Speech neuroprosthesis: revision 2 to 3

Revision 2

Revision 3

trl: 5

categories: ["cybernetics"]

tags: ["bci", "speech decoding", "als", "anarthria", "neural decoding", "communication"]

summary: "A brain–computer interface that reconstructs intended speech from motor cortex activity, producing text or synthesized voice for people who cannot speak."

summary: "A brain-computer interface that reconstructs intended speech from motor cortex activity, producing text or synthesized voice for people who cannot speak."

updated: "2026-08-23"

updated: "2026-10-29"

humanEvidence: "Published results come from one or a few implanted participants, mostly with ALS or brainstem stroke; no multi-participant efficacy trial has been run and no device is approved anywhere."

humanEvidence: "Published results come from one or a few implanted participants, mostly with ALS or brainstem stroke; no multi-participant efficacy trial has been run and no device is approved anywhere."

access: "Not obtainable: a handful of research implants at academic centres under investigational device exemptions, with no approved or purchasable system."

access: "Not obtainable: a handful of research implants at academic centres under investigational device exemptions, with no approved or purchasable system."

reversibility: "difficult"

{ "label": "Recording site", "value": "Ventral sensorimotor cortex" },

{ "label": "Sensors used", "value": "ECoG grids, penetrating arrays" },

{ "label": "First sentence decoding", "value": "2021" },

{ "label": "Best reported rate", "value": "~60–80 words per minute" },

{ "label": "Best reported rate", "value": "~60-80 words per minute" },

{ "label": "Participants per study", "value": "One to a few" },

{ "label": "Regulatory status", "value": "Investigational" }

]

}

```

**Speech neuroprosthesis** is a brain–computer interface that recovers intended speech from

**Speech neuroprosthesis** is a brain-computer interface that recovers intended speech from neural

neural activity and renders it as text, synthesized voice, or an animated face. It is aimed at

activity and renders it as text, synthesized voice, or an animated face. It is aimed at people who

people who retain the intention and the neural machinery for speech but have lost the motor

retain the intention and the neural machinery for speech but have lost the motor pathway: those with

pathway: those with amyotrophic lateral sclerosis, brainstem stroke, or severe anarthria from other

amyotrophic lateral sclerosis, brainstem stroke, or severe anarthria from other causes. Between 2021

causes. Between 2021 and 2025 the field moved from decoding a fifty-word vocabulary at a

and 2025 the field moved from decoding a fifty-word vocabulary at a conversational crawl to

conversational crawl to near-real-time output over vocabularies of a hundred thousand words.

near-real-time output over vocabularies of a hundred thousand words.

Crucially, these systems decode attempted articulation, not thought. The signals come from the

These systems decode attempted articulation rather than thought. The signals come from the region of

region of cortex that would drive the lips, jaw, tongue, and larynx; the decoder recovers a motor

cortex that would drive the lips, jaw, tongue and larynx. The decoder recovers a motor plan, and

plan, and language models convert that noisy plan into plausible sentences. That distinction

language models turn that noisy plan into plausible sentences. The distinction governs both what the

governs both what the technology can do and what privacy risks it does and does not create.

technology can do and which privacy risks it does and does not create.

```keyfacts

[

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

people who have been unable to speak for years; the map does not disappear when the output pathway

fails. That persistence is the enabling fact for the whole field.

fails. The whole field depends on that persistence.

A speech neuroprosthesis records from that cortex, extracts features (high-gamma power for

A speech neuroprosthesis records from that cortex and extracts features: high-gamma power for

[[ecog-interfaces]], threshold crossings and band power for a penetrating [[utah-array]]), and

[[ecog-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

passes them to a sequence model that emits phonemes or subword units. A language model then converts

converts the phoneme stream into words, using the statistics of English to resolve ambiguity that

the phoneme stream into words, using the statistics of English to resolve ambiguity that the neural

the neural signal alone leaves open. These are among the few clinical research systems in which a

signal alone leaves open. These are among the few clinical research systems in which a statistical

statistical language model of the kind central to debates about

language model of the kind central to debates about [[artificial-general-intelligence]] sits inside

[[artificial-general-intelligence]] sits inside a control loop with a person's nervous system. The

a control loop with a person's nervous system. The user attempts to speak, silently or with whatever

user attempts to speak, silently or with whatever residual movement remains, and the pipeline runs

residual movement remains, and the pipeline runs in near real time.

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. This is the same [[neural-decoding]] architecture used for

moves with the decoded articulation. The underlying [[neural-decoding]] architecture is the same one

cursor control, applied to a much higher-dimensional output space.

used for cursor control, applied to a much higher-dimensional output space.

> [!key] 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 — human listeners do the same thing — but it

> noisy neural signal. This is legitimate engineering, and human listeners do the same thing. It

> means the reported word error rate is a property of the whole system, and that unusual or

> 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.[^moses2021] Slow, error-prone, and unmistakably speech.

wrong.[^moses2021] 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

with ALS reported 62 words per minute over a 125,000-word vocabulary, with a word error rate near 24

24 percent — and under 10 percent when the vocabulary was restricted to fifty words.[^willett2023] A

percent, falling under 10 percent when the vocabulary was restricted to fifty words.[^willett2023] 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, together with a

stroke, reported a median 78 words per minute over a 1,024-word vocabulary, along with a synthesized

synthesized voice built from pre-injury recordings and an avatar that moved its

voice built from pre-injury recordings and an avatar that moved its face.[^metzger2023]

face.[^metzger2023]

A 2024 report from a UC Davis and Brown collaboration pushed accuracy rather than speed, reporting

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

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. The rate in self-paced

participant with ALS, with the system usable from the first session. In self-paced conversation the

conversation was about 32 words per minute over more than 248 cumulative hours — slower than either

rate was about 32 words per minute over more than 248 cumulative hours, slower than either 2023

2023 result.[^card2024] Subsequent work has attacked latency: streaming architectures that emit

result.[^card2024] Later work has gone after latency, with streaming architectures that emit words

words as they are formed rather than at the end of a sentence, and voice synthesis fast enough to

as they are formed rather than at the end of a sentence, and voice synthesis fast enough to preserve

preserve

conversational turn-taking and some prosody.[^littlejohn2025]

A related line uses imagined handwriting rather than speech. Decoding attempted pen strokes from

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, which remains the

motor cortex produced roughly 90 characters per minute in one participant. That remains the fastest

fastest character-level result and demonstrates that the choice of imagined movement is itself a

character-level result, and it shows that the choice of imagined movement is itself a design

design parameter.[^willett2021]

parameter.[^willett2021]

## What has and has not been shown

Shown: fluent, large-vocabulary decoding in individual participants; stable enough performance for

The published work has shown fluent, large-vocabulary decoding in individual participants,

daily use over months; voice synthesis with recognisable identity; decoding from both surface and

performance stable enough for daily use over months, and voice synthesis with recognisable identity.

penetrating electrodes; usable results from the first day of calibration in at least one case.

It has shown decoding from both surface and penetrating electrodes, and in at least one case usable

results from the first day of calibration.

Not shown: any of this in more than a handful of people. Nearly every result above comes from a

None of this has been shown in more than a handful of people. Nearly every result above comes from a

single participant, and different participants have different implant sites, different disease

single participant, and participants differ in implant site, disease stage and residual

stages, and different residual capabilities. There is no multi-participant efficacy trial, no

capabilities. There is no multi-participant efficacy trial and no approved device. There are no data

approved device, no data on how the systems perform in noisy real-world settings with fatigue and

on how the systems perform in noisy real-world settings with fatigue and changing medication, and

changing medication, and almost no work outside English. Most systems remain tethered through a

almost no work outside English. Most systems remain tethered through a percutaneous connector to

percutaneous connector to laboratory hardware.

laboratory hardware.

A further limitation is that these decoders address speech *production*. They do not help people

These decoders also address only speech *production*. They do not help people whose language

whose language comprehension or formulation is impaired, which excludes many stroke and dementia

comprehension or formulation is impaired, which excludes many stroke and dementia patients. A

patients. A [[cochlear-implant]] restores an input channel; a speech neuroprosthesis restores an

[[cochlear-implant]] restores an input channel and a speech neuroprosthesis restores an output

output channel; neither touches the linguistic system between them.

channel, but neither touches the linguistic system between them.

Within the wider field of [[neuroprosthetics]], speech decoding is unusual in producing an output

Within the wider field of [[neuroprosthetics]], speech decoding is unusual in producing output that

that is unambiguously interpretable. Efforts at cognitive prosthetics further upstream — the

can be interpreted without ambiguity. Cognitive prosthetics further upstream, such as the

hippocampal work described under [[memory-prosthesis]] — report effects that are small, contested,

hippocampal work described under [[memory-prosthesis]], report effects that are small, contested,

and difficult to evaluate precisely because there is no equivalent of a transcript to score.

and hard to evaluate, precisely because they have no equivalent of a transcript to score.

> [!caution] Every headline figure is one person's decoder

> The 2021, 2023 and 2024 results each come from a single implanted participant, and the largest

> The 2021, 2023 and 2024 results each come from a single implanted participant. The largest group

> group in any speech-decoding paper is the four BrainGate2 participants — T12, T15, T16 and T17 —

> in any speech-decoding paper is the four BrainGate2 participants (T12, T15, T16 and T17), reported

> reported together in 2025.[^kunz2025] Each decoder is trained on that individual's own attempted

> together in 2025.[^kunz2025] Each decoder is trained on that individual's own attempted speech and

> speech and does not transfer: models trained jointly across participants have appeared only

> does not transfer. Models trained jointly across participants have appeared only recently, and

> recently, and they still align each new person's cortex into the shared space with a per-subject

> they still align each new person's cortex into the shared space with a per-subject

> transform.[^boccato2026] 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

## Inner speech and privacy

The obvious worry — that a device might transcribe thoughts a person did not choose to utter — is

The obvious worry is that a device might transcribe thoughts a person did not choose to utter. That

neither science fiction nor imminent. Attempted speech and imagined speech produce overlapping but

is neither science fiction nor imminent. Attempted speech and imagined speech produce overlapping

distinguishable patterns in motor cortex, and a 2025 study reported that imagined speech is

but distinguishable patterns in motor cortex. A 2025 study reported that imagined speech can be

decodable from the same electrodes, while also demonstrating a keyword-gated mode designed so that

decoded from the same electrodes, and also demonstrated a keyword-gated mode in which the decoder

the decoder ignores neural activity until the user internally produces a chosen

ignores neural activity until the user internally produces a chosen password.[^kunz2025]

password.[^kunz2025]

That is a reassuring engineering answer to a real problem, and it establishes the shape of the

That is a reassuring engineering answer to a real problem, and it shows the shape of the issue: the

issue: the safeguard has to be designed in, because the signal is there whether or not the system

safeguard has to be designed in, because the signal is present whether or not the system is asked to

is asked to use it. The broader questions belong to [[mental-privacy]] and the developing

use it. The broader questions belong to [[mental-privacy]] and the developing [[neurorights]]

[[neurorights]] frameworks — who holds the neural recordings, whether decoded text has legal status

frameworks. They include who holds the neural recordings, whether decoded text has legal status as

as speech, and what happens when a decoder outputs a sentence the user did not intend. Error

speech, and what happens when a decoder outputs a sentence the user did not intend. Error correction

correction in a speech prosthesis is not a cosmetic matter; a wrong word attributed to a person who

in a speech prosthesis matters for more than appearance, because a wrong word attributed to a person

cannot verbally repudiate it is a distinctive harm.

who cannot verbally repudiate it is a distinctive harm.

## Outlook

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 that shape every implanted

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 sitting between cortex and the world, doing

[[human-ai-merger]]. A machine-learning model sits between cortex and the world, doing inferential

inferential work the user's damaged nervous system cannot, with the user unable to distinguish

work that the user's damaged nervous system cannot, and the user cannot tell their own contribution

their own contribution from the model's.

from the model's.

The scientific question underneath is how much of language is reachable this way at all. Decoding

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

converted into a motor plan, is represented in distributed cortical patterns that no implanted

turned into a motor plan, is represented in distributed cortical patterns that no implanted

electrode array currently samples, and it is not known whether a device could recover it without

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