Neuralink ALS Speech Data: 75 Words per Minute, an iPhone and a Grok Agent, Claimed vs Verified

Short answer: On October 28, 2026 Neuralink reported that one participant in its VOICE speech-restoration study reached 75 words per minute at a 5.8% word error rate on a 125,000-word vocabulary during their first week of use. A day later, social posts said a man with ALS can make phone calls and operate an iPhone by thought, and that Neuralink is prototyping a Grok agent triggered by mental commands. These are company-reported or secondhand figures from a single participant, with no peer-reviewed paper. Promising, but not independently verified.

TL;DR: claimed versus verified

What Neuralink actually posted

The headline number comes from Neuralink's own post on X on October 28: "Early Neuralink speech restoration data: We are enabling speech to text and voice at conversational rates in participants with ALS. One participant achieved 75 WPM at 5.8% WER on a 125k-word vocabulary in their first week." A fan account, Neura Pod, amplified it and compared it with average adult speech of around 150 words per minute on a 40,000-word vocabulary; that comparison is the account's, not Neuralink's. Other outlets said the data was presented at a neuromuscular disease meeting.

What we could not find is the underlying dataset: no preprint, no clinical report, no information on how many sentences were tested, whether the sentences were prompted or free, or how word error rate was scored. A 5.8% error rate on a large vocabulary is notable, but it depends heavily on the task. Reading prompted sentences is easier than open conversation.

The iPhone and Grok details

A green key fitting a narrow opening in a fence, illustrating permission limits for a thought-triggered Neuralink agent

The next day, reports said Neuralink spent 72 hours at a man's home testing a system that converts neural signals into text for mobile apps and passes brain-generated speech into phone calls, without a button press or an unlocked phone. They also said a prototype Grok agent can run computer tasks triggered by specific mental commands.

We treat all of that as a reported demonstration. We found no Neuralink technical post describing the phone interface, the latency, or how the agent is permissioned. That matters, because an agent that acts from neural signals raises the same permission questions as any AI agent with account access, plus new ones about accidental triggers.

What is verifiable: the VOICE trial

The VOICE study is registered on ClinicalTrials.gov as NCT07224256. An independent evidence review by Managing Expectations, which read the registry on September 23, 2026, summarizes it this way:

• Early feasibility study of a robotically implanted brain-computer interface for communication restoration; Neuralink is the sponsor.
• Single group, open label, estimated enrollment of six; not a randomized efficacy trial.
• Primary outcomes are device- and procedure-related adverse events at 12 months, with preliminary speech performance followed up to 48 months.
• Estimated primary completion October 2028 and study completion October 2031. These are registry estimates, not promises.
• Neuralink states its devices are investigational and not approved by the FDA or other regulators.

Notebookcheck's September report adds that the study has FDA Breakthrough Device Designation, which speeds review but is not approval, and that Neuralink had implanted more than 20 patients worldwide, two of them in VOICE at the time.

How the numbers compare with earlier work

Speech brain-computer interfaces did not start with Neuralink. Academic groups have reported decoding speech attempts at tens of words per minute with different electrodes and participants. Reports on the Neuralink numbers cite a 2024 academic study with roughly 97.5% word accuracy over 8.4 months at about 32 words per minute. The new figure is higher on both speed and vocabulary size, but it uses different hardware, tasks and participants, so we would not call it a like-for-like win.

What to watch for in a real comparison: the number of participants, whether performance holds beyond week one, the test conditions (prompted versus free speech), and whether results are reported with confidence intervals.

Where AI fits

Decoding intended speech from neural activity is a machine learning problem: a model trained with the participant maps recorded signals to words. Neuralink has described a 50,000-hour pretraining effort for a brain-computer interface foundation model, and the voice output side is a speech model. In September Neuralink credited Grok Voice for the voice output in a demonstration, per the Managing Expectations review, while noting the full architecture and per-component performance were not established.

Elsewhere, researchers are decoding images from fMRI with AI; see our post on Weizmann's one-hour mind-reading image reconstruction. Together these show how quickly neural decoding is improving, and why neural data privacy deserves attention now.

How to read a brain-computer interface accuracy claim

Three numbers are easy to confuse, and each hides different assumptions.

Words per minute measures speed of output, not understanding. A system that outputs 75 words per minute with frequent corrections can feel slower than one that outputs 50 cleanly. Ask whether the figure counts only correct words, whether pauses between sentences are included, and whether the participant is reading prompts or composing their own.

Word error rate counts substitutions, deletions and insertions against a reference transcript. A 5.8% rate means roughly one word in seventeen is wrong, which is understandable in context but noticeable in a long message. The rate depends on the vocabulary size and on whether a language model is helping to fix mistakes. With a 125,000-word vocabulary and a strong language model behind the decoder, a few errors can be silently repaired, which flatters the result without being dishonest.

Time in use matters most. Many implants perform best early and then drift as the body responds to the electrodes or as recording quality changes. A first-week figure is encouraging, but a six-month figure from the same participant is the number that predicts whether this helps in daily life.

The privacy question nobody has answered

A device that reads signals tied to intended speech produces neural data, and policy on that data is still being written. Questions worth asking any BCI developer: Is raw neural data stored, and for how long? Is it used to train models that serve other participants? Can a participant delete it? Who can access it if the company is acquired? Neuralink says its devices are investigational and the trial is registered, but we have not seen a public data-handling policy for VOICE participants. When a prototype agent is added on top, the question widens to what the agent can see on the phone and what it can do without confirmation.

What a good independent test would look like

A credible follow-up would report results for every VOICE participant, not only the best one. It would include free conversation as well as prompted sentences, publish the decoder's error profile, and show performance over months. An outside lab replicating the evaluation protocol on shared test sentences would settle most of the open questions. Until then, the most defensible reading is that Neuralink has shown a striking early result in one person, consistent with the direction of academic speech-BCI research, and that everything beyond that is still to be demonstrated.

Why caution is warranted

• One participant. A first-week result from one person tells us what is possible, not what is typical.
• Company-produced. The numbers come from the sponsor, without independent audit.
• Short follow-up. Implant performance can change over months as signal quality shifts.
• Not a cure. Even the best case restores communication; it does not alter the disease.
• Hype risk. Headlines like "thought-controlled iPhone" blur a supervised trial demo with a consumer product.

Questions we cannot answer yet

1. How many participants are in the speech cohort and what are their individual results?
2. What exactly was the 5.8% error rate measured on?
3. How does the phone and agent interface handle accidental or unintended commands?
4. Who has access to the neural data, and what are the retention rules?
5. When will a paper or full dataset appear?

What to do with this information

If you follow assistive technology, bookmark the registry entry and watch for a preprint. If you build voice or agent products, note that low-latency, personalized speech models and permissioned agents are exactly the pieces being combined here; our coverage of AssemblyAI's text-to-speech model looks at the voice-output side. If you or a family member are considering a trial, speak to a clinician and the trial team rather than relying on social posts.

Related reading

• Neuralink pretrains a BCI foundation model on 50,000 hours
• Weizmann Brain-IT: fMRI image reconstruction
• Grok Bot Vault: saved logins and cards
• Is ChatGPT Dots safe to leave unattended?
• AssemblyAI TTS-1

Sources: Neuralink on X, Neura Pod on X, Managing Expectations evidence review, Notebookcheck, ClinicalTrials.gov NCT07224256.

Figures are company-reported as of October 29, 2026 and may change. This post is not medical advice.