Speech neuroprosthesis

Speech neuroprosthesis is a brain-computer interface that recovers intended speech from neural activity and renders it as text, synthesized voice, or an animated face. It is aimed at people who retain the intention and the neural machinery for speech but have lost the motor pathway: those with amyotrophic lateral sclerosis, brainstem stroke, or severe anarthria from other causes. Between 2021 and 2025 the field moved from decoding a fifty-word vocabulary at a conversational crawl to near-real-time output over vocabularies of a hundred thousand words. By 2026 one participant had used such a system at home for nearly two years, and two companies had begun small feasibility trials of fully implanted wireless versions.

These systems decode attempted articulation rather than thought. The signals come from the region of cortex that would drive the lips, jaw, tongue and larynx. The decoder recovers a motor plan, and language models turn that noisy plan into plausible sentences. The distinction governs both what the technology can do and which privacy risks it does and does not create.

§How it works

Speech production has a somatotopic map. Along the ventral part of the precentral gyrus, cortical populations encode the movements of individual articulators, and this organization persists in people who have been unable to speak intelligibly for years; the map does not disappear when the output pathway fails.¹ The whole field depends on that persistence.

A speech neuroprosthesis records from that cortex and extracts features: high-gamma power for Electrocorticography interfaces, or threshold crossings and band power for a penetrating Utah array. It passes them to a sequence model that emits phonemes or subword units. A language model then converts the phoneme stream into words, using the statistics of English to resolve ambiguity that the neural signal alone leaves open. These are among the few clinical research systems in which a statistical language model of the kind central to debates about Artificial general intelligence sits inside a control loop with a person's nervous system. The user attempts to speak, silently or with whatever residual movement remains, and the pipeline runs in near real time.

Output can be text on a screen, a synthesized voice, or both. Some systems personalize the voice from recordings made before the person lost speech, and one has driven a digital avatar whose face moves with the decoded articulation. The underlying Neural decoding architecture is the same one used for cursor control, applied to a much higher-dimensional output space.

The language model does a lot of the work

Raw phoneme decoding is error-prone, with about one phoneme in five wrong in the 2023 Stanford system, and the language model often corrected those errors to recover the intended word.¹ A substantial share of performance comes from the language model choosing the most probable sentence consistent with a noisy neural signal. This is legitimate engineering, and human listeners do the same thing. It does mean that the reported word error rate is a property of the whole system, and that unusual or unpredictable utterances are decoded worse than the aggregate figure suggests.

§Development history

The 2021 report that established feasibility involved a man with anarthria following a brainstem stroke, implanted with a surface grid over sensorimotor cortex. He produced sentences from a fifty-word vocabulary at roughly fifteen words per minute, with about a quarter of words wrong.² The output was slow and error-prone, but it was recognisably speech.

Two 2023 papers changed the scale. A Stanford group using four penetrating arrays in a participant with ALS reported 62 words per minute over a 125,000-word vocabulary, with a word error rate near 24 percent, falling under 10 percent when the vocabulary was restricted to fifty words.¹ A UCSF and Berkeley team, using a 253-channel surface array in a woman paralysed by a brainstem stroke, reported a median 78 words per minute over a 1,024-word vocabulary, along with a synthesized voice built from pre-injury recordings and an avatar that moved its face.³

A 2024 report from a UC Davis and Brown collaboration pushed accuracy rather than speed. It reported 97.5 percent word accuracy over a 125,000-word vocabulary, sustained across 8.4 months in a participant with ALS, with the system usable from the first session. In self-paced conversation the rate was about 32 words per minute over more than 248 cumulative hours, slower than either 2023 result.⁴ Later work went after latency. The UCSF group built a streaming decoder that synthesizes speech in the participant's pre-injury voice in 80-millisecond increments rather than at the end of a sentence.⁵ The UC Davis group, working with the same man as the 2024 report, synthesized voice from intracortical activity within 10 milliseconds and decoded enough paralinguistic information for him to change intonation and sing short melodies. Naive listeners transcribing that synthesized voice still got a median of about 44 percent of words wrong, against roughly 96 percent for his own dysarthric speech.⁶

In 2026 the same UC Davis team reported long-term independent home use. Over nearly two years the participant used the system for more than 3,800 hours at home with no researchers present, communicating about 1.96 million words at an average of 56 words per minute and holding a full-time job. He rated 92 percent of sentences as at least mostly correct, a self-report the authors note they could not check against ground truth; in formal tests word accuracy stayed above 99 percent on a 125,000-word vocabulary.⁷

A related line uses imagined handwriting instead of speech. Decoding attempted pen strokes from motor cortex produced roughly 90 characters per minute in one participant, faster than any earlier brain-computer interface typing result, and it shows that the choice of imagined movement is itself a design parameter.⁸

§Company trials

Two companies opened early feasibility studies of fully implanted wireless speech devices: Neuralink's began in October 2025, and Paradromics' was approved in November 2025 and implanted its first participant in June 2026. Neuralink's study, VOICE, tests the skull-mounted N1 implant, whose electrode threads are placed by the company's R1 robot. Registered with ClinicalTrials.gov, it started on 3 October 2025 at UT Southwestern Medical Center in Dallas, with a registered target of six adults with severe speech impairment and impaired upper-limb function from ALS, primary lateral sclerosis, stroke or spinal cord injury. Its primary outcomes are device- and procedure-related adverse events at twelve months; efficacy is a secondary outcome measured as the monthly maximum of correct words per minute. The registry lists "Telepathy" among the implant's other names.⁹ In May 2025 Neuralink announced that the US Food and Drug Administration had granted the speech device a breakthrough device designation, a programme meant to speed development and review that does not amount to approval.¹⁰

Neuralink's trial page says participants attempt or imagine speaking, with output as text or a re-created voice, and that its median participant decodes about 50 words per minute.¹¹ On 28 October 2026 the company stated that one participant with ALS reached 75 words per minute with a 5.8 percent word error rate on a 125,000-word vocabulary in the first week, citing a conference abstract presented at the Northeast ALS Consortium meeting.¹² Those figures have not appeared in a peer-reviewed paper. A blog post the next day described features built during a 72-hour visit to one participant's home, among them decoding that starts and stops without a button press and synthesized speech fed into phone calls.¹³

Paradromics received approval for an investigational device exemption in November 2025 for Connect-One, a feasibility study of its Connexus device that initially enrols two participants across sites at UC Davis, Massachusetts General Hospital and the University of Michigan.¹⁴ The study's first implant took place at Michigan in June 2026, in a woman with motor neuron disease; the device has 421 microelectrodes connected to a transceiver in the chest.¹⁵ In September 2026 the company said she had used it for real-time text and synthesized speech, but the report gave no speed figure; its only accuracy claims were the company's statements that selected sentences were decoded without error, and it cited no publication.¹⁶

Musk's bandwidth forecast

Elon Musk wrote on 30 October 2026 that Neuralink's first product, Telepathy, "enables high-speed communication with computers or humans simply by thinking" and that its "[o]utput bandwidth will vastly exceed the fastest human speech or typing in a few years."¹⁷ In the speech study his post quotes, the implant decodes attempted or imagined speech from speech motor cortex, not thought in general. Natural conversation runs at roughly 150-160 words per minute.¹¹¹ Peer-reviewed decoding has reached a median 78 with a quarter of words wrong,³ the longest home use averaged 56,⁷ and Neuralink gives its own median participant about 50,¹¹ so matching ordinary conversation would take roughly a two- to threefold gain, and exceeding the fastest speakers more than that; no published result shows either. The trial behind the post lists safety, not speed, as its primary outcome.

§What has and has not been shown

The published work has shown fluent, large-vocabulary decoding in individual participants, performance stable enough for daily use over months, and voice synthesis with recognisable identity. It has shown decoding from both surface and penetrating electrodes, and in at least one case usable results from the first day of calibration. In one man it has shown nearly two years of independent home use.⁷

None of this has been shown in more than a handful of people. Nearly every result above comes from a single participant, and participants differ in implant site, disease stage and residual capabilities. There is no multi-participant efficacy trial and no approved device. The only long-term home data come from one participant; the authors name the single participant as a limitation and note that they did not systematically assess fatigue.⁷ There are still almost no data on noisy settings or changing medication. Most work remains in English, although one team decoded both Spanish and English from a bilingual participant¹⁸ and a Chinese group decoded Mandarin syllables, tones included, in real time.¹⁹ The system with the longest record of independent home use connects through percutaneous wired connectors to research computers that trained care partners set up each day.⁷

These decoders also address only speech production. They do not help people whose language comprehension or formulation is impaired, which excludes many stroke and dementia patients. A Cochlear implant restores an input channel and a speech neuroprosthesis restores an output channel, but neither touches the linguistic system between them.

Within the wider field of Neuroprosthetics, speech decoding is unusual in producing output that can be interpreted without ambiguity. Cognitive prosthetics further upstream, such as the hippocampal work described under Memory prosthesis, report effects that are small, contested, and hard to evaluate, precisely because they have no equivalent of a transcript to score.

Every headline figure is one person's decoder

The 2021, 2023, 2024 and 2026 results each come from a single implanted participant. The 2025 inner-speech study analysed four BrainGate2 participants (T12, T15, T16 and T17).²⁰ The decoders behind those results were each trained on one person's own attempted speech. Models trained across people are newer: a 2026 reanalysis of two public datasets aligned each person's cortex into a shared space with a per-subject transform,²¹ and a 2026 preprint that pooled six intracortical participants reported that a shared model beat every individual one and adapted to a new user from fewer than 200 sentences, reaching a word error rate below 7 percent.²² BrainGate2 is a pilot clinical trial and the company studies are early feasibility studies, so the published figures describe individuals rather than a cohort, and there is no denominator for how often an implant reaches conversational performance. A trial that reported performance for every participant it enrolled would supply one.

§Inner speech and privacy

The obvious worry is that a device might transcribe thoughts a person did not choose to utter. That is neither science fiction nor imminent. Attempted speech and imagined speech produce overlapping but distinguishable patterns in motor cortex. A 2025 study decoded instructed inner speech from the same electrodes with word error rates of 26-54 percent on a large vocabulary, but could not decode complete, intelligible sentences during free-form thinking. It also demonstrated a keyword-gated mode in which the decoder stays locked until the user imagines saying a rare word, "chittychittybangbang"; in real-time tests with one participant the keyword was correctly detected or correctly ignored 98.75 percent of the time.²⁰

That is a reassuring engineering answer to a real problem, and it shows the shape of the issue: the safeguard has to be designed in, because the signal is present whether or not the system is asked to use it. The broader questions belong to Mental privacy and the developing Neurorights frameworks. The ethicist Marcello Ienca places speech decoding within the narrow sense of mental privacy, the decoding of present thoughts, which he distinguishes from inferences about traits and risks drawn from brain data.²³ The questions include who holds the neural recordings, whether decoded text has legal status as speech, and what happens when a decoder outputs a sentence the user did not intend. Error correction in a speech prosthesis matters for more than appearance, because a wrong word attributed to a person who cannot verbally repudiate it is a distinctive harm.

§Outlook

The company studies are early feasibility trials, the stage before any pivotal trial, and both enrol people with ALS among other causes of severe paralysis, the population with the clearest need and the strongest existing results.⁹¹⁴ Progress there arguably depends less on decoding accuracy, which the academic results suggest is already adequate for conversation, than on hardware longevity, surgical throughput, and the unglamorous work of making a device that a clinician can implant and a family can support at home. The same constraints shape every implanted Brain–computer interface, from the endovascular Stentrode of Synchron to the penetrating threads of Neuralink.

These systems are also the strongest existing evidence for the modest version of Human–AI merger. A machine-learning model sits between cortex and the world, doing inferential work that the user's damaged nervous system cannot, and the output does not separate the user's contribution from the model's.

The scientific question underneath is how much of language can be reached this way at all. Decoding articulation works because articulation is motor. Semantic content, the words a person has not yet turned into a motor plan, is represented in distributed cortical patterns that no implanted electrode array currently samples. It is not known whether a device could recover it without covering far more of the brain than any surgery would justify.

§See also

• Brain–computer interface
• Electrocorticography interfaces
• Utah array
• Neural decoding
• Neuroprosthetics
• Mental privacy
• Neurorights
• Brain-to-brain interfaces

§References