Neuralink Thought-to-Text: What the VOICE Trial Has Shown So Far
Neuralink's speech and thought-to-text work in 2026: the VOICE trial, the FDA designation, typing speeds, and how the claims compare with published research.
Neuralink, Elon Musk's brain-implant company, is now testing whether its implant can restore speech to people who have lost it. The trial is called VOICE, it began in October 2025, and its stated goal is conversation at 140 words per minute. Videos of participants producing words have circulated in 2026. What has not appeared is published accuracy data. This article sets out what is on the public record, what comes only from the company, and how it compares with the academic systems that have published their results.
What is on the record
The FDA designation. In May 2025 Neuralink announced that the US Food and Drug Administration had granted Breakthrough Device Designation "to help restore communication for individuals with severe speech impairment", including people with ALS, stroke, spinal cord injury, cerebral palsy and multiple sclerosis. A breakthrough designation gives a device priority review and more contact with the FDA. It is not an approval, and it says nothing about whether the device works.
The trial registration. The VOICE study is registered on ClinicalTrials.gov as NCT07224256. What the registry says:
- It is an early feasibility study, the first stage of human testing, with an estimated enrolment of six people.
- It started on 3 October 2025, with primary completion estimated for October 2028. The listed site is UT Southwestern in Dallas.
- Participants are adults aged 22 to 75 with severe speech impairment from ALS, primary lateral sclerosis, stroke or spinal cord injury, who also have impaired use of their arms, and communicate in English.
- The device is Neuralink's N1 implant, which is wireless and rechargeable and sits flush in the skull, with flexible electrode threads inserted by the R1 surgical robot.
- The primary outcome is safety: adverse events at 12 months. The measure of whether it works is secondary: the monthly maximum of correct words per minute.
That design is normal for a first study. It also means the trial is not set up to produce an accuracy figure you could compare with a published paper.
What Neuralink has said
In its January 2026 update, "Two Years of Telepathy", Neuralink reported:
- 21 participants implanted worldwide across its trials, most of them in the second half of 2025.
- For cursor control, some participants reaching or exceeding the 8 to 10 bits per second of an able-bodied person with a mouse.
- A "ten finger keyboard for the mind", where imagined finger movements map to letters, with typing speeds of up to 40 words per minute. The example user has ALS but can still speak.
- For VOICE, the aim of "conversational speeds of 140 words per minute" by reading signals from brain regions involved in speech production, and plans to increase implants from 1,000 to 3,000 electrodes.
- No serious device-related adverse events.
Since then, videos and posts about VOICE participants producing speech, including a September 2026 clip of a participant saying "I love you" in a reconstructed voice, have spread through press and social media. We have not been able to confirm the details of those demonstrations from a primary source. None comes with a word error rate, a vocabulary size or latency figures.
How that compares with published speech BCIs
The yardstick is the academic research, where results come with methods and numbers:
| System | Speed | Accuracy | Status |
|---|---|---|---|
| UC Davis (NEJM 2024, Nature Medicine 2026) | 56 words/min average in daily home use | 97.5% over 8 months; >99% in formal tests | Peer reviewed, one participant, ~2 years |
| Stanford (Nature 2023) | 62 words/min | 76% (24% WER) on 125,000 words | Peer reviewed, one participant |
| UCSF (Nature 2023) | 78 words/min | 75% (25% WER) on 1,024 words | Peer reviewed, one participant |
| Neuralink VOICE | Goal: 140 words/min | Not published | Company claims and videos |
Two things are worth knowing when reading about Neuralink:
- Neuralink's hardware has real advantages. It is fully implanted and wireless. The academic systems with the best numbers use wired connectors that pass through the skin, which limits life at home and carries infection risk. If Neuralink can match their accuracy without the wires, that is significant.
- The software problem is shared and largely solved in research. Decoding speech from motor cortex, turning neural activity into phonemes and then sentences with a language model, is a published recipe (explained in how speech BCIs work). Neuralink's likely challenge is not inventing a new decoder but showing reliable accuracy across participants and years.
"Thought-to-text" versus attempted speech
Headlines tend to say "thought". The trial targets attempted speech: signals from the areas that would move the mouth. In practice the line is blurred, because some participants cannot move at all and "attempted" speech then produces no movement. Research from Stanford in 2025 showed that purely imagined speech is also partly decodable from the same area, less accurately, and tested a mental password to keep private thoughts private. Whether Neuralink decodes attempted or imagined speech, and how it handles that privacy question, is one of the things its eventual publications should answer.
Neuralink is not alone
Other companies are running speech trials at the same time. Paradromics announced on 14 September 2026 that the first participant in its Connect-One trial communicated in real time with text and synthesised speech, also through a fully implanted wireless device. Synchron works with Apple's BCI input protocol for controlling iPhones and iPads. None of the three has yet published peer-reviewed speech accuracy. For the full landscape, see brain-to-text in 2026.
What to watch for
- A peer-reviewed paper or conference data with word error rate, vocabulary size and words per minute, ideally in free conversation as well as copying prompted sentences.
- How many participants reach usable accuracy, not just the best one.
- Stability over months without daily recalibration.
- Whether the voice is intelligible to people who do not know the sentence in advance.
Frequently asked questions
Can Neuralink turn thoughts into text?
Neuralink has shown participants typing up to 40 words per minute with imagined finger movements, and has released videos from its VOICE speech trial. It has not published accuracy figures for speech, and its trial targets attempted speech signals, not free-form thought.
What is the Neuralink VOICE trial?
It is an early feasibility study of Neuralink's implant for people with severe speech impairment from ALS, stroke, primary lateral sclerosis or spinal cord injury. It started in October 2025, plans to enrol six people, is registered as NCT07224256, and measures safety as its primary outcome.
Is Neuralink approved for speech?
No. In May 2025 the FDA gave it Breakthrough Device Designation for speech restoration, which speeds up review. That is not an approval. The device is only available inside clinical trials.
How fast can Neuralink users type?
Neuralink says participants have reached up to 40 words per minute with its imagined-finger keyboard, and that some exceed able-bodied mouse speeds in cursor control. Its goal for speech is 140 words per minute. These are company figures and have not been independently published.
Who has the most accurate speech brain implant?
On published results, UC Davis's system: over 97% word accuracy on a 125,000-word vocabulary, and nearly two years of daily use at home at an average of 56 words per minute, reported in NEJM in 2024 and Nature Medicine in 2026.
Sources
- Neuralink: Breakthrough Device Designation for speech (May 2025)
- Neuralink: Two Years of Telepathy (January 2026)
- ClinicalTrials.gov NCT07224256: Neuralink VOICE study
- ClinicalTrials.gov NCT07357428: Paradromics Connect-One study
- Card et al., "Long-term independent use of an intracortical BCI for speech and cursor control", Nature Medicine (2026)
- 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)
- Kunz et al., "Inner speech in motor cortex and implications for speech neuroprostheses", Cell (2025)
