University of California, San Francisco and UC Berkeley
Verified evidenceCreator to judgeAttempted-speech brain signals can produce real-time text, audio, and facial animation as a faster channel.
Clinical operations · California, United States
Collections: Human still decides · Regulated autonomy
Executive brief
A breakthrough accessibility outcome can coexist with material error, invasiveness, and single-user limits; all belong in the case.
AI value · Communication rate
VerifiedMedian real-time decoding rate of 78.3 words per minute.
Single-participant investigational system, not an approved general product.
Before
Uses head tracking to select letters on a commercial assistive device. → Reads the slowly composed text.
After
Attempts to speak while 253 cortical electrodes capture speech-motor signals. → Decodes text and synthesizes speech and avatar movements in real time.
Human boundary
The participant owns intended communication and decides use; clinical teams control the investigational device and safety.
Why it matters
Attempted-speech brain signals can produce real-time text, audio, and facial animation as a faster channel.
Before
How the work ran before the change.
Participant
Uses head tracking to select letters on a commercial assistive device.
ControlUser intent and device selection.
Communication partner
Reads the slowly composed text.
ControlHuman clarification.
What changed
Attempted-speech brain signals can produce real-time text, audio, and facial animation as a faster channel.
Decision rightHuman moves from creator to judge
After
How the same work runs now.
Participant
Attempts to speak while 253 cortical electrodes capture speech-motor signals.
ControlParticipant-initiated use and clinical trial protocol.
AI decoder
Decodes text and synthesizes speech and avatar movements in real time.
ControlLanguage model, confidence, and user correction.
Exception path
Incorrect text is corrected or repeated; the participant can revert to the existing assistive device.
Work removed
Decision authority
The participant owns intended communication and decides use; clinical teams control the investigational device and safety.
Before
Participant
Uses head tracking to select letters on a commercial assistive device.
Control: User intent and device selection.
Communication partner
Reads the slowly composed text.
Control: Human clarification.
After
Participant
Attempts to speak while 253 cortical electrodes capture speech-motor signals.
Control: Participant-initiated use and clinical trial protocol.
AI decoder
Decodes text and synthesizes speech and avatar movements in real time.
Control: Language model, confidence, and user correction.
Work that left the path
Human role before
The participant selected letters slowly with head tracking.
Human role after
The participant communicates through attempted speech and judges or corrects decoded output; clinicians maintain the implant and trial.
AI roleDecision mode: assistive neural decoding. Maps brain activity to communication outputs; it does not infer consent or make care decisions.
14.2 words per minute with the participant's commercial head-tracking device.→Median real-time decoding rate of 78.3 words per minute.
Nature study reported August 2023. · One clinical-trial participant with severe paralysis.
Single-participant investigational system, not an approved general product.
Naturalistic speech speed and accuracy were previously elusive for neuroprostheses.→Median word error rate of 25%.
Study evaluation. · Large-vocabulary real-time decoding for one participant.
A 25% error rate requires correction and limits high-stakes communication.
Anti-pattern
Calling an investigational single-participant result a deployable cure.
Questions
Portability conditions
Reputation risk
medium
Evidence and authority
Current · updated
1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID c42c26c667a37013
Johns Hopkins Health System · Creator to judge
An alert only changes care when a clinician confirms it quickly.
Three Italian endoscopy centers using GI Genius · Creator to judge
That the endoscopist must be the only visual observer capable of flagging a lesion during withdrawal.
Google · Creator to judge
Large migrations can combine machine-found locations, LLM edits, automated checks, and engineer review.
Rising Academies schools · Creator to judge
Students can receive adaptive, low-bandwidth math practice without a scarce one-to-one tutor.