Speech neuroprosthesis: revision 1 to 4

Revision 1

Revision 4

type: "technology"

status: "experimental"

horizon: "2030s"

trl: 5

trl: 6

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: "2026-07-27"

updated: "2026-10-30"

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: "Peer-reviewed results come from single implanted participants or groups of up to four, mostly with ALS or brainstem stroke; one man has used a system at home for nearly two years, and no efficacy trial has reported."

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

access: "Not obtainable: research implants at academic centres and in small company feasibility studies, all investigational, with no approved or purchasable system as of October 2026."

reversibility: "difficult"

issues: ["The 2024 UC Davis result is given as accuracy only; no words-per-minute figure is stated."]

issues: ["Neuralink's 75 wpm figure rests on a company post about a paywalled conference abstract that was not read directly.", "Paradromics' September 2026 speech claims are sourced only to a local news report with no speed figure."]

---

```infobox

{ "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": "Company trials", "value": "Neuralink, Paradromics (from 2025)" },

{ "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. 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.

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

[

  { "value": "~15 wpm", "label": "First sentence decoding", "note": "2021, 50-word vocabulary, roughly a quarter of words wrong" },

  { "value": "~15 wpm", "label": "First sentence decoding", "note": "2021, 50-word vocabulary, roughly a quarter of words wrong" },

{ "value": "62 wpm", "label": "Penetrating-array result", "note": "2023, Stanford, 125,000-word vocabulary" },

{ "value": "78 wpm", "label": "Surface-array result", "note": "2023, UCSF, with synthesized voice and avatar" }

{ "value": "78 wpm", "label": "Surface-array result", "note": "2023, UCSF, with synthesized voice and avatar" },

{ "value": "56 wpm", "label": "Average rate in home use", "note": "2026 report, one man with ALS, nearly two years" }

]

```

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 intelligibly for years; the map does not disappear when the output pathway

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

fails.[^willett2023] 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

> Raw phoneme decoding is error-prone, with about one phoneme in five wrong in the 2023 Stanford

> performance comes from the language model choosing the most probable sentence consistent with a

> system, and the language model often corrected those errors to recover the intended word.[^willett2023]

> noisy neural signal. This is legitimate engineering — human listeners do the same thing — but it

> A substantial share of performance comes from the language model choosing the most probable

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

> sentence consistent with a noisy neural signal. This is legitimate engineering, and human listeners

> unpredictable utterances are decoded worse than the aggregate figure suggests.

> 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.[^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

word accuracy near 98 percent over a very large vocabulary in a participant with ALS, with the

97.5 percent word accuracy over a 125,000-word vocabulary, sustained across 8.4 months in a

system usable from the first session and improving over months of daily self-directed

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

use.[^card2024] Subsequent work has attacked latency: streaming architectures that emit words as

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

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

result.[^card2024] Later work went after latency. The UCSF group built a streaming decoder that

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

synthesizes speech in the participant's pre-injury voice in 80-millisecond increments rather than at

the end of a sentence.[^littlejohn2025] 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.[^wairagkar2025]

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

In 2026 the same UC Davis team reported long-term independent home use. Over nearly two

motor cortex produced roughly 90 characters per minute in one participant, which remains the

years the participant used the system for more than 3,800 hours at home with no researchers present,

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

communicating about 1.96 million words at an average of 56 words per minute and holding a full-time

design parameter.[^willett2021]

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.[^card2026]

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.[^willett2021]

## Company trials

Two companies opened early feasibility studies of fully implanted wireless speech devices:

[[neuralink|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.[^voice2025] 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.[^reuters2025]

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.[^neuralinkspeech] 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.[^neuralinkneals2026] 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.[^neuralinkhack2026]

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.[^paradromics2025] 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.[^michigan2026] 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.[^cbsaustin2026]

> [!debate] 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."[^musk2026] 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.[^willett2023][^neuralinkspeech]

> Peer-reviewed decoding has reached a median 78 with a quarter of words wrong,[^metzger2023] the

> longest home use averaged 56,[^card2026] and Neuralink gives its own median participant about

> 50,[^neuralinkspeech] 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

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. In one man it has shown nearly two years of independent

home use.[^card2026]

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. The only long-term

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

home data come from one participant; the authors name the single participant as a limitation and note

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

that they did not systematically assess fatigue.[^card2026] There are still almost no data on noisy

percutaneous connector to laboratory hardware.

settings or changing medication. Most work remains in English, although

one team decoded both Spanish and English from a bilingual participant[^silva2024] and a Chinese

group decoded Mandarin syllables, tones included, in real time.[^qian2025] 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.[^card2026]

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, 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).[^kunz2025]

> 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,[^boccato2026] 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.[^fogg2026] 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 — 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 decoded instructed inner speech from the

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

same electrodes with word error rates of 26-54 percent on a large vocabulary, but could not decode

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

complete, intelligible sentences during free-form thinking. It also demonstrated a keyword-gated

password.[^kunz2025]

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.[^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. The ethicist Marcello Ienca places speech decoding within the narrow sense of mental

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

privacy, the decoding of present thoughts, which he distinguishes from inferences about traits and

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

risks drawn from brain data.[^ienca2026] 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 near-term path is a pivotal trial in ALS, the population with the clearest need and the

The company studies are early feasibility trials, the stage before any pivotal trial, and both

strongest existing results, using a fully implanted wireless system rather than a percutaneous one.

enrol people with ALS among other causes of severe paralysis, the population with the clearest need and the strongest existing

Progress there depends less on decoding accuracy, which is already adequate, than on hardware

results.[^voice2025][^paradromics2025] Progress there arguably depends less on decoding accuracy,

longevity, surgical throughput, and the unglamorous work of making a device that a clinician can

which the academic results suggest is already adequate for conversation, than on hardware longevity,

implant and a family can support at home — the same constraints that shape every implanted

surgical throughput, and the unglamorous work of making a device that a clinician can implant and a

[[brain-computer-interface]], from the endovascular [[stentrode]] to the penetrating threads of

family can support at home. The same constraints shape every implanted

[[neuralink]].

[[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 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 output does not separate the user's

their own contribution from the model's.

contribution 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

[^littlejohn2025]: `paper` Littlejohn, K. T. et al. "A streaming brain-to-voice neuroprosthesis to restore naturalistic communication." *Nature Neuroscience*, 2025.

[^littlejohn2025]: `paper` Littlejohn, K. T. et al. "A streaming brain-to-voice neuroprosthesis to restore naturalistic communication." *Nature Neuroscience*, 2025.

[^kunz2025]: `paper` Kunz, E. M. et al. "Inner speech in motor cortex and implications for speech neuroprostheses." *Cell*, 2025. {The password-gated mode is a laboratory demonstration in already implanted research participants, not a safeguard validated in any deployed device.}

[^wairagkar2025]: `paper` Wairagkar, M. et al. "An instantaneous voice-synthesis neuroprosthesis." *Nature* 644, 145-152, 2025. https://doi.org/10.1038/s41586-025-09127-3

[^card2026]: `paper` Card, N. S. et al. "Long-term independent use of an intracortical brain-computer interface for speech and cursor control." *Nature Medicine* 32, 2504-2510, 2026. https://doi.org/10.1038/s41591-026-04414-6 {One participant, T15; the 92 percent figure is his own rating of his sentences, which the authors could not verify independently.}

[^kunz2025]: `paper` Kunz, E. M. et al. "Inner speech in motor cortex and implications for speech neuroprostheses." *Cell*, 2025. {The keyword lock was tested in real time with one already implanted participant, T12; it is a laboratory demonstration, not a safeguard validated in any deployed device.}

[^boccato2026]: `paper` Boccato, T., Olak, M. and Ferrante, M. "Cross-subject decoding of human neural data for speech brain computer interfaces." *Journal of Neural Engineering*, 2026. {A reanalysis of the two largest public intracortical speech datasets, so its own sample is four participants; the authors name that as the study's fundamental limitation.}

[^fogg2026]: `preprint` Fogg, Z. M. et al. "A generalizable speech neuroprosthesis." *bioRxiv*, 2026. https://doi.org/10.64898/2026.07.23.739430

[^silva2024]: `paper` Silva, A. B. et al. "A bilingual speech neuroprosthesis driven by cortical articulatory representations shared between languages." *Nature Biomedical Engineering* 8, 977-991, 2024. https://doi.org/10.1038/s41551-024-01207-5

[^qian2025]: `paper` Qian, Y. et al. "Real-time decoding of full-spectrum Chinese using brain-computer interface." *Science Advances* 11, eadz9968, 2025. https://doi.org/10.1126/sciadv.adz9968 {Several authors are employees of Neuroxess, the company that supplied the implant used in the study.}

[^ienca2026]: `paper` Ienca, M. "Two kinds of mental privacy: what speech decoding does and does not threaten." *Hastings Center Report* 56(5), 25-27, 2026. https://doi.org/10.1002/hast.70116

[^voice2025]: `trial` Neuralink Corp. "VOICE: An Early Feasibility Study of a Precise Robotically Implanted Brain-Computer Interface for Communication Restoration." ClinicalTrials.gov NCT07224256, first posted 4 November 2025. https://clinicaltrials.gov/study/NCT07224256

[^reuters2025]: `news` Reuters. "Neuralink's speech restoration device gets FDA's 'breakthrough' tag." 5 May 2025, syndicated by Tuoi Tre News. https://news.tuoitre.vn/neuralinks-speech-restoration-device-gets-fdas-breakthrough-tag-103250505161133851.htm

[^neuralinkspeech]: `statement` Neuralink. "Speech Restoration." Clinical trials page, accessed 30 October 2026. https://neuralink.com/trials/speech-restoration/

[^neuralinkneals2026]: `statement` Neuralink. Post on X reporting early speech restoration data, 28 October 2026. https://x.com/neuralink/status/2115534593147187397 {A company summary of a conference abstract, not a peer-reviewed paper; the figure is one participant's first-week result.}

[^neuralinkhack2026]: `statement` Neuralink. "A 72-Hour Engineering Sprint for Translating Thought to Speech." Company blog, 29 October 2026. https://neuralink.com/updates/72-hour-voice-hackathon/

[^musk2026]: `statement` Musk, E. Post on X quoting Neuralink's hackathon thread, 30 October 2026. https://x.com/elonmusk/status/2116018780111597911

[^paradromics2025]: `statement` Paradromics. "Paradromics Receives FDA Approval for the Connect-One Clinical Study with the Connexus Brain-Computer Interface." Press release, 20 November 2025. https://paradromics.com/news/paradromics-receives-fda-approval-for-the-connect-one-clinical-study-with-the-connexus-brain-computer-interface/

[^michigan2026]: `statement` Michigan Medicine. "University of Michigan implants first-in-human Paradromics wireless brain-computer interface, designed to restore communication." News release, 17 June 2026. https://www.michiganmedicine.org/news-release/university-michigan-implants-first-human-paradromics-wireless-brain-computer-interface-designed

[^cbsaustin2026]: `news` Maciborski, W. "'I have a lot to say': Austin brain implant helps patient communicate in real time." CBS Austin, 14 September 2026. https://cbsaustin.com/news/local/i-have-a-lot-to-say-austin-brain-implant-helps-patient-communicate-in-real-time {The report relies on the company and the implanting surgeon and quotes no independent expert.}