Speech Neuroprosthesis
When a person tries to speak, the motor cortex still sends its signals. Implanted electrodes read them; a neural network and a language model turn them into words on a screen.
Chapters · 14
- 01From Attempted Speech to Text
- 02What Is Lost and What Is Left
- 03Where the Arrays Go
- 04One Microelectrode Array
- 05From Spikes to Features
- 06A Network Reads Out Speech Sounds
- 07A Language Model Picks the Words
- 08A Bigger Vocabulary, More Errors
- 09Training and Recalibration
- 10From Signals to a Voice
- 11Inner Speech
- 12What It Is Not Yet
- 13An Array, Taken Apart
- 14Laboratory: Vocabulary Size
Scroll
From Attempted Speech to Text
A participant with ALS could no longer speak so that others understood him. In July 2023, surgeons placed four microelectrode arrays in the precentral gyrus of his left hemisphere. On its first day of use, 25 days after surgery, 30 minutes of recordings while he tried to speak were enough to decode a 50-word vocabulary with a word error rate of 0.44%.
After 1.4 more hours of training on the second day, a 125,000-word vocabulary gave 9.8%. As data built up, it fell to 2.5% by the 15th session. Over 8.4 months he used it to talk for more than 248 hours.
He spoke at his own pace, about 32 words a minute; conversation runs at about 160. In another study published in 2023, a participant with ALS reached 62 words a minute, with a word error rate of 23.8% on 125,000 words.
50 words: 0.44% · 125,000 words: 9.8%, later 2.5% · 248 hours in all
What Is Lost and What Is Left
ALS (amyotrophic lateral sclerosis) attacks motor neurons, the nerve cells in the brain and spinal cord that control voluntary movement. As they degenerate they stop sending messages to the muscles, and speaking, swallowing and walking grow harder. People with ALS usually remain able to reason, remember and understand.
A stroke in the brainstem can leave a person locked in: awake, able to think and understand, but unable to move the body or speak. Most can blink and move their eyes up and down.
In these participants’ motor cortex, the movements of lips, tongue, jaw and larynx in speech can still be read years after paralysis; the signals just no longer reach the muscles. A speech brain–computer interface reads them from the cortex, around the break.
Motor neurons stop signalling · locked in · the cortex still carries the code
Where the Arrays Go
The precentral gyrus lies in front of the central sulcus. Its lower part, area 6v above the lateral fissure, is the motor cortex of the mouth and face; a single electrode there picks up lips, jaw, tongue and larynx intermixed. It is painted vermilion on the model.
The 2023 study placed four arrays: two in area 6v and two in area 44 of the inferior frontal gyrus, part of Broca’s area. Area 44 carried almost no information about speaking, so decoding used only the 128 electrodes in 6v.
The 2024 participant’s four arrays are all in the left precentral gyrus, in v6v, d6v, area 4 and 55b: 256 electrodes. Brains differ, and the papers say more work is needed to confirm that the speech regions can be targeted reliably.
6v: motor cortex of the face · area 44 carried little · 4 arrays, 256 electrodes
One Microelectrode Array
Each array is a small piece of silicon, 3.2 × 3.2 mm, with 8 × 8 = 64 needles standing on it, each 1.5 mm long; the maker gives the spacing as 400 µm. Needles and platform are one piece of silicon.
The shafts are insulated with Parylene-C and the tips coated with iridium oxide. Each tip carries one recording site of about 50 µm, designed to pick up one or a few nearby neurons.
In surgery, a window of about 5 × 5 cm is opened in the skull, and a high-speed pneumatic inserter pushes the array 1.5 mm into the cortex. Wires from its back run as a bundle of gold wires through the skull to a pedestal held to the outside of the skull by titanium screws. The enlarged array beside the bust is drawn at 20 : 1, the pedestal at 2 : 1.
8 × 8 needles · 3.2 mm square · 1.5 mm long · 400 µm apart
From Spikes to Features
The voltage on each electrode is recorded 30,000 times a second. When a neuron nearby fires, the voltage drops sharply for a moment.
The decoder does not try to tell neurons apart. It counts two things every 20 ms: how often the voltage drops below a threshold (−4.5 times the noise’s root mean square), and the power in the spiking band. The paper notes that, for decoding, these do about as well as sorted single neurons.
With 256 electrodes and two features each, that is 512 numbers every 20 ms. Each is standardised by its own mean and spread, smoothed, and passed to the neural network.
30,000 a second · threshold −4.5 × RMS · 512 numbers every 20 ms
A Network Reads Out Speech Sounds
The decoder is a recurrent neural network: five layers of gated recurrent units (GRU), 512 units each. Every 80 ms it reads the last 280 ms of features and gives 41 probabilities: 39 English phonemes, the pause between words, and a “blank” used in training.
Training has a catch: the participant cannot speak clearly, so nobody knows which sound he is making at each moment. The network is trained with CTC (connectionist temporal classification): it is told only the order of the sentence’s phonemes and finds the timing itself.
Signals drift from day to day, so the network starts with an input layer learned separately for each day. Taking the most likely sound at each step, the network alone got about 20% of phonemes wrong in the 2023 study.
5 layers × 512 · every 80 ms · 41 outputs · about 20% of phonemes wrong
A Language Model Picks the Words
A string of phonemes is not yet a sentence. A language model knows which words exist and which tend to go together, and finds the likeliest words given the phoneme probabilities. The 2023 study used the 125,000 words of the CMU Pronouncing Dictionary and counted three-word sequences in a large body of web text.
It takes a step every 80 ms, keeping the best candidates with a beam search, in under 1 ms a step, and the words on the screen update as it goes. The 2024 system also rescored the 100 likeliest sentences with a large language model of 6.7 billion parameters (OPT).
It corrects errors, and sometimes makes them. An example from the paper: the intended sentence was “we don’t listen to the radio at work at all”; four sounds were decoded wrongly; “listen” still came out right, but “radio” became “reader”.
125,000 words · three-word sequences · under 1 ms a step · rescored by a 6.7-billion-parameter model
A Bigger Vocabulary, More Errors
With 50 everyday words, the 2023 study’s word error rate was 9.1%; with 125,000 words it was 23.8% (on different test sentences). The network itself got about the same share of phonemes wrong in both, around 20%.
One reason: the more words, the more neighbours one sound away. In the CMU dictionary, the middle word of the 50 has about 50 other words within one sound; “I” has 142, and “eye” sounds exactly the same. Within the 50, most words have no other word one sound away.
The paper decoded the same data again with vocabularies of many sizes: only very small ones, 50 to 100 words, were clearly more accurate; by about 1,000 words the error rate was as high as with the full vocabulary.
50 words: 9.1% · 125,000 words: 23.8% · much the same above 1,000
Training and Recalibration
In the 2023 study, each test day began with the participant attempting 260 to 480 sentences (about 41 minutes), and the network was retrained on these together with all earlier data: about 140 minutes a day in all. Measured offline, without retraining the word error rate was about 30%.
The 2024 system went live on a 50-word vocabulary after 30 minutes and 213 sentences on the first day, and on 125,000 words after 1.4 more hours on the second. Later, the network was fine-tuned after every sentence he finished.
Over the 29 days when he used it only for himself, he asked three times to recalibrate because it made more errors than he was used to; each took about 7.5 minutes, attempting 20 sentences shown on the screen.
About 41 minutes a day (2023) · 30 minutes on day one (2024) · 7.5 minutes to recalibrate
From Signals to a Voice
Another team, in 2023, used electrodes that lie on the surface of the cortex without going in: 253 on one sheet, 3 mm apart. The participant was a woman who could not speak for many years after a brainstem stroke. Decoded text reached a median of 78 words a minute on a 1,024-word vocabulary, with a median word error rate of 25.5%; the system also synthesised speech and moved the face of a digital avatar.
In 2025, the 2024 participant used a decoder that speaks directly: with no text in between, it computes a short slice of sound from the neural signals every 10 ms, in a voice trained on recordings from before his illness. The paper reports under 10 ms from signal acquisition to synthesised samples, not counting the sound card’s playback.
Listeners picking his sentence out of six chose right 94% of the time on average; writing down what they heard, their median word error rate was 43.75%, against 96.43% for his own speech. He could also ask questions, stress a word, and sing simple three-pitch tunes.
253 surface electrodes · under 10 ms to sound · 43.75% when transcribed
Inner Speech
A 2025 study asked four participants to speak only in their heads, without trying to move. Their motor cortex showed signals too, like a weaker version of attempted speech; decoded in real time on 50 words, the word error rates were 14% to 33%.
That raises a question: could words only thought be decoded? The researchers tried two safeguards: training the decoder to label silently imagined sentences as “silence”, so far tested only offline; and a password, “chittychittybangbang”, imagined first before the decoder gives any output. In real-time tests the password was judged correctly in 79 of 80 trials.
The paper also states that complete, intelligible sentences could not be decoded during free-form thinking. These systems read speech-related activity in the motor cortex, not thoughts as such.
Inner speech: 14–33% on 50 words · password 79 of 80 · free thought not decoded
What It Is Not Yet
It is not a product. The studies run in clinical trials with investigational devices: by 2024 the BrainGate trials had 19 participants, most of them decoding arm and hand movements; each speech paper has one to four participants.
It takes brain surgery and leaves a pedestal through the scalp. From 2004 to 2021, BrainGate’s 14 participants had their devices for 12,203 days in all, with 68 device-related adverse events, 6 of them serious; the most common was skin irritation around the pedestal, and none required removing the device.
It reads the motor cortex during attempted speech, not thoughts; the network needs regular retraining or recalibration; and the 2024 participant still needed a researcher to connect and launch the system. A cochlear implant sends signals into the auditory nerve; a speech brain–computer interface works the other way, reading signals out of the cortex.
Clinical trials · investigational devices · brain surgery · attempted speech
An Array, Taken Apart
The enlarged array is drawn at 20 : 1. Needles and platform are one piece of silicon and do not come apart; the shafts are insulated, leaving only the iridium-oxide tips bare.
On the back, each needle has its own lead wire, sealed in silicone; the 64 leads gather into a bundle of gold wires that runs to the pedestal.
The pedestal is drawn at 2 : 1: 16.5 mm high with a base 19 mm across, screwed to the skull with titanium screws; a detachable connector and cable fit on top. The pads, the silicone and the pedestal’s outline are this page’s drawing.
Needles and platform · pads and silicone · gold wire bundle · pedestal
Laboratory: Vocabulary Size
Pick a sentence, a vocabulary and a phoneme error rate, and press “Decode”. The model first gets some sounds wrong at random (red), then picks the vocabulary word whose pronunciation is closest. Each press is a fresh run.
With 50 words, one or two wrong sounds usually still land closest to the right word. With 125,000, a single wrong sound often lands on another word, and words that sound exactly the same, like “I” and “eye”, are a coin toss.
This is a model: the phoneme error rates are the networks’ own from the papers (19.7% or 10.3%); the large vocabulary counts only neighbours one sound away; and there is no language model of context, so the large vocabulary goes wrong far more often than in the papers. Context is what brought the papers’ 125,000-word error rates down to 23.8% and lower.
Readings: wrong sounds · wrong words · the model’s average
Try thisSwitch the vocabulary from 50 words to 125,000: which words of a sentence most often come out as other words, and why do short words like “I” go wrong more?
What the model leaves out
The bust is a life-size porcelain head, not any participant. The brain’s main sulci (lateral fissure, central, precentral and postcentral sulci, superior and inferior frontal sulci, superior temporal and intraparietal sulci) follow a standard lateral view; the small folds are generated. The vermilion part is the lower precentral gyrus. The four arrays are drawn 3.2 mm square on it; “all in the left precentral gyrus” is the 2024 study’s description, and their exact places are this page’s. Two arrays share one pedestal; the wires, the pedestals’ outline, the cables and the computer are schematic. The enlarged array is at 20 : 1 and the pedestal at 2 : 1, two different scales. The travelling lights show only the order of things, not the signal’s speed or form. The sentences on the screen are the laboratory’s, all in the 50-word set.
Some easy misreadings
It reads activity in the motor cortex during attempted speech, not thoughts; during free-form thinking no complete sentences could be decoded. A word error rate counts how many words must be changed to match the intended sentence; it is not “how much was understood”. The 50-word and 125,000-word results used different test sentences, so they cannot be divided one by the other. The “97.5% accuracy” in the 2024 abstract is a 2.5% word error rate on prompted sentences; in free conversation it was 3.7%. Each paper has one to four participants, all speaking English, and all these systems are still in clinical trials.
Common questions · 5
- How accurate was a speech brain–computer interface on its first day?
On its first day of use, 25 days after surgery, 30 minutes of recordings while he tried to speak were enough to decode a 50-word vocabulary with a word error rate of 0.44%.
01From Attempted Speech to Text ↑- Do people who can no longer speak still have speech signals in the brain?
In these participants’ motor cortex, the movements of lips, tongue, jaw and larynx in speech can still be read years after paralysis; the signals just no longer reach the muscles.
02What Is Lost and What Is Left ↑- How does the neural network of a speech brain–computer interface work?
The decoder is a recurrent neural network: five layers of gated recurrent units (GRU), 512 units each.
06A Network Reads Out Speech Sounds ↑- Why does a bigger vocabulary make more decoding errors?
With 50 everyday words, the 2023 study’s word error rate was 9.1%; with 125,000 words it was 23.8% (on different test sentences).
08A Bigger Vocabulary, More Errors ↑- Can a speech brain–computer interface read thoughts?
The paper also states that complete, intelligible sentences could not be decoded during free-form thinking.
11Inner Speech ↑
Sources
- Willett et al. (2023): A high-performance speech neuroprosthesis, Nature 620, 1031–1036
- Card et al. (2024): An Accurate and Rapidly Calibrating Speech Neuroprosthesis, N Engl J Med 391, 609–618
- Wairagkar et al. (2025): An instantaneous voice-synthesis neuroprosthesis, Nature 644, 145–152
- Metzger et al. (2023): A high-performance neuroprosthesis for speech decoding and avatar control, Nature 620, 1037–1046
- Kunz et al. (2025): Inner speech in motor cortex and implications for speech neuroprostheses, Cell 188, 4658–4673
- Moses et al. (2021): Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria, N Engl J Med 385, 217–227
- Rubin et al. (2023): Interim Safety Profile From the Feasibility Study of the BrainGate Neural Interface System, Neurology 100, e1177–e1192
- NINDS: Amyotrophic Lateral Sclerosis (ALS)
- NINDS: Locked-in syndrome
- ClinicalTrials.gov: NCT00912041 (BrainGate2)
- Blackrock Neurotech: Utah Array (specifications)
- Carnegie Mellon University: The CMU Pronouncing Dictionary 0.7b
Data
The bust is this page’s porcelain head at life size, not any participant. The hemisphere’s main sulci are drawn after a standard lateral view; the small folds are generated. The four arrays (64 electrodes, 3.2 × 3.2 mm, 1.5 mm long) sit on the left precentral gyrus as Card et al., NEJM 391, 609 (2024) describe; their exact places, the wire bundles, the pedestals’ outline (CerePort pedestal: 16.5 mm high, 19 mm base, Blackrock Neurotech’s specifications), the cables and the computer are drawn. The decoder’s sizes (features every 20 ms, a 5-layer GRU with 512 units, 41 outputs every 80 ms, CTC) and the language model (125,000 words, trigram) are from Willett et al., Nature 620, 1031 (2023) and its Supplementary Methods; the example of Fig. 2d is quoted from it. The laboratory is this page’s model: the 50-word set of Moses et al. (2021) as used by Willett et al. (2023); pronunciations from the CMU Pronouncing Dictionary 0.7b (Copyright © 1993–2015 Carnegie Mellon University, BSD-style licence); phoneme error rates of 19.7 % and 10.3 % from Willett et al. (2023) and Card et al. (2024); the error mix and the nearest-word rule are the model’s. Material presets, parts and the softbox studio of the site’s model kit are ported from AirsupHQ/machine-skill (MIT, © 2026 Air Supply Technologies Inc.).
The text, pictures and models on this page are licensed CC BY-NC 4.0: free to use and adapt for teaching and other non-commercial purposes, with credit to world.lapla.si. Third-party material listed above keeps its own licence.
Cite this plate
Laplace’s Mini World, Plate F59, “Speech Neuroprosthesis”, https://world.lapla.si/en/speechbci/, CC BY-NC 4.0, updated 2026-10-28.
Follow the question
Plate 53How the Cochlea Takes Sound ApartSound entering the ear ends up in a coiled tube. A membrane inside it is pushed into a wave that swells highest near the entrance for high notes and far along for low ones. A few thousand hair cells turn that motion into nerve signals.
Plate F03MRI ScannerA clinical MRI scanner: a superconducting magnet, gradient coils and RF coils, one inside another. Hydrogen nuclei precess in the field; their signal fills k-space, and a Fourier transform turns it into an image.
Plate F13Attention in TransformersA two-layer, eight-head Transformer reads a sentence in your browser, one character at a time, and guesses the next. Each thread is one act of attention; its brightness is the weight.
Plate F22Embeddings: The Geometry of MeaningA black-lacquered, gilt-lined globe set with tiny pictures and brass name tags. Each is a direction from a small model: a picture and its caption point almost the same way.
Plate F18How a Neural Network LearnsAn aluminium relief milled to a network’s real error, a steel ball that steps down it by the real gradient, and a small network in glass and brass: light runs forward with the numbers and back with each weight’s share of the blame.
