Casey Harrell knew exactly what he wanted to say. For years, the problem was getting it out. Amyotrophic lateral sclerosis — ALS — had slowly taken the muscles he used to speak, until only his wife and his assistant could fully understand him. Then researchers at the University of California, Davis, connected a computer to tiny sensors implanted in his brain, and the words started coming back — in a voice built from recordings of his own, before the disease changed it.
New results from that team, reported by The New York Times on November 14 and due to be presented this week at the Society for Neuroscience conference in Washington, go further. The researchers say refined software can now make the decoded words far easier to understand — and a second patient, a woman who could no longer speak intelligibly at all, has used a similar implant to vary her tone of voice and even sing a musical scale. It is early, preliminary work. But it is the clearest picture yet of how a brain implant can return something ALS takes away: not just words, but the way words are said.
How the implant reads a word before it is spoken
The device is called a brain-computer interface, and it works by skipping the broken part of the chain. In ALS, the motor neurons that carry commands from the brain to the muscles die. The intention to speak survives; the pathway out does not. The implant sits on top of that intention.
In July 2023, neurosurgeon David Brandman implanted four microelectrode arrays — each about as wide as a peppercorn — into the speech motor cortex on the left side of Harrell's brain, the strip of tissue that sends commands to the jaw, tongue, larynx and lips. Together the arrays carry 256 electrodes that listen to the electrical activity that arises when he tries to talk. Software trained on his particular neural patterns translates that activity into the 39 phonemes — the sound units — that make up every word in American English, then assembles the phonemes into words and sentences, and a synthesised voice speaks them aloud.
The phoneme approach is what makes the system scalable. The team did not have to teach it every word: on the first day it was switched on, in August 2023, Harrell reached 99.6 per cent accuracy on a 50-word vocabulary after just 30 minutes of training. The dictionary was then expanded to 125,000 words — essentially every word an English speaker is likely to need. The voice that speaks them is modelled on recordings Harrell made before ALS affected his speech — TechTimes reports his young daughter barely remembered his original voice — so his family hears something close to the voice they knew.
It is one of the most striking examples yet of artificial intelligence working inside medicine — a field where patients, as our earlier report on patients' distrust of AI medical advice found, remain deeply cautious about trusting machine judgement with their health.
It is important to be precise about what the implant does not do. It does not read thoughts. Harrell has to try to say something; the software recognises the pattern of that attempt. It does not cure ALS, and it does not restore natural speech — his own vocal cords produce nothing. What it restores is communication.
Two million words at home
The most striking evidence is not a laboratory demonstration but a life lived with the device. In a study published in Nature Medicine in June 2026, the UC Davis team reported that Harrell had used the implant at home for more than 3,800 hours over nearly two years — roughly five and a half hours a day, every day, with no researcher in the room. In that time he communicated more than 183,000 sentences, close to two million words, at an average pace of 56 words a minute: about a third of the speed of ordinary conversation, but enough for real exchanges. He kept working full-time as a climate campaigner, took video calls, sent emails signed "Sent from my brain", played Wordle with his wife and daughter, and made jokes — including, when asked whether he could swear with the device, answering that he most certainly could. The team's Sergey Stavisky described him as "the first power user of a speech brain-computer interface."
That long record matters because earlier brain implants mostly worked in short, supervised sessions. A device that only functions while scientists are watching is a demonstration. One that works at home, for years, is a tool.
What is new this week — and what is still unproven
The new results come in two parts, and they need to be kept separate. First, the team reanalysed Harrell's data from a 2025 study in Nature, in which his brain signals were converted directly into expressive speech with a delay of about 10 milliseconds — fast enough to hear his own synthesised voice as he spoke, and to sing simple three-pitch melodies. The problem then was clarity: listeners could accurately transcribe only a little more than half of the words he produced. With refined algorithms, the researchers now report that the intelligibility of the words could be increased by as much as 94 percent. That is a reanalysis of existing data, not a live test — and as brain-computer interface researcher Mariska Vansteensel of University Medical Center Utrecht noted, the improvement still needs to be tested in real use.
Second, a new participant — a woman identified only as T21 — received an implant with 384 electrodes. She could no longer speak intelligibly when she entered the study. According to the researchers, the system allowed her to produce features of speech such as questioning or joyful intonation, and to sing a musical scale. Those results are preliminary and have not yet been published.
The singing matters more than it might sound. Intonation — the rise at the end of a question, the emphasis on a single word, the lift of a joke — is what turns decoded words into a person speaking. In the 2025 study Harrell could modulate his voice to ask questions with 90.5 per cent accuracy and emphasise specific words with 95.7 per cent accuracy. A system that only produces flat text is a typewriter; one that carries tone is a voice.
What this means — and what it does not
For all the promise, the honest limits are stark. The long-term evidence comes from a single participant. The device is investigational, tested inside the BrainGate2 clinical trial, and it requires external equipment and help with setup. It has not been shown to work equally well for everyone with ALS, and there is no version of it a patient can buy. Researchers say the system needs to be studied in more people, and made smaller and easier to use, before it can help beyond a handful of trial participants.
But the direction of travel is now visible. A decade ago, a speech brain implant was a laboratory curiosity. Today one man has spoken nearly two million words through his, at home, in his own voice — and this week, a second person sang a scale through hers. The technology is still at what one of its pioneers calls "the very beginning." The beginning, at least, now has a soundtrack.
This article is general information only and is not medical advice.
