In August 2023, a team at the University of California, San Francisco led by neurosurgeon Edward Chang published in Nature a high-performance neuroprosthesis that turned a paralyzed woman’s attempted speech directly into communication. The participant, severely paralyzed by a brainstem stroke, had electrodes placed over the speech areas of her cortex; as she tried to speak silently, deep-learning models decoded her neural activity in real time.
What distinguished the work was that it produced three outputs at once from the same brain signals: text on a screen, synthesized speech audio personalized to sound like her pre-injury voice, and the facial movements of an animated avatar. Text was decoded at a median rate of 78 words per minute over a large vocabulary, approaching the cadence of natural conversation and far faster than earlier letter-by-letter spelling interfaces.
The result built on a line of intracortical and surface-electrode work, including the BrainGate consortium’s speech and handwriting decoders, that together moved brain-computer interfaces from moving cursors toward restoring language itself. A companion study from Stanford published the same week reported similar speech-decoding speeds, marking 2023 as a turning point for the field.
These systems are, at heart, applied machine learning. The neural recordings are noisy and high-dimensional; turning them into words, voice, and a moving face is a sequence-modeling problem handled by the same families of neural networks used elsewhere in AI, trained on data collected as the participant attempted to speak.