Brain-to-Text in 2026: Where Thought-to-Speech Really Stands
Brain-to-text and brain-to-speech in 2026, with sources: implants restoring speech in ALS, Neuralink's VOICE trial, Meta's non-invasive decoders, and the limits.
In 2026 a man with ALS has used a brain implant at home, without researchers present, for nearly two years, writing close to two million words at an average of 56 words per minute. Other implanted systems now give people who cannot speak a synthetic version of their own voice with a delay of a few milliseconds. Several companies have begun speech trials, and Meta has shown that sentences can be partly decoded without surgery at all. This overview explains what has actually been achieved, what is still a claim, and what the technology cannot do, with links to the original papers throughout.
The short version
- Implanted speech BCIs work, for a small number of people, in research settings. The best systems decode attempted speech into text with word error rates in the low single digits on large vocabularies.
- Voice synthesis straight from the brain works in real time, but the peer-reviewed voice is still often hard to understand.
- Inner speech, words only imagined, can be partly decoded from the same implants. Free-form thought cannot.
- Non-invasive decoding (no surgery) has moved from picking words from a list to open sentences, but only in a lab scanner and with far more errors.
- Company speech trials (Neuralink, Paradromics) have started, and have shown demonstrations, but have published no accuracy figures.
Who these systems are for
A speech brain-computer interface (BCI) is built for people who have lost the ability to speak through paralysis: ALS (motor neurone disease), brainstem stroke, or spinal cord injury. In most of them, the brain still plans speech normally; the signal just never reaches the muscles. An implant placed over the part of the motor cortex that controls the lips, tongue, jaw and larynx can record those plans, and software translates them into text or sound.
The benchmark is not a keyboard; it is natural conversation, about 150 to 160 words per minute. The participants in these studies were typically communicating at a handful of words per minute before the implant, using eye-tracking or a head mouse.
Milestones with the numbers
| Year | Team | Result | Source |
|---|---|---|---|
| 2021 | UCSF (Chang) | 50-word vocabulary, 15 words/min, 26% WER, brainstem stroke | NEJM |
| 2023 | Stanford (Willett) | 62 words/min, 24% WER on a 125,000-word vocabulary, ALS | Nature |
| 2023 | UCSF (Metzger) | 78 words/min, 25% WER on 1,024 words, plus a talking avatar | Nature |
| 2024 | UC Davis (Card) | 97.5% word accuracy sustained over 8 months, 125,000 words, ALS | NEJM |
| 2025 | UCSF/Berkeley (Littlejohn) | Streaming voice and text in 80 ms steps | Nature Neuroscience |
| 2025 | UC Davis (Wairagkar) | Voice synthesised within 10 ms, with intonation and singing | Nature |
| 2025 | Stanford (Kunz) | Inner speech decoded; a mental password to protect privacy | Cell |
| 2026 | UC Davis (Card) | ~2 years of independent home use, 56 words/min, >99% accuracy in tests | Nature Medicine |
| 2026 | Meta (Lévy, Zhang) | Non-invasive MEG: 29% character error; v2 with LLM: 39% word error | Nature Neuroscience, preprint |
WER means word error rate: the share of words substituted, missed or added. Lower is better. Comparing across rows needs care, because vocabulary size and the language model used change the WER enormously.
Implants: from lab demos to daily use
The biggest story of 2026 is durability. Casey Harrell, the participant in UC Davis's 2024 study, has four small electrode arrays (256 electrodes) in his speech motor cortex. The June 2026 follow-up in Nature Medicine reports more than 3,800 hours of use at home over nearly two years: 183,060 sentences, used for texting, email and a full-time job. He rated 92% of his sentences as at least mostly correct, and in formal tests word accuracy was above 99%. The decoder recalibrates continuously in the background, so it does not need a daily training session.
Speed and voice are the other frontier. UCSF and Berkeley's 2025 streaming system decodes in 80-millisecond steps rather than waiting for a full sentence. UC Davis went further: a 2025 Nature paper converts neural activity directly to sound within 10 ms, with no text in between, so the participant hears himself as he speaks and can change pitch to ask a question. The honest number: listeners transcribing that synthetic voice got 43.75% of words wrong, better than his own speech (96% wrong) but far from easy listening. A July 2026 preprint reports 5.24% on the same benchmark data with a new model. That result has not been peer reviewed.
Accuracy is also starting to generalise across people. A July 2026 preprint trained one decoder on six participants and, with fewer than 200 sentences from a new person, reached a word error rate below 7%. Until now, every participant needed their own model trained from scratch.
For how these decoders work, from electrodes to phonemes to sentences, see how speech brain-computer interfaces work.
Inner speech and mental privacy
If an implant can decode speech you attempt, can it decode speech you only imagine? Stanford's 2025 Cell paper, with four participants, found that inner speech is clearly represented in the same motor cortex, as a weaker version of attempted speech. On a 50-word vocabulary, imagined sentences were decoded in real time with 14% to 33% word error; on a large vocabulary, 26% to 54%. Complete sentences during free, unprompted thinking could not be decoded.
The same paper tested two safeguards: training the decoder to output nothing for inner speech (tested offline), and a mental password, "chittychittybangbang", which the participant imagines to switch decoding on. The password was recognised correctly 98.75% of the time. These are early answers to a question every speech BCI will have to settle.
Companies: trials started, numbers not yet published
- Neuralink received FDA Breakthrough Device Designation for speech in May 2025 and started its VOICE trial in October 2025, aiming for 140 words per minute. It has shared videos and company posts but no peer-reviewed speech data. We cover what is and is not known in Neuralink and thought-to-text.
- Paradromics announced on 14 September 2026 that the first participant in its Connect-One trial, a woman with motor neurone disease, communicated in real time as text and synthesised speech, including a phone call with her grandchildren. The device is fully implanted and transmits wirelessly. No accuracy figures have been released.
- Synchron, whose Stentrode is placed through a blood vessel instead of open surgery, became in 2025 the first BCI to use Apple's new BCI Human Interface Device protocol, letting a participant with ALS operate an iPad by thought. That is selection and navigation, not speech decoding, but it is the first sign of brain interfaces being supported by a mainstream operating system.
Without surgery: real progress, real limits
The strongest non-invasive result is Meta's Brain2Qwerty. It records brain activity with MEG, a magnetic scanner in a shielded room, while volunteers type. The peer-reviewed version reached a 29% character error rate. A second version from June 2026 adds a large language model and decodes whole English sentences with 39% word error. EEG, the only wearable option, roughly doubled the errors. Details are in Meta Brain2Qwerty explained.
fMRI-based decoders from UT Austin and NTT can recover the gist of a story someone hears or a video they recall, but not the exact words. They need hours in a scanner and the person's cooperation. Meta's EMG wristband, which lets Ray-Ban Display wearers write by moving a finger, reads muscle signals, not thoughts. The full picture is in can AI read your thoughts.
What to watch next
- Peer-reviewed company data. The first published accuracy figures from Neuralink or Paradromics will show whether wireless, fully implanted devices match the wired academic systems.
- Cross-person decoders. If a new user needs minutes rather than hours of training, the technology becomes practical for many more people.
- Intelligible real-time voice. The gap between the 43.75% peer-reviewed result and the 5.24% preprint is the most important number to watch.
- Locked-in users. A 2026 Cell Reports study found decoding works above chance but is less accurate for people with no remaining speech movement at all, the group that needs it most.
- Wearable MEG. Newer optically pumped magnetometers could take Meta-style decoding out of the shielded room.
Frequently asked questions
Is brain-to-text technology real?
Yes, for a small number of people with paralysis who have electrode arrays implanted in their speech motor cortex. The best systems decode attempted speech into text on a 125,000-word vocabulary with over 97% word accuracy, and one participant has used his daily for nearly two years. Everything else, including non-invasive and consumer devices, is research.
How fast can a brain implant produce text?
Published systems range from about 30 to 78 words per minute in tests, and one participant averages 56 words per minute in daily home use. Natural speech is about 150 to 160 words per minute. Neuralink has stated a goal of 140 words per minute but has not published results.
Can a brain implant read thoughts?
Not in the everyday sense. Implants decode speech a person attempts, and partly speech they deliberately imagine. A 2025 Stanford study could not decode complete sentences during free-form thinking and tested a mental password so that imagined speech is only decoded when the user chooses.
Can brain-to-text work without surgery?
Partly. Meta's Brain2Qwerty decodes typed sentences from MEG recordings with about 61% of words correct on average, but it needs a scanner in a magnetically shielded room, the person has to be typing, and it does not work in real time. Wearable EEG is much less accurate.
When will speech BCIs be available to patients?
No speech BCI is approved for general use. Neuralink's VOICE and Paradromics' Connect-One are early feasibility trials with a handful of participants, designed to measure safety. Approval for wider use is years away.
Sources
- Card et al., "Long-term independent use of an intracortical BCI for speech and cursor control", Nature Medicine (2026)
- Card et al., "An Accurate and Rapidly Calibrating Speech Neuroprosthesis", NEJM (2024)
- Willett et al., "A high-performance speech neuroprosthesis", Nature (2023)
- Metzger et al., "A high-performance neuroprosthesis for speech decoding and avatar control", Nature (2023)
- Moses et al., "Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria", NEJM (2021)
- Littlejohn et al., "A streaming brain-to-voice neuroprosthesis to restore naturalistic communication", Nature Neuroscience (2025)
- Wairagkar et al., "An instantaneous voice-synthesis neuroprosthesis", Nature (2025)
- Kunz et al., "Inner speech in motor cortex and implications for speech neuroprostheses", Cell (2025)
- Jude et al., Cell Reports (2026): speech decoding in locked-in ALS
- Neuralink: Breakthrough Device Designation for speech
- ClinicalTrials.gov NCT07224256: Neuralink VOICE trial
- ClinicalTrials.gov NCT07357428: Paradromics Connect-One
- Lévy et al., "Noninvasive decoding of typed sentences from human brain activity", Nature Neuroscience (2026)
