Skip to content
ContentLora

    Tip: press / anywhere to search.

    concept

    Speech neuroprostheses

    Also known as speech BCI, speech brain-computer interface, brain-to-voice neuroprosthesis

    A speech neuroprosthesis is a brain-computer interface that decodes the brain activity of attempted speech into text or a synthesized voice, for people who have lost the ability to speak.[1][2] Research systems have reached 97.5 percent word accuracy over eight months in one man with ALS and real-time voice synthesis, and in 2025 and 2026 companies including Neuralink and Paradromics began dedicated speech trials.[3][4][5][6]

    Editor reviewedStrict sourcingUpdated Neuroscience and brain-computer interfacesHealth and medicineLife sciences
    Key facts

    What it is

    Speech neuroprostheses record from the parts of the motor cortex that control the speech articulators. When a person with paralysis tries to speak, these areas still produce patterns that a decoder can turn into words.[7] Systems use either penetrating microelectrode arrays or high-density arrays on the brain’s surface.[1][8]

    Landmark results

    2023, speed. Two Nature papers set the pace. A Stanford team decoded attempted speech from intracortical arrays at 62 words per minute, with a 23.8 percent word error rate on a 125,000-word vocabulary.[7] A UCSF team used surface recordings to reach a median 78 words per minute with a median 25 percent word error rate, and also drove a talking facial avatar.[8]

    2024, accuracy. At UC Davis, a 45-year-old man with ALS received four arrays with 256 electrodes.[1] After 30 minutes of calibration the system was 99.6 percent accurate on 50 words; it later sustained 97.5 percent accuracy over 8.4 months, at about 32 words per minute, in more than 248 hours of conversation.[9][3]

    2025, voice. UCSF researchers streamed synthesized speech in the participant’s pre-injury voice, decoding in 80-millisecond steps.[4] UC Davis researchers synthesized a man’s voice instantly from 256 electrodes and let him change intonation and sing short melodies.[2]

    2025, inner speech. A study in four participants found that imagined (inner) speech is represented in motor cortex and can be decoded in real time.[10]

    The same decoding approach restores other functions. Before speech decoding took off, the fastest text BCI decoded attempted handwriting: in 2021 a participant with a paralysed hand reached 90 characters per minute with 94.1 percent raw accuracy.[11] In 2006 a man with tetraplegia controlled a cursor and a prosthetic hand through a motor-cortex array.[12] In 2023 a brain-spine interface let a man with tetraplegia walk; after rehabilitation with it, he could walk with crutches even when it was switched off.[13][14]

    From labs to companies

    The results above came from academic clinical studies; the UC Davis work ran within the BrainGate2 trial.[15] GAO notes that AI, including language models that predict what a user intends to say, can cut calibration time and make devices faster and more accurate.[16] Companies are now running their own speech studies. Neuralink‘s VOICE study began in October 2025 and aims for 140 words per minute.[5] Paradromics received FDA approval for its Connect-One speech study in November 2025 and said in September 2026 that its first participant was communicating in real time.[6][17] Speech-decoding intellectual property from Stanford and the University of California has been licensed to Blackrock Neurotech, Neuralink and Paradromics.[18]

    Limits

    Each landmark result comes from one or a few participants.[1][10] The best systems remain well below natural conversation at about 160 words per minute.[19] Decoding inner speech raises mental-privacy questions; the 2025 study showed that some free-form inner speech could be picked up and tested safeguards against unintended decoding.[20] No speech BCI was authorized by FDA for marketing as of the latest GAO review in December 2024.[21]

    Questions readers ask

    How accurate are speech BCIs?

    In a 2024 study of one man with ALS, the system sustained 97.5 percent word accuracy over 8.4 months, and he used it for more than 248 hours of conversation.[3]

    How fast are they compared with normal speech?

    Landmark 2023 systems reached 62 to 78 words per minute, against about 160 words per minute for natural conversation.[7][8][19]

    Could a speech BCI decode private thoughts?

    A 2025 study found some aspects of free-form inner speech could be decoded during specific tasks, and showed strategies that prevent unintended decoding of private inner speech.[20]

    Can I get one?

    Not as an approved product. Speech BCIs are available only within clinical studies such as BrainGate2, Neuralink's VOICE and Paradromics' Connect-One.[15][5][6][21]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      In the UC Davis speech neuroprosthesis study, a 45-year-old man with ALS received four microelectrode arrays with 256 electrodes in his left ventral precentral gyrus. confirmedas of 2024-08-15

    2. [2]

      A 2025 UC Davis study synthesized a man's voice in real time from 256 implanted microelectrodes, and let him change intonation and sing short melodies. confirmedas of 2025-06-12

    3. [3]

      The UC Davis neuroprosthesis reached 99.6 percent accuracy on a 50-word vocabulary on its first day of use and sustained 97.5 percent accuracy over 8.4 months, with the participant conversing at about 32 words per minute for more than 248 hours. confirmedas of 2024-08-15

    4. [4]

      A 2025 UCSF study used high-density surface recordings to drive a streaming speech synthesizer in the participant's pre-injury voice, decoding in 80-millisecond increments. confirmedas of 2025-03-31

    5. [6]

      In November 2025 FDA granted Paradromics an Investigational Device Exemption for the Connect-One early feasibility study of its Connexus BCI to restore speech and computer control, initially enrolling two participants. confirmedas of 2025-11-20

    6. [7]

      A 2023 Stanford study decoded the attempted speech of a participant with ALS from intracortical arrays at 62 words per minute, with word error rates of 9.1 percent on a 50-word vocabulary and 23.8 percent on a 125,000-word vocabulary, and found a detailed articulatory representation of phonemes persisting years after paralysis. confirmedas of 2023-08-23

    7. [8]

      A 2023 UCSF study used high-density recordings from the surface of the speech cortex to decode text at a median 78 words per minute with a median word error rate of 25 percent, and to drive a speaking facial avatar. confirmedas of 2023-08-23

    8. [9]

      Calibrating the UC Davis neuroprosthesis took 30 minutes of recordings on the first day, and after 1.4 more hours of training it reached 90.2 percent accuracy on a 125,000-word vocabulary. confirmedas of 2024-08-15

    9. [10]

      A 2025 study in four participants found that inner speech is robustly represented in motor cortex and that imagined sentences can be decoded in real time. confirmedas of 2025-08-21

    10. [11]

      In a 2021 Stanford study, a participant whose hand was paralysed by spinal cord injury used an intracortical BCI that decoded attempted handwriting to type 90 characters per minute with 94.1% raw accuracy, close to typical smartphone typing speeds for his age group (115 characters per minute). confirmedas of 2021-05-12

    11. [12]

      In a 2006 Nature study, a man with tetraplegia used a 96-microelectrode array implanted in his primary motor cortex to move a computer cursor, open simulated e-mail and operate a prosthetic hand. confirmedas of 2006-07-13

    12. [13]

      A 2023 study reported a fully implanted brain-spine interface that linked cortical signals to spinal cord stimulation and let a man with chronic tetraplegia stand and walk in community settings. confirmedas of 2023-05-24

    13. [14]

      The 2023 brain-spine interface was calibrated within minutes, stayed reliable for over a year including at home, and rehabilitation with it improved recovery so the participant could walk with crutches even when it was switched off. confirmedas of 2023-05-24

    14. [15]

      The UC Davis speech study was part of the BrainGate2 clinical trial (NCT00912041). confirmedas of 2024-08-15

    15. [16]

      GAO reports that BCIs must be individually calibrated to each user's brain signals, and that AI, including language models that predict what a user intends to say, can cut calibration time and make devices faster and more accurate. confirmedas of 2024-12-17

    16. [17]

      In September 2026 Paradromics said its first Connect-One participant had used Connexus for real-time speech and text communication, choosing her own words and answering open-ended questions; the company did not publish accuracy or speed figures. confirmedas of 2026-09-14

    17. [18]

      According to the competing-interest disclosures of a 2025 Nature paper, Stanford University speech-decoding intellectual property has been licensed to Blackrock Neurotech and Neuralink, and University of California speech-BCI intellectual property has been licensed to Paradromics. confirmedas of 2025-06-12

    18. [19]

      Natural conversation runs at about 160 words per minute, according to the 2023 Stanford speech BCI paper. confirmedas of 2023-08-23

    19. [20]

      The same study found some aspects of free-form inner speech could be decoded during recall and counting tasks, and demonstrated strategies that prevent speech BCIs from unintentionally decoding private inner speech. confirmedas of 2025-08-21

    20. [21]

      As of December 2024, FDA had not granted marketing authorization for any implantable device for BCI use. confirmedas of 2024-12-17

    Revision history (2)
    1. Page created.
    2. Refresh: added the 2021 handwriting BCI and the role of language models in calibration; the Paradromics speech result now has a second source.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "Speech neuroprostheses." ContentLora, updated Oct 10, 2026. https://contentlora.com/wiki/speech-neuroprostheses

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    Connections
    Utah array Typical hardwareNeuralink Company trialParadromics Company trial
    ☰ Contents