Speech neuroprosthesis gave Casey Harrell his voice at 97.5%
A speech neuroprosthesis can turn the words a paralysed person tries to say into text and into a synthesized voice. In the clearest case so far, Casey Harrell, a 45-year-old man with ALS, received 4 microelectrode arrays in the speech motor cortex of his brain in July 2023. On the first day the decoder was 99.6% accurate on a 50-word set. Over 8.4 months it held 97.5% across a vocabulary of 125,000 words. Nicholas Card, Sergey Stavisky, David Brandman and colleagues at UC Davis reported the result in the New England Journal of Medicine in August 2024. This article follows that study and two that came after it, and asks what a device that hears attempted speech can tell us about consciousness.
The short answer is that the implant reads the motor output of an intention, at 256 points out of about 86 billion neurons. It does not read feelings, and it does not read free thought. A second line of work shows why that limit is narrower than it sounds. Silent inner speech leaves a fainter copy of the same motor pattern, and in unresponsive patients the same kind of measurement can find a mind that the body cannot show.
- The Card et al. decoder reached 97.5% word accuracy on a 125,000-word vocabulary, at about 32 words per minute, over more than 248 hours of use.
- In 2025 Maitreyee Wairagkar and the same team turned the signals straight into a voice, decoded within 10 ms, with intonation for questions and stress.
- Erin Kunz and colleagues at Stanford found inner speech in motor cortex and protected it with a silent password that the decoder recognised more than 98% of the time.
A 100-second film follows the three studies on a glass brain, from the four arrays in the speech motor cortex to the decoder, the synthesized voice and the silent password. Watch Can a machine hear the words your brain tries to say? on YouTube.
How a speech neuroprosthesis decodes attempted speech
ALS destroys the motor neurons that carry commands from the brain to the muscles of the face, tongue and larynx. The cortex that plans those movements keeps working. The UC Davis team placed 4 arrays, 256 intracortical electrodes in total, in the left ventral precentral gyrus, the part of the motor cortex that controls the speech articulators. You can find that gyrus, and the precentral strip it belongs to, in the Human Brain Atlas on the MNI template.
When Harrell tries to speak, the neurons under the arrays still fire, although his muscles barely move. The decoding pipeline has three stages.
- The electrodes record spiking activity from small populations of neurons in the speech motor cortex.
- A recurrent neural network turns that activity into a probability for each speech sound (phoneme) many times per second. A small input layer, retrained each session, absorbs the day-to-day drift of the recordings.
- A language model joins the phoneme probabilities into the most likely sequence of words, and the words appear on a screen and are spoken by a computer voice built from recordings of Harrell’s voice before ALS.
The calibration was fast. After 30 minutes of recordings on the first day, the system reached 99.6% on 50 words. On the second day, with 1.4 more hours of data, it reached 90.2% on the full 125,000-word vocabulary. Harrell then used it in his own conversations.
The three studies in numbers
| Study | Participants | Signal | Output | Main result | Main limit |
|---|---|---|---|---|---|
| Card et al., NEJM 2024 | 1 (ALS, BrainGate2) | 256 electrodes, ventral precentral gyrus | Text, then a computer voice | 99.6% on 50 words on day 1, 97.5% on 125,000 words over 8.4 months, about 32 words per minute | One participant, attempted speech only |
| Wairagkar et al., Nature 2025 | 1 (same participant) | The same 256 electrodes | A synthesized voice, sound by sound | Decoded within 10 ms, about 60% of words understood by listeners against 4% without the device | Intelligibility far below natural speech |
| Kunz et al., Cell 2025 | 4 (ALS or brainstem stroke) | Microelectrode arrays in motor cortex | Text from silent inner speech | Up to 74% on a 125,000-word vocabulary, password recognised over 98% | Free inner speech still decodes with many errors |
From text to a synthesized voice
Text on a screen loses most of what a voice carries. In the second study, Maitreyee Wairagkar, Nicholas Card and colleagues built a decoder that skips words altogether. It maps the neural activity straight to sound features and drives a voice synthesizer in closed loop. The paper reports that the activity was decoded causally within 10 ms. UC Davis describes the delay to audible output as about one fortieth of a second, close to the delay a person hears between speaking and hearing their own voice.
The decoder also recovered paralinguistic features. Harrell could raise the pitch at the end of a sentence to ask a question, stress a chosen word, and sing short, simple melodies. Listeners understood almost 60% of the words he produced with the device, against about 4% of his own speech without it. The authors are clear that this is an early result in one participant.
Can a speech neuroprosthesis read your thoughts?
The third study asked the question most people ask first. Erin Kunz and colleagues recorded from motor cortex in four participants while they attempted to speak, and while they only imagined speaking. Inner speech was represented in the same motor cortex. Its pattern was highly correlated with attempted speech and weaker in magnitude, and one dimension of the activity, which the authors call motor intent, separated the two.
Imagined sentences could be decoded in real time, at up to 74% accuracy on a 125,000-word vocabulary. In some tasks the system also picked up inner speech the participants had not been asked to produce, such as numbers while they counted shapes on a screen. That is the privacy problem in concrete form. The authors describe safeguards against it. In one, the decoder stays locked until the user silently thinks a chosen keyword, “chitty chitty bang bang”, which the system recognised more than 98% of the time.
So a speech neuroprosthesis does not read thoughts in general. It reads speech-like motor patterns, and the inner speech that shares them, in one small patch of cortex.
What a speech neuroprosthesis tells us about consciousness
All three studies decode the output side of a mind. The arrays sit in motor cortex, which turns an intention to speak into commands for the muscles. What they read is the result of processing that happened elsewhere, in a person who knows what they want to say. That is a measurement of access, the information a person can report and act on. It is silent on experience, what it is like to be that person. The numbers make the scale plain, 256 electrodes against about 86 billion neurons in the human brain.
The finding still bears on two open questions on this site. The first is the substrate question that runs through the Brain Emulation section. Intention, words, and the tone of a voice leave physical patterns precise enough for a machine to learn. A position that treats a mind as a pattern of activity predicts exactly that, and every gain in decoding narrows the distance between a pattern and its readout. The bandwidth gap that the Paradromics and Neuralink comparison measures, thousands of channels against hundreds of millions of axons, is the same gap seen from the hardware side.
The second is the measurement problem. In 2024 Yelena Bodien and colleagues used task-based fMRI and EEG in 241 adults who could not follow commands at the bedside. Sixty of them, 25%, showed the brain activity of following a command. This is cognitive motor dissociation, a mind present without the motor output to show it. A speech neuroprosthesis is the same idea pushed further, a readout that bypasses the failed motor path. Detecting awareness when behaviour is absent is also what the perturbational complexity index tries to do, and the Zap and Zip console runs that index on open mouse EEG, awake and under anaesthesia.
Comparison to The Consciousness AI
The Consciousness AI studies consciousness as an emergent property of the universe, substrate independent, and asks which measurable dynamics accompany it. The speech studies do not test that position, and nothing in the project decodes human speech. They are relevant to it in one specific way. They show a working method for reading an internal state from population activity, with its accuracy, its calibration cost, and its blind spots stated in numbers. The indicator approach that the field uses for artificial systems, set out in the current state of AI consciousness research, faces the same three problems. A readout can be accurate on what it measures and still say nothing about experience. The code and the consoles of the project are open at github.com/tlcdv/the_consciousness_ai.
Limits and open questions
The speech neuroprosthesis results are strong and narrow. Two of the three studies rest on one participant, and all three use investigational devices in clinical trials. The arrays sample one small region and need recalibration as recordings drift. The voice decoder is intelligible about 60% of the time, far from natural speech. Free inner speech decodes with many errors, and the safeguards against unintended decoding are first designs. None of the studies claims to measure experience. Each gives a person who lost their voice a way to speak again, and each turns the question of how much of a mind is a readable pattern into one that can be answered with data.
Researchers covered here
- Nicholas CardUC Davis Neuroprosthetics LabFirst author of the 2024 NEJM speech neuroprosthesis study that decoded attempted speech at 97.5% accuracy across 125,000 words
- Maitreyee WairagkarUC Davis Neuroprosthetics LabFirst author of the 2025 Nature study on an instantaneous voice-synthesis neuroprosthesis
- Erin KunzStanford UniversityFirst author of the 2025 Cell study that found inner speech in motor cortex and tested a mental password for speech neuroprostheses
- Sergey StaviskyUC Davis Neuroprosthetics Lab, Department of Neurological SurgeryCo-director of the UC Davis Neuroprosthetics Lab and senior author of the speech neuroprosthesis that restored Casey Harrell's voice
- David BrandmanUC Davis Health, Department of Neurological SurgeryNeurosurgeon who implanted the speech neuroprosthesis arrays in Casey Harrell and co-directs the UC Davis Neuroprosthetics Lab