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.

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 for years; the map does not disappear when the output pathway fails. 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 accuracy is well below the headline word accuracy. 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.¹ 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.² 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 has gone after latency, with streaming architectures that emit words as they are formed rather than at the end of a sentence, and voice synthesis fast enough to preserve conversational turn-taking and some prosody.⁵

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. That remains the fastest character-level result, and it shows that the choice of imagined movement is itself a design parameter.⁶

§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.

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. There are no data on how the systems perform in noisy real-world settings with fatigue and changing medication, and almost no work outside English. Most systems remain tethered through a percutaneous connector to laboratory hardware.

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 and 2024 results each come from a single implanted participant. The largest group in any speech-decoding paper is the four BrainGate2 participants (T12, T15, T16 and T17), reported together in 2025.⁷ Each decoder is trained on that individual's own attempted speech and does not transfer. Models trained jointly across participants have appeared only recently, and they still align each new person's cortex into the shared space with a per-subject transform.⁸ BrainGate2 is a registered feasibility study, 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 reported that imagined speech can be decoded from the same electrodes, and also demonstrated a keyword-gated mode in which the decoder ignores neural activity until the user internally produces a chosen password.⁷

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. They 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 near-term path is a pivotal trial in ALS, the population with the clearest need and the strongest existing results, using a fully implanted wireless system rather than a percutaneous one. Progress there depends less on decoding accuracy, which is already adequate, 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 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 user cannot tell their own 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