UCSF brain-computer interface decodes speech and body gestures simultaneously from a single cortical implant

By Inside BCI · September 18, 2026

Two people with paralysis, one from a brainstem stroke at age twenty and one from ALS diagnosed six years ago, spoke and gestured through a virtual avatar in the same session, in real time, from a single implant on the surface of their brain. Nature Neuroscience published the result on 14 September 2026 from Samantha C. Brosler, Jessie R. Liu, Alexander B. Silva and senior author Edward F. Chang at the University of California, San Francisco. The paper introduces the first single-implant, single-session, simultaneous real-time decoding of speech and upper-body gesture in people with paralysis.

What the participants could do

Two BRAVO-trial participants completed the full avatar experiments. Bravo-1r is a man in his early forties who lost intelligible speech and most upper-body function after a pontine brainstem stroke at age twenty. He has been in the UCSF BRAVO trial since 2019, originally implanted with a 128-channel ECoG array; his device was replaced with a 253-channel high-density array in April 2024. Bravo-6 is a man in his early sixties diagnosed with ALS at age fifty-six. Dysarthria began at fifty-seven and progressed to complete loss of intelligible speech within nine months; he was implanted with a 253-channel array in March 2025.

Both drove a personalized virtual avatar that produced speech and upper-body gestures at the same time, from two decoders running in parallel on the same array. The gestures decoded across the two participants included wave, nod, hand shake, clap, shrug, thumbs up, and fist pump. Bravo-1r’s vocabulary was five phrases, four gestures, twelve speech-and-gesture combinations. Bravo-6’s vocabulary was ten phrases, ten gestures, one hundred speech-and-gesture combinations.

The real-time conversation paradigm was the most demanding condition. Bravo-1r reached a median 100 percent accuracy on both the speech and the gesture decoder across three conversational blocks, against chance rates of 16.7 percent for speech and 20 percent for gesture. Bravo-6 reached 85 percent on the gesture decoder and 75 percent on the speech decoder.

A third participant, Bravo-3, contributed data to the earlier isolated multi-effector recordings but withdrew from the trial before the avatar experiments began. The paper does not state the reason.

The design finding that changes multi-effector BCI training

The most important finding is not the accuracy number. It is that neural signals produced when a participant attempts speech and gesture simultaneously are not the linear sum of the isolated speech signal and the isolated gesture signal. They are a distinct multimodal pattern. Decoders trained on simultaneous data outperform decoders trained on isolated data across all behavioural contexts.

Practically, this means any BCI aiming at multi-effector communication cannot train a speech decoder alone and a gesture decoder alone and expect them to work in combination. Multi-effector output requires multi-effector training data. Without cross-training, false-positive rates on the opposite modality were 76 percent for Bravo-1r and 31 percent for Bravo-6. Cross-training reduced both to zero.

Edward Chang framed the significance: “Conversation is about much more than the words being spoken. It’s a multilayered, dynamic process involving the whole motor cortex. This proof-of-concept shows us it’s possible for a BCI to restore some of this freedom and flexibility.”

Samantha Brosler, one of three co-first authors, framed the field context: “But most of these studies have focused on restoring one of these functions at a time.”

Where this fits in the Chang lab BRAVO sequence

The UCSF Chang lab has now published four peer-reviewed BRAVO-trial communication-BCI papers in five years. Moses et al. (NEJM, July 2021) decoded a 50-word vocabulary from a 128-channel ECoG in Bravo-1 after brainstem stroke. Metzger et al. (Nature, August 2023) decoded 78 words per minute from a 253-channel high-density ECoG in Bravo-3, with a 1,024-word general-English vocabulary, driving an avatar with facial expression. Littlejohn et al. (Nature Neuroscience, 2025) reported a streaming brain-to-voice neuroprosthesis on the same Bravo-3 array with lower latency. Brosler et al. (Nature Neuroscience, September 2026) adds simultaneous speech and gesture decoding on Bravo-1r and Bravo-6, from a single 253-channel high-density ECoG.

The scaling story is coherent. Vocabulary expanded from fifty words to over a thousand between 2021 and 2023. Output modality expanded from text to synthesised voice to voice-plus-face-avatar between 2021 and 2023. Latency dropped to streaming between 2023 and 2025. Now the output has expanded to voice-plus-face-plus-upper-body-gesture without any additional implanted hardware.

The hardware footprint

The array is a 253-channel high-density ECoG surface implant manufactured by PMT Corporation of Chanhassen, Minnesota, with 3 millimetre pitch and 1 millimetre contact diameter, placed subdurally on the pial surface of the left hemisphere at UCSF Medical Center. Percutaneous pedestal connectors, secured to the skull in the same operation, carry signals from the array to an external Blackrock Microsystems CerePlex E256 digital headstage.

The system is wired and uses a transcutaneous pedestal. It is not a Neuralink Threads array. It is not a Paradromics Connexus microelectrode. It is not a Synchron Stentrode. The surgical footprint and regulatory pathway for ECoG surface implants are meaningfully different from penetrating microelectrode and endovascular platforms.

Chang has stated that the next step is a fully implantable wireless successor. No commercial licensee for the UCSF Chang system is publicly named at time of publication.

Where this lands in the invasive speech-decoding field

The dominant competitor result at the invasive speech-decoding frontier is Stanford BrainGate, most recently Card et al. (NEJM 2024) and Card et al. (Nature Medicine 2026), using Utah microelectrode arrays. Willett et al. (Cell 2020) demonstrated compositional hand-knob body representation on the same platform. Wairagkar et al. (Nature 2025) reported an instantaneous voice-synthesis result.

The narrow claim that survives adversarial pressure is not that no one has decoded arm movement and speech before. Stanford BrainGate has done both, in separate participants and separate sessions on Utah arrays. What holds is single-implant, single-session, simultaneous, real-time decoding of both modalities from one cortical implant. The design finding that simultaneous training beats isolated training is the generalisable result.

Whether the same pattern holds on Utah-style penetrating microelectrode arrays, on Paradromics Connexus, or on Synchron Stentrode is an empirical question. The relevant labs have the participants and the platforms to test it.

Not disclosed

The reason Bravo-3 withdrew from the trial before the avatar experiments began is not stated in the paper’s public materials. Specific NIH grant numbers were not visible in the primary source materials accessed at publication. Chang’s competing-interests disclosure for this specific paper should be read directly from the Nature Neuroscience author-info block before any commercial statements are made.

What to watch

Whether UCSF Chang lab and PMT Corporation publish or announce a fully implantable wireless successor to the 253-channel array. Chang has stated the wireless system is the near-term next step.

Whether Stanford BrainGate publishes a simultaneous speech-and-gesture result on Utah microelectrode arrays. The design finding on simultaneous training is testable on their existing participants.

Whether Paradromics, Synchron, Precision Neuroscience, Neuralink, or CorTec report simultaneous multi-effector decoding results on their respective platforms in the coming quarters. The paper reframes multi-effector BCI decoders as a training-data problem in addition to a hardware-channel-count problem.

Whether an independent motor-cortex neuroscience result surfaces on why multimodal signals are not the linear sum of isolated modality signals. The neural interaction pattern is a fundamental motor-cortex finding, not only a BCI engineering result.

Whether a commercial licensee is named for the UCSF Chang system in the next twelve months as the wireless successor moves toward clinical use.

Sources

Primary:

Prior UCSF Chang lab BRAVO papers:

Competitor invasive speech-decoding work:

Cross-reference to prior InsideBCI coverage:

research UCSF Edward-Chang Samantha-Brosler Jessie-Liu Alexander-Silva Karunesh-Ganguly BRAVO-trial ECoG high-density-ECoG PMT-Corporation Blackrock-Neurotech speech-decoding gesture-decoding multi-effector-BCI Nature-Neuroscience brainstem-stroke ALS avatar wireless-BCI