A stroke took her speech in 2005. An implant now turns her attempts to speak into a synthesised voice. It arrives from a speaker about a second after she starts trying to talk, built to sound like her own, reconstructed from a short clip of her speaking at her wedding before the brainstem stroke left her unable to move the muscles that produce speech. She does not mouth the words. An electrode array resting on the surface of her brain picks up the signals that would have gone to her lips, tongue, jaw and larynx, and a model converts them into audio in 80 millisecond increments.

The system was described by Kaylo Littlejohn, Cheol Jun Cho and colleagues in a paper titled “A streaming brain-to-voice neuroprosthesis to restore naturalistic communication”, published in Nature Neuroscience on 31 March 2025, with Edward Chang of UCSF and Gopala Anumanchipalli of UC Berkeley as joint senior authors. The participant, known in the trial as BRAVO-3 and referred to publicly as Ann, is a former high school mathematics teacher from Canada who had her stroke in 2005 at the age of 30. She is enrolled in the BRAVO clinical trial at UCSF, registration NCT03698149.

This is one study, not settled consensus.

What the implant actually reads

The array itself is not new. It was placed during an earlier phase of the same trial, and UCSF described it in 2023 as a paper-thin rectangle carrying 253 electrodes, positioned over the speech sensorimotor cortex. What changed in the 2025 paper is the software sitting behind it.

A distinction gets lost in most coverage, and it concerns what the signal represents. This decoder is not reading thoughts or inner monologue. Cho, one of the co-lead authors, described the approach in the UC Berkeley release as intercepting signals after the decision to speak has already been made, at the point where intention becomes motor control.

If she does not attempt to articulate, there is nothing to decode.

Because she has no usable residual vocalisation, the team had no recordings of her attempted speech to train against. They generated synthetic target audio using a text-to-speech model, then conditioned a voice conversion step on a pre-injury clip so the output resembled her. Ann told the UCSF team in 2023 that hearing it was “like hearing an old friend”.

What “almost in real time” means in the numbers

An earlier version of this system produced roughly an eight second delay for a single sentence. That is long enough to make ordinary conversation impossible, and it is the problem the paper sets out to address.

In the extended data, the authors report a median latency of 1.67 seconds between the go cue on screen and the onset of synthesised speech on a 1,024-word general sentence set, with a 99 per cent confidence interval of 1.57 to 1.81 seconds. On a smaller 50-phrase set, median latency was 2.61 seconds. Anumanchipalli’s often-quoted line in the same announcement, that “within 1 second, we are getting the first sound out”, starts the clock elsewhere: at the detected speech intent signal, not the visual cue. Both figures are accurate. They measure separate intervals.

Offline, applied to unprompted blocks where she attempted free-form silent speech, streaming synthesis ran at a median 46.5 words per minute against 24.8 for the older delayed method. Accuracy is harder to summarise. In one comparison condition, where text was decoded and then spoken by an off-the-shelf voice engine instead of being synthesised directly, the median word error rate on the 1,024-word set was 40.8 per cent for the audio and 30.7 per cent for the text.

What the paper does not show

Nothing here demonstrates that the method works for anyone else. One participant, one implantation site, one lesion type. The team’s own analysis of training volume found meaningful gains between 8.6 and 29.4 hours of recorded data and none between 29.4 and 58.4 hours, which says something about the calibration burden and nothing about how any of it transfers to a second brain.

Expressive speech is also absent. Pitch, loudness and intonation are not decoded, and Littlejohn has described that as ongoing work, an unsolved problem in audio synthesis generally and not only in neuroprosthetics.

The device is investigational, wired to a bank of computers, and not approved by any regulator. Access to the underlying material is partial. The ECoG neural recordings are deposited in the Harvard Dataverse, and the decoding code is published to a GitHub repository named in the paper. Wider clinical trial data sits behind restricted request, and her personalised voice model is withheld to protect her anonymity, with a generic voice supplied in its place.

Competing interests are worth reading alongside the results: several authors hold pending patents on the decoding methods, Edward Chang is a co-founder of Echo Neurotechnologies, and David Moses, the UCSF neurological surgery researcher who co-managed the project, is a director there.

What has happened since March 2025

This paper is now more than a year old, and the field has moved in ways that clarify what it did and did not settle.

In June 2025, Maitreyee Wairagkar and colleagues at UC Davis published an instantaneous voice-synthesis neuroprosthesis in Nature, working with a man with ALS and using 256 penetrating microelectrodes in the ventral precentral gyrus rather than a surface array. Their system decoded some paralinguistic features, allowing the participant to change intonation and sing short melodies. Different lab, different hardware, similar direction of travel.

More consequential, in our reading, is a Nature Medicine paper published on 15 June 2026 by Nicholas Card and colleagues, reporting on the same UC Davis participant using a brain-computer interface at home for more than 3,800 hours across nearly two years with no researchers present, at an average of 56 words per minute. It decodes text, not voice. What it does address is the bottleneck that latency figures obscure entirely: whether any of this works on a Tuesday afternoon with no research team in the room.

What to watch

Three things. Whether a streaming voice decoder replicates in a second BRAVO participant, since everything published so far rests on one person. Whether a wireless version reaches trial, which both teams have named as a priority. And whether accuracy holds up outside prompted sentence sets, where the language model has less to work with.

The BRAVO trial is still running and BrainGate2, registration NCT00912041, is still enrolling. Neither group has yet published a streaming voice system used independently at home, and that is the result that would change what these devices are for.