How Does a Speech Neuroprosthesis Work?

By neuraspeak editorial · Updated 2026-10-07 · 4 min read

A speech neuroprosthesis is an implanted brain-computer interface that records activity from the parts of the cortex that control the lips, tongue, jaw and larynx, then uses machine-learning decoders to turn a person's attempted speech into text, a synthesized voice or an animated avatar. In the published systems, the user tries to speak and the device decodes the movement commands the brain still sends, even though paralysis keeps them from reaching the muscles.

What part of the brain does a speech neuroprosthesis read?

Most speech neuroprostheses record from the speech motor cortex, the strip of sensorimotor cortex that plans and sends commands to the muscles used for speaking. In conditions such as amyotrophic lateral sclerosis (ALS) or a brainstem stroke, this cortex can keep generating speech commands even after the pathway to the muscles has failed. The device picks up those commands at their source.

The main research groups target slightly different spots. In the Stanford study published in Nature in 2023, Willett and colleagues placed four microelectrode arrays in a woman with ALS: two in the ventral premotor area known as 6v and two in area 44, part of Broca's area. Nearly all of the useful speech information came from area 6v. The UC Davis team that published in the New England Journal of Medicine (NEJM) in 2024 put four arrays, 256 electrodes in total, into the left ventral precentral gyrus of a man with ALS. The UCSF group led by Edward Chang lays a sheet of surface electrodes across the speech sensorimotor cortex; its 2023 Nature study used 253 electrodes.

How are the brain signals recorded?

Brain signals are recorded either by tiny electrodes that go into the cortex or by a thin electrode sheet resting on its surface. These are the two hardware families behind every high-performance speech result so far.

Penetrating microelectrode arrays, such as the Utah-style arrays used by the BrainGate consortium, sit a millimeter or so into the cortex. They pick up spikes from individual neurons or small groups of neurons. Electrocorticography (ECoG) grids sit on the brain's surface without penetrating it and record the summed activity of larger populations of neurons. Decoders lean heavily on the high-gamma frequency band of that activity.

In most academic studies to date, the implant connects to a pedestal on the skull, and a cable runs from the pedestal to external computers. That is one reason use has mostly been limited to supervised sessions. Fully implanted wireless systems from companies such as Paradromics and Neuralink are now in early feasibility studies for communication, but neither company had published peer-reviewed speech results as of October 2026.

How does the decoder turn neural activity into words?

The decoder is a trained neural network that maps short windows of brain activity to speech sounds, usually phonemes, and then hands those predictions to a language model that picks the most likely words. Working at the level of phonemes lets the system produce words it never saw during training, which is how vocabularies grew from 50 words to more than 125,000.

Training the system starts with calibration. The participant sees a sentence on a screen and attempts to say it while the system records the neural activity. A recurrent neural network learns which activity patterns go with which phonemes. A language model then sorts the noisy stream of phoneme probabilities into real words and sentences. The UC Davis system in the 2024 NEJM paper combined an n-gram language model with a large language model for this step.

Calibration has become much shorter. In the UC Davis study, 30 minutes of recordings on the first day of use gave 99.6% word accuracy on a 50-word vocabulary. After 1.4 more hours of training on day two, the system reached 90.2% accuracy on a 125,000-word vocabulary. With continued training, accuracy held at 97.5% over 8.4 months.

How does the system produce a voice rather than text?

There are two approaches to producing a voice: decode text first and send it to a text-to-speech engine, or synthesize sound directly from the neural signals. The text route is simpler and very accurate. The UC Davis system reads its decoded sentences aloud in a voice modeled on recordings of the participant made before ALS affected his speech.

Direct synthesis is newer. UCSF's 2023 Nature study decoded text, synthesized audio personalized to the participant's voice before her injury, and drove a digital avatar's facial movements. A 2025 follow-up in Nature Neuroscience by Littlejohn and colleagues streamed speech in 80-millisecond increments. In that study, the first sound came within about a second of the attempt to speak, compared with roughly eight seconds in the earlier system. Also in 2025, a UC Davis study in Nature reported an intracortical system that turned neural activity into voice in about 25 milliseconds. The participant could change his intonation and sing simple melodies. Listeners understood almost 60% of the synthesized words, compared with 4% of his unassisted speech.

Can a speech neuroprosthesis read private thoughts?

Current systems are trained to decode attempted speech, not free-form thought, but research shows that inner speech can also leave a decodable trace in motor cortex. In a study published in Cell in August 2025, Kunz and colleagues recorded from four participants with ALS or brainstem stroke. They found that imagined speech produced patterns similar to attempted speech, only weaker. The team decoded imagined sentences with up to 74% accuracy.

The same study found that some unprompted inner speech, such as counting, could be partially decoded. That raises a privacy concern for future devices. The researchers tested two safeguards. One trains the decoder to ignore inner speech. The other unlocks decoding only when the user imagines a chosen password phrase, which the system recognized with more than 98% accuracy.

What are the main limitations today?

Speech neuroprostheses are still investigational. Each major result comes from one or a few participants, the devices need brain surgery, and none has FDA marketing approval. Error rates on large vocabularies have dropped quickly, but performance varies between people, and neural signals can drift from day to day, so decoders have to adapt.

Long-term home use is just beginning to be documented. In June 2026, Nature Medicine reported on the UC Davis participant, who used his system nearly every day at home for almost two years: more than 3,800 hours, at an average of about 56 words per minute. This is the strongest evidence so far that the technology can work outside the lab, but it comes from one person with one device design.

The bottom line

A speech neuroprosthesis records speech commands from the motor cortex, decodes them into phonemes with a neural network, and assembles words with a language model, then outputs text or a synthesized voice. The best research systems now reach high accuracy on vocabularies of more than 125,000 words, but they remain experimental and have been tested in very few people.

Questions

Glossary: Speech neuroprosthesisNeural decoderPhoneme decodingElectrocorticography (ECoG)Utah arrayNeural signal drift

Sources

1. A high-performance speech neuroprosthesis (Willett et al., Nature) · 2023-08-23
2. A high-performance neuroprosthesis for speech decoding and avatar control (Metzger et al., Nature) · 2023-08-23
3. An Accurate and Rapidly Calibrating Speech Neuroprosthesis (Card et al., NEJM) · 2024-08-15
4. A streaming brain-to-voice neuroprosthesis to restore naturalistic communication (Littlejohn et al., Nature Neuroscience) · 2025
5. An instantaneous voice-synthesis neuroprosthesis (Wairagkar et al., Nature) · 2025
6. Inner speech in motor cortex and implications for speech neuroprostheses (Kunz et al., Cell) · 2025-08-14
7. Long-term independent use of an intracortical brain-computer interface for speech and cursor control (Nature Medicine) · 2026-06-15

Facts on this page are drawn from the sources listed and were accurate to the best of our knowledge on the update date shown. Company and product names are trademarks of their owners and are used for identification only. Nothing here is medical, investment or legal advice. Companies and individuals can request a correction via the corrections page.

Read next