distribution. d Distribution of single-electrode decoding values were correlated with HG-ESNR for all subjects. Linear model fits indicated moderate but significant correlations for S1 (purple - Pearson r = 0.23, p = 0.0087), S2 (green - r = 0.23, p = 0.0077), and S3 (blue - r = 0.19, p = 0.0303), and a stronger correlation for S4 (red - r = 0.44, p = 6.5e-14, F-statistic test against constant model). e Decoding errors decreased with phoneme distance for all positions indicating higher confusion between phonemes with shared articulatory features (see methods and supplement for how error/phoneme distance was computed). Solid lines indicate linear model fit obtained for each position (P1: F1,79 = 90.4, p = 9.9e-15, P2: F1,79 = 19.7, p = 2.8e-5, P3: F1,79 = 28.6, p = 8.4e-7; F-statistic test against constant model) and shaded regions indicate the 95% confidence interval of the linear model fit obtained using bootstrap regression (n = 1000 bootstraps). f Phoneme errors at all three positions in the non-words and all phonemes combined across position show decoding errors (bits) well below chance (chance = 2.4 bits for a uniform confusion matrix). Source data are provided as a Source Data file.
Fig. 4 High accuracy phoneme decoding from µECoG. a Confusion matrix for an example subject (S1) for phoneme position one (P1) using singular value decomposition linear discriminant analysis (SVD-LDA) decoding demonstrated high accuracy for each phoneme. b Decoding was also highly accurate for each position within the non-word (P1, P2, P3), as well as for all phonemes combined across position (mean P1 = 49.9%, P2 = 45%, P3 = 38.8%, All = 36.6%, and chance = 11.11%). c This accurate decoding outperformed standard IEEG in a 4-way vowel decoder (P1: µECoG (n = 4): 71%, vs. ECoG (n = 4): 36.7%, p = 0.015, vs. SEEG (n = 7): 36.5%, p = 0.011; chance = 25%, one-sided permutation test with 10,000 iterations). This nearly doubled performance increase further demonstrates the benefit of higher resolution neural recordings for speech decoding. The red lines and the blue boxes indicate the median and 25th/75th percentile and dashed lines represent the full-range of the distribution. d Distribution of single-electrode decoding values were correlated with HG-ESNR for all subjects. Linear model fits indicated moderate but significant correlations for S1 (purple - Pearson r = 0.23, p = 0.0087), S2 (green - r = 0.23, p = 0.0077), and S3 (blue - r = 0.19, p = 0.0303), and a stronger correlation for S4 (red - r = 0.44, p = 6.5e-14, F-statistic test against constant model). e Decoding errors decreased with phoneme distance for all positions indicating higher confusion between phonemes with shared articulatory features (see methods and supplement for how error/phoneme distance was computed). Solid lines indicate linear model fit obtained for each position (P1: F1,79 = 90.4, p = 9.9e-15, P2: F1,79 = 19.7, p = 2.8e-5, P3: F1,79 = 28.6, p = 8.4e-7; F-statistic test against constant model) and shaded regions indicate the 95% confidence interval of the linear model fit obtained using bootstrap regression (n = 1000 bootstraps). f Phoneme errors at all three positions in the non-words and all phonemes combined across position show decoding errors (bits) well below chance (chance = 2.4 bits for a uniform confusion matrix). Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 5 Accurate decoding is dependent on the high-spatial resolution, high-channel count capability, and micro-scale recording of µECoG. a Accuracy of P1 decoding increased as a function of spatial resolution and coverage as demonstrated by sub-sampling electrodes and decreased with contact size as demonstrated by spatial averaging. Each box plot shows the distribution of accuracy metrics from the subsampled electrodes (Electrode subsampling: n ≥ 50 electrode subsamples obtained using Poisson Disk sampling, electrode subgrids: n = all possible subsample grids defined by the coverage, spatial averaging: n = 10 instances of spatial averaged data). The horizontal center lines and the boxes indicate the median and 25th/75th percentile and vertical solid lines represent the full-range of the distribution. Blue lines connect the median accuracy values from the boxplots. Electrode pitch less than 2 mm along with micro-scale sampling was required to obtain decoding accuracy within 95% of the maximum decoding for the entire array (red dotted line) and demonstrated 29% improvement from the 10 mm pitch simulation and a 17% improvement from the 4 mm pitch simulation. b Classification for articulators was spatially dependent, resulting in millimeter-scale spatially specific articulator maps shown here in example subjects S1 and S4. Darker colors show higher area under the curve for receiver operating characteristic curves (ROC-AUC) for the univariate linear decoding values for each articulatory feature. Performance for vowels (top) was spatially diffuse, whereas performance for consonant (bottom) formed spatially distinct clusters. The contour points identify 90th percentile of AUC values across all articulators, and the marker indicates corresponding centroids for each articulator. Source data are provided as a Source Data file.
Fig. 5 Accurate decoding is dependent on the high-spatial resolution, high-channel count capability, and micro-scale recording of µECoG. a Accuracy of P1 decoding increased as a function of spatial resolution and coverage as demonstrated by sub-sampling electrodes and decreased with contact size as demonstrated by spatial averaging. Each box plot shows the distribution of accuracy metrics from the subsampled electrodes (Electrode subsampling: n ≥ 50 electrode subsamples obtained using Poisson Disk sampling, electrode subgrids: n = all possible subsample grids defined by the coverage, spatial averaging: n = 10 instances of spatial averaged data). The horizontal center lines and the boxes indicate the median and 25th/75th percentile and vertical solid lines represent the full-range of the distribution. Blue lines connect the median accuracy values from the boxplots. Electrode pitch less than 2 mm along with micro-scale sampling was required to obtain decoding accuracy within 95% of the maximum decoding for the entire array (red dotted line) and demonstrated 29% improvement from the 10 mm pitch simulation and a 17% improvement from the 4 mm pitch simulation. b Classification for articulators was spatially dependent, resulting in millimeter-scale spatially specific articulator maps shown here in example subjects S1 and S4. Darker colors show higher area under the curve for receiver operating characteristic curves (ROC-AUC) for the univariate linear decoding values for each articulatory feature. Performance for vowels (top) was spatially diffuse, whereas performance for consonant (bottom) formed spatially distinct clusters. The contour points identify 90th percentile of AUC values across all articulators, and the marker indicates corresponding centroids for each articulator. Source data are provided as a Source Data file.CC BY · PubMed Central
Fig. 6 µECoG enables the sequential temporal decoding of utterances. a A temporal decoding model was trained on data aligned to the speech utterance onset for phonemes at each position. Results are shown for two example subjects (S1, S2) revealing temporal patterns of successful decoding for each phoneme position (P1: blue, P2: orange, P3: yellow). Dotted lines reflect peak decoding. This ability demonstrates that the temporal sequence of phoneme positions can be reconstructed via the information present in the neural signal aligned to the utterance. b We leveraged this ability by constructing a nonlinear decoding model to resolve temporal speech sequences of entire utterances. We used a long short-term memory (LSTM) Seq2Seq model with an encoder-decoder recurrent neural network (RNN) to translate HG spatio-temporal activation patterns (1 second window centered at utterance onset) to phoneme sequences, with three layers: a temporal convolution (green), a bi-directional encoder RNN (blue), and a decoder RNN (orange) layer. The numbers in the boxes represent the number of LSTM units (c) Our LSTM model outperformed a linear model for decoding phonemes across all positions, shown here for two example subjects, S1 (mean accuracy: 46.4% vs. 34.4%, mean phoneme error: 0.99 bits vs. 1.68 bits, n = 10 decoding instances on test trials) and S2 (accuracy: 50.8% vs. 43%, phoneme error: 0.91 bits vs. 1.4 bits, n = 10 decoding instances on test trials). Error bars indicate standard deviations, and the data points indicate performance metrics from each decoding instance on test trials. d Subsampling of electrodes (n = 25 instances of electrode subsamplings) reduced the LSTM performance in decoding phoneme sequences as validated by one-way ANOVA (S1: F3, 99 = 40.47, p = 5.4e-17, S2: F3, 99 = 99.97, p = 2e-79) and post-hoc t-tests (see Main text). The red lines and the blue boxes indicate the median and 25/75th percentile and dashed lines represent the full-range of the distribution. This improvement in decoding demonstrated the benefit of combining nonlinear decoding models that leverage high spatio-temporal neural activation with µECoG recordings that can sample the human brain at the micro-scale. Source data are provided as a Source Data file.
Fig. 6 µECoG enables the sequential temporal decoding of utterances. a A temporal decoding model was trained on data aligned to the speech utterance onset for phonemes at each position. Results are shown for two example subjects (S1, S2) revealing temporal patterns of successful decoding for each phoneme position (P1: blue, P2: orange, P3: yellow). Dotted lines reflect peak decoding. This ability demonstrates that the temporal sequence of phoneme positions can be reconstructed via the information present in the neural signal aligned to the utterance. b We leveraged this ability by constructing a nonlinear decoding model to resolve temporal speech sequences of entire utterances. We used a long short-term memory (LSTM) Seq2Seq model with an encoder-decoder recurrent neural network (RNN) to translate HG spatio-temporal activation patterns (1 second window centered at utterance onset) to phoneme sequences, with three layers: a temporal convolution (green), a bi-directional encoder RNN (blue), and a decoder RNN (orange) layer. The numbers in the boxes represent the number of LSTM units (c) Our LSTM model outperformed a linear model for decoding phonemes across all positions, shown here for two example subjects, S1 (mean accuracy: 46.4% vs. 34.4%, mean phoneme error: 0.99 bits vs. 1.68 bits, n = 10 decoding instances on test trials) and S2 (accuracy: 50.8% vs. 43%, phoneme error: 0.91 bits vs. 1.4 bits, n = 10 decoding instances on test trials). Error bars indicate standard deviations, and the data points indicate performance metrics from each decoding instance on test trials. d Subsampling of electrodes (n = 25 instances of electrode subsamplings) reduced the LSTM performance in decoding phoneme sequences as validated by one-way ANOVA (S1: F3, 99 = 40.47, p = 5.4e-17, S2: F3, 99 = 99.97, p = 2e-79) and post-hoc t-tests (see Main text). The red lines and the blue boxes indicate the median and 25/75th percentile and dashed lines represent the full-range of the distribution. This improvement in decoding demonstrated the benefit of combining nonlinear decoding models that leverage high spatio-temporal neural activation with µECoG recordings that can sample the human brain at the micro-scale. Source data are provided as a Source Data file.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

The citation lists are queued

They are fetched later and appear here then.

References 70 · Cited by 75

Research works printed but not linked: 4

  1. Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv Prepr. arXiv1412.6980 (2014).
  2. Anumanchipalli, G. K., Chartier, J. & Chang, E. F. Speech synthesis from neural decoding of spoken sentences. Nature 568, 493–498 (2019). · DOI · Paper page
  3. Kaufman, L. & Rousseeuw, P. J. Finding groups in data: an introduction to cluster analysis. (John Wiley & Sons, 2009).
  4. Abadi, M. et al. Tensorflow: A system for large-scale machine learning. In 12th {USENIX} symposium on operating systems design and implementation ({OSDI} 16) 265–283 (2016).

Metadata

Authors
Duraivel, S., Rahimpour, S., Chiang, C.-H., Trumpis, M., Wang, C., Barth, K., Harward, S. C., Lad, S. P., Friedman, A. H., Southwell, D. G., Sinha, S. R., Viventi, J., & Cogan, G. B.
Affiliations
  • Duke University (US)
  • University of Pennsylvania (US)
Corresponding author
  • Viventi, Jonathan
  • Cogan, Gregory B.
Venue
Nature Communications
Published
2023-11-06
Volume
14
Issue
1
Pages
6938-6938
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4388413005
Cited by
75
References
70
Citations by year
2026: 21 · 2025: 36 · 2024: 17
FWCI
8.95
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • U.S. Department of Health & Human Services | NIH | National Institute of Neurological Disorders and Stroke (R01NS129703, UG3NS120172)
  • U.S. Department of Health & Human Services | NIH | National Institute on Deafness and Other Communication Disorders (R01DC019498)
  • United States Department of Defense | United States Army | Army Medical Command | Congressionally Directed Medical Research Programs (W81XWH-21-0538)
Topics
  • Life Sciences › Neuroscience › Cellular and Molecular Neuroscience › Neuroscience and Neural Engineering
  • Life Sciences › Neuroscience › Cognitive Neuroscience › EEG and Brain-Computer Interfaces
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
Keywords
neural speech decoding, micro-electrocorticography, high-density neural recording, brain-computer interface, signal-to-noise ratio
MeSH terms
Electrocorticography, Brain, Communication, Humans, Quality of Life, Speech, Brain-Computer Interfaces

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: High-resolution neural recordings improve the accuracy of speech decoding. https://app.cochat.ai/papers/10.1038/s41467-023-42555-1
The paper
Duraivel, S., Rahimpour, S., Chiang, C.-H., Trumpis, M., Wang, C., Barth, K., Harward, S. C., Lad, S. P., Friedman, A. H., Southwell, D. G., Sinha, S. R., Viventi, J., & Cogan, G. B. (2023). High-resolution neural recordings improve the accuracy of speech decoding. Nature Communications. https://doi.org/10.1038/s41467-023-42555-1

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.