{"success":true,"data":{"author":"Erin M. Kunz, Benyamin Abramovich Krasa, Foram Kamdar et al. (Stanford NPTL, BrainGate2); Hugging Face mirror","bucketAccessType":"public","bucketName":"raei/Kunz-2025-InnerSpeechMotorCortexImplicationsSpeechNeuroprostheses","bucketPrefix":"","category":"Neuroscience","cloudProvider":"huggingface","createdAt":"2026-11-22T01:09:18.514Z","description":"Intracortical neural recordings released with Kunz et al., Inner speech in motor cortex and implications for speech neuroprostheses (Cell 2025, Stanford). Four BrainGate2 clinical trial participants (T12, T15, T16 and T17) with microelectrode (Utah) arrays in motor cortex performed attempted, inner (silently imagined) and perceived speech tasks, so the data show how inner speech is represented in motor cortex and whether it can be decoded by a speech BCI. Features are threshold-crossing counts (binnedTX) and spike-band power (spikePow) in MATLAB files, about 10.8 GB, in eleven experiment folders: isolated and interleaved verbal behaviors, sentence data, the chittychittybangbang experiment, conjunctive counting and six sequence-recall tasks. Each folder has its own README with field definitions, and DRYAD_README.md is the original documentation; load with scipy.io.loadmat and the nptl-stanford inner_speech code. This unofficial mirror extracts Dryad's ZIP archives without modifying data files. The Dryad dataset (DOI 10.5061/dryad.gf1vhhn1j) is released under CC0 1.0. Cite Kunz, Abramovich Krasa, Kamdar et al., Cell 2025 (doi 10.1016/j.cell.2025.06.015).","displayTitle":"Inner Speech in Motor Cortex: Utah-Array Recordings of Attempted, Inner and Perceived Speech in Four BrainGate2 Participants (Kunz et al., Cell 2025, Dryad)","doi":"10.5061/dryad.gf1vhhn1j","huggingFaceUrl":"https://huggingface.co/datasets/raei/Kunz-2025-InnerSpeechMotorCortexImplicationsSpeechNeuroprostheses","id":"6fc38b34-53a4-549a-b967-2ab648315496","license":"CC0-1.0","licenseEvidenceUrl":"https://api.datacite.org/dois/10.5061/dryad.gf1vhhn1j","objectCountEstimate":84,"probeStatus":"ok","referenceUrl":"https://doi.org/10.1016/j.cell.2025.06.015","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"raei/Kunz-2025-InnerSpeechMotorCortexImplicationsSpeechNeuroprostheses\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"raei/Kunz-2025-InnerSpeechMotorCortexImplicationsSpeechNeuroprostheses\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=raei, license:cc0-1.0, downloads=677"],"format":"mixed","generatedAt":"2026-11-22T01:10:17.515Z","layout":{"example":"chittychittybangbang/t12.2024.12.06_chittychittybangbang_50words.mat","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"chittychittybangbang/t12.2024.12.06_chittychittybangbang_50words.mat","truncatedAt":512}],"sampleObjects":["DRYAD_README.md","chittychittybangbang/chittychittybangbang_readme.txt","chittychittybangbang/t12.2024.12.06_chittychittybangbang_50words.mat","chittychittybangbang/t12.2024.12.17_chittychittybangbang_50words.mat","chittychittybangbang/t12.2024.12.19_chittychittybangbang_50words.mat"],"sidecars":[{"bytes":"---\nlicense: cc0-1.0\npretty_name: Kunz et al. (2025) Inner Speech in Motor Cortex\nsize_categories:\n  - 10GB<n<100GB\ntags:\n  - neuroscience\n  - brain-computer-interface\n  - intracortical-recordings\n  - inner-speech\n  - motor-cortex\n  - matlab\n  - dryad\n---\n\n# Inner speech in motor cortex and implications for speech neuroprostheses\n\nThis is an **unofficial mirror** of the public dataset accompanying Kunz et al.\n(2025). It is not affiliated with or endorsed by the dataset authors, their\ninstitutions, or Dryad.\n\n## Source and documentation\n\n- Original dataset: [Dryad DOI 10.5061/dryad.gf1vhhn1j](https://doi.org/10.5061/dryad.gf1vhhn1j)\n- Source version: 2, published July 30, 2025\n- Article: [Kunz et al., *Cell* (2025)](https://doi.org/10.1016/j.cell.2025.06.015)\n- Original dataset documentation: [`DRYAD_README.md`](./DRYAD_README.md)\n- Official analysis code: [`nptl-stanford/inner_speech`](https://github.com/nptl-stanford/inner_speech)\n\nThe scientific files are the files distributed in Dryad's eleven ZIP archives.\nFor this mirror, the archives were extracted, each redundant archive-name\nwrapper directory was removed, and macOS `__MACOSX`/`.DS_Store` metadata was\nexcluded. The original ZIP archives are not included, and the scientific data\nfiles were not modified.\n\n## Dataset summary\n\nThe dataset contains approximately 10.8 GB of intracortical Utah-array neural\nrecordings from four BrainGate2 participants performing attempted, inner, and\nperceived speech tasks. The recordings include threshold-crossing counts\n(`binnedTX`) and spike-band-power features (`spikePow`) stored in MATLAB files.\nThey are not scalp EEG or surface ECoG recordings.\n\nThe top-level directories correspond to the eleven experiment bundles:\n\n- `isolatedVerbalBehaviors` and `interleavedVerbalBehaviors`\n- `sentenceData` and `chittychittybangbang`\n- `conjunctiveCounting`\n- six `seqRecall*` task directories\n\nEach directory contains its task-specific documentation. The MATLAB files can\nbe loaded in Python with `scipy.io.loadmat`; field definitions and indexing\nconventions are documented in those README files.\n\n## Human-participant data\n\nThe original authors state that the data are anonymized, participants are\nreferenced only by the codes T12, T15, T16, and T17, and informed consent for\ndata sharing was obtained. Users should nevertheless handle these human neural\nrecordings responsibly, comply with applicable institutional and legal\nrequirements, and not attempt to identify participants.\n\n## License\n\nThe Dryad dataset is released under the\n[Creative Commons Zero v1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/)\npublic-domain dedication (`CC0-1.0`). CC0 permits reuse, modification, and\nredistribution, including commercial use. Although attribution is not a CC0\ncondition, scholarly citation of the original dataset is strongly encouraged.\n\n## Citation\n\n> Kunz, Erin; Abramovich Krasa, Benyamin; Kamdar, Foram et al. (2025).\n> Inner speech in motor cortex and implications for speech neuroprostheses\n> [Dataset]. Dryad. https://doi.org/10.5061/dryad.gf1vhhn1j\n\n```bibtex\n@dataset{kunz2025innerspeech,\n  author    = {Kunz, Erin and Abramovich Krasa, Benyamin and Kamdar, Foram and others},\n  title     = {Inner speech in motor cortex and implications for speech neuroprostheses},\n  year      = {2025},\n  publisher = {Dryad},\n  doi       = {10.5061/dryad.gf1vhhn1j},\n  url       = {https://doi.org/10.5061/dryad.gf1vhhn1j}\n}\n```\n","encoding":"utf8","key":"README.md","name":"README.md","size":3450,"truncated":false},{"bytes":"Overview: The `seqRecallSpeech` task is an instructed delay task in which participants attempted to speak words associated with the directions of arrows to compare attempted speech behavior to the `seqRecallVerbalmemory` task (Figures S2 S3EF). \n\nTask design: In an instructed delay task, participants were presented with a text cue and attempted to speak those words. On each trial, the participant first saw a red square with text cue above it indicating what words to say (See \"cueList\" for each cue). Then the square turned green and the text cue was removed as a \"go cue\" after which the participant spoke the cued words. For T12, a sequence of direction words were cued that aligned with arrow sequences shown in the `seqRecallVerbalMemory` task (e.g. speech cue: \"up right up\", arrow sequence cue: ↑ → ↑). For T16 due to limited session time only single words were cued (e.g \"up\"). See \"cueList\" for the set of cues shown. Data was collected in a series of 'blocks', in between which the participant completed blocks of the `seqRecallVerbalMemory` task. Neural data between blocks is not recorded here. Data can be loaded with MATLAB or Python (scipy.io.loadmat).\n\nNeural data files (file name indicates participant t12 or t16):\n`t12.2023.09.28_binnedTX_spikePow_20ms_speak.mat`\n`t16.2024.07.17_binnedTX_spikePow_20ms.mat`\n\nData fields:\nspikePow: T x F matrix of binned spike band power, where T = number of time steps in the experiment and F = number of channels. Spike band power was defined as the mean of the squared voltages observed on the channel after high-pass filtering (250 Hz cutoff; units of microvolts squared). The data was denoised with a linear regression reference technique and block-wise mean subtracted to remove any nonstationairities which may be present in Utah array data. \n\nbinnedTX: T x F matrix of binned threshold crossing counts, where T = number of time steps in the experiment and F = number of channels. The data was denoised with a linear regression reference technique and block-wise mean subtracted to remove any nonstationairities which may be present in Utah array data. Either a -3.5 or -4.5 x RMS threshold was used (see methods for individual participant threshold parameters). \n\ngoTrialEpochs: N x 2 matrix of go period epochs (N = number of trials), where each row describes the starting and ending time step of the go period for that trial. (1-based indexing of 1st dimension of spikePow and binnedTX fields).\n\ndelayTrialEpochs: N x 2 matrix of delay period epochs (N = number of trials), where each row describes the starting and ending time step of the delay period for that trial. (1-based indexing of 1st dimension of spikePow and binnedTX fields).\n\ntrialCues: N x 1 vector of cue numbers (N = number of trials), where each entry describes the movement cue given on that trial. To get the text of that cue, use cueList (1-based indexing). For example to see the cue given on the 5th trial, use MATLAB code \"cueList{trialCue(5)}\" or Python code \"cueList[0,trialCues[5]-1]\".\n\ncueList: C x 1 vector of cues (C = number of unique cues in the experiment), where each entry contains the text cue corresponding to that movement condition.\n\nchanSets: A x 64 List of sets of channel indices organized by individual arrays (A = number of 64 channel Utah Arrays) (1-based indexing of 2nd dimension of spikePow and binnedTX fields).\n\nchanSetNames: A x 1 List of string names of the array (A = number of 64 channel Utah arrays) corresponding to the sets of indices in chanSets. Names determined by the HCP cortical parcellation.\n\nbinSize: (int) Size in ms of the binned time steps, T in `spikePow` and `binnedTX` neural features. Either 20 or 10 depending on the participant. \n\n","encoding":"utf8","key":"seqRecallSpeech/README.txt","name":"README.txt","size":3726,"truncated":false},{"bytes":"Overview: The `seqRecallUninstructed3ElementArrow` task is an instructed delay task designed to elicit inner speech for the cognitive task of sequence recall without explicit instruction to engage inner speech (Fig 4AB Fig S3AB)\n\nTask design: In an instructed delay task, participants were presented with cue consisting of a sequence of arrows which indicated a sequence of hand movements. On each trial, the participant first saw a red square with the arrow cue above it indicating the instructed movement sequence (e.g. ↑ → ↑ - move \"up right up\", See \"cueList\" for each cue). Then the square turned green and the cue was removed after which the participant moved their hands in the directions that were cued. T12 retained significant upper extremity dexterity and was able to maneuver a physical joystick to align with the arrow cue directions. T16 was significantly more paralyzed in her arm and hand and therefore attempted to draw sequences of lines in the cued directions. Both participants used their dominant right arm. See \"cueList\" for the set of cues shown. Data was collected in a series of 'blocks', in between which the participant completed blocks of the `seqRecallUninstructed3ElementLines` and for T12 only also the `seqRecallUninstructedSingleArrow` tasks. Neural data between blocks is not recorded here. Data can be loaded with MATLAB or Python (scipy.io.loadmat).\n\nNeural data files (file name indicates participant t12 or t16):\n`t12.2023.05.16_binnedTX_spikePow_20ms_joystickSeq.mat`\n`t16.2024.04.08_binnedTX_spikePow_20ms.mat`\n\nData fields:\nspikePow: T x F matrix of binned spike band power, where T = number of time steps in the experiment and F = number of channels. Spike band power was defined as the mean of the squared voltages observed on the channel after high-pass filtering (250 Hz cutoff; units of microvolts squared). The data was denoised with a linear regression reference technique and block-wise mean subtracted to remove any nonstationairities which may be present in Utah array data. \n\nbinnedTX: T x F matrix of binned threshold crossing counts, where T = number of time steps in the experiment and F = number of channels. The data was denoised with a linear regression reference technique and block-wise mean subtracted to remove any nonstationairities which may be present in Utah array data. Either a -3.5 or -4.5 x RMS threshold was used (see methods for individual participant threshold parameters). \n\ngoTrialEpochs: N x 2 matrix of go period epochs (N = number of trials), where each row describes the starting and ending time step of the go period for that trial. (1-based indexing of 1st dimension of spikePow and binnedTX fields).\n\ndelayTrialEpochs: N x 2 matrix of delay period epochs (N = number of trials), where each row describes the starting and ending time step of the delay period for that trial. (1-based indexing of 1st dimension of spikePow and binnedTX fields).\n\ntrialCues: N x 1 vector of cue numbers (N = number of trials), where each entry describes the movement cue given on that trial. To get the text of that cue, use cueList (1-based indexing). For example to see the cue given on the 5th trial, use MATLAB code \"cueList{trialCue(5)}\" or Python code \"cueList[0,trialCues[5]-1]\".\n\ncueList: C x 1 vector of cues (C = number of unique cues in the experiment), where each entry contains the text cue corresponding to that movement condition.\n\nchanSets: A x 64 List of sets of channel indices organized by individual arrays (A = number of 64 channel Utah Arrays) (1-based indexing of 2nd dimension of spikePow and binnedTX fields).\n\nchanSetNames: A x 1 List of string names of the array (A = number of 64 channel Utah arrays) corresponding to the sets of indices in chanSets. Names determined by the HCP cortical parcellation.\n\nbinSize: (int) Size in ms of the binned time steps, T in `spikePow` and `binnedTX` neural features. Either 20 or 10 depending on the participant. \n","encoding":"utf8","key":"seqRecallUninstructed3ElementArrow/README.txt","name":"README.txt","size":3946,"truncated":false},{"bytes":"Overview: The `seqRecallUninstructed3ElementLines` task is an instructed delay task designed NOT to elicit inner speech during sequence recall when no explicit mental strategy instruction was given. It serves as a motor sequencing control for the `seqRecallUninstructed3ElementArrow` task. (Fig 4AB Fig S3AB)\n\nTask design: In an instructed delay task, participants were presented with an image cue showing three line segments with a start and finish which instructed a sequence of hand movements. On each trial, the participant first saw an image cue denoting three line segments with an \"x\" at the start (see `example_image_cue_right_up_right.png`). Then the image cue was removed and replaced with a green square after which the participant attempted to reproduce the image shown with a sequence of three hand movements. T12 retained significant upper extremity dexterity and was able to hold a stylus and draw. T16 was significantly more paralyzed in her arm and hand and therefore attempted to draw lines while holding a stylus. Both participants used their dominant right arm. See \"cueList\" for the set of cues shown. Data was collected in a series of 'blocks', in between which the participant completed blocks of the `seqRecallUninstructed3ElementArrow` and, for T12 only, also the `seqRecallUninstructedSingleArrow` task. Neural data between blocks is not recorded here. Data can be loaded with MATLAB or Python (scipy.io.loadmat).\n\nNeural data files (file name indicates participant t12 or t16):\n`t12.2023.09.19_binnedTX_spikePow_20ms_lineSequenes.mat`\n`t16.2024.04.08_binnedTX_spikePow_20ms_lines.mat`\n\nData fields:\nspikePow: T x F matrix of binned spike band power, where T = number of time steps in the experiment and F = number of channels. Spike band power was defined as the mean of the squared voltages observed on the channel after high-pass filtering (250 Hz cutoff; units of microvolts squared). The data was denoised with a linear regression reference technique and block-wise mean subtracted to remove any nonstationairities which may be present in Utah array data. \n\nbinnedTX: T x F matrix of binned threshold crossing counts, where T = number of time steps in the experiment and F = number of channels. The data was denoised with a linear regression reference technique and block-wise mean subtracted to remove any nonstationairities which may be present in Utah array data. Either a -3.5 or -4.5 x RMS threshold was used (see methods for individual participant threshold parameters). \n\ngoTrialEpochs: N x 2 matrix of go period epochs (N = number of trials), where each row describes the starting and ending time step of the go period for that trial. (1-based indexing of 1st dimension of spikePow and binnedTX fields).\n\ndelayTrialEpochs: N x 2 matrix of delay period epochs (N = number of trials), where each row describes the starting and ending time step of the delay period for that trial. (1-based indexing of 1st dimension of spikePow and binnedTX fields).\n\ntrialCues: N x 1 vector of cue numbers (N = number of trials), where each entry describes the movement cue given on that trial. To get the text of that cue, use cueList (1-based indexing). For example to see the cue given on the 5th trial, use MATLAB code \"cueList{trialCue(5)}\" or Python code \"cueList[0,trialCues[5]-1]\".\n\ncueList: C x 1 vector of cues (C = number of unique cues in the experiment), where each entry contains the text cue corresponding to that movement condition.\n\nchanSets: A x 64 List of sets of channel indices organized by individual arrays (A = number of 64 channel Utah Arrays) (1-based indexing of 2nd dimension of spikePow and binnedTX fields).\n\nchanSetNames: A x 1 List of string names of the array (A = number of 64 channel Utah arrays) corresponding to the sets of indices in chanSets. Names determined by the HCP cortical parcellation.\n\nbinSize: (int) Size in ms of the binned time steps, T in `spikePow` and `binnedTX` neural features. Either 20 or 10 depending on the participant. ","encoding":"utf8","key":"seqRecallUninstructed3ElementLines/README.txt","name":"README.txt","size":4000,"truncated":false}],"sourceObjectKey":"chittychittybangbang/t12.2024.12.06_chittychittybangbang_50words.mat","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"raei/Kunz-2025-InnerSpeechMotorCortexImplicationsSpeechNeuroprostheses\", filename=\"chittychittybangbang/t12.2024.12.06_chittychittybangbang_50words.mat\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"chittychittybangbang/t12.2024.12.06_chittychittybangbang_50words.mat","totalObjects":84},"summary":"Intracortical neural recordings released with Kunz et al., Inner speech in motor cortex and implications for speech neuroprostheses (Cell 2025, Stanford).","title":"Inner Speech in Motor Cortex: Utah-Array Recordings of Attempted, Inner and Perceived Speech in Four BrainGate2 Participants (Kunz et al., Cell 2025, Dryad)","updatedAt":"2026-11-22T04:03:46.696Z","visibility":"public","websiteUrl":"https://github.com/nptl-stanford/inner_speech"}}