--- title: "A paralyzed woman's attempt to speak became text at 62 words per minute" date: 2026-10-28 category: Brain–Computer Interface site: NeuroAI canonical: https://neuroai.site/a/na-brain-braingate-speech-bci language: en --- # A paralyzed woman's attempt to speak became text at 62 words per minute > Two Nature papers in 2023 showed brain-computer interfaces decoding attempted speech into text and voice far faster and more accurately than before — a quiet leap for assistive communication. A woman who has not spoken clearly in years sits still while sensors watch the part of her brain that once formed words. She does not make a sound. On the screen, a sentence appears — produced from the attempt to speak, not from her vocal cords. ## Key takeaways - A Stanford-led team decoded attempted speech to text at 62 words per minute — about 3.4 times the prior record. - In a 50-word test, the system erred only 9.1% of the time; with a 125,000-word vocabulary, 23.8%. - A separate UC San Francisco group synthesized a patient's voice and animated a speaking avatar. - Both results were published in Nature in August 2023 and built on decades of BrainGate research. - Each headline result came from a single participant; the path from one patient to a usable device is long. ## The study behind the numbers The work was led by Francis Willett and Jaimie Henderson at Stanford and published in Nature on 23 August 2023. Their participant — identified in the paper as T12 — has bulbar-onset ALS, a form of the disease that robbed him of intelligible speech while leaving the neural plan for speech largely intact. Four tiny microelectrode arrays were placed in his brain: two in the ventral premotor cortex and two in a region of Broca's area linked to speech production. When T12 tried to speak, the arrays recorded the firing of individual neurons. A neural network trained on those patterns learned to map attempted articulation to words. The result: attempted speech decoded to text at 62 words per minute, approaching the roughly 160 words per minute of natural conversation and far above earlier brain-to-text systems. ## Why accuracy matters as much as speed Speed alone would be a parlor trick if the words were wrong. On a constrained 50-word vocabulary, the system's word-error rate was 9.1% — about 2.7 times fewer errors than the previous best speech BCI. Stretched to a 125,000-word vocabulary, the error rate rose to 23.8%, but the authors noted this was, to their knowledge, the first successful demonstration of large-vocabulary decoding. For a locked-in patient, a 9% error rate on everyday phrases is the difference between being understood and being exhausting to talk to. ## The sibling study: a voice, not just text In the same issue of Nature, Edward Chang's group at UC San Francisco reported a different approach. Instead of implanting electrodes inside the cortex, they placed a sheet of electrodes over the brain surface of a patient who was severely paralyzed by a brainstem stroke. That patient attempted to speak silently while the system translated the neural signal into three outputs at once: text, a synthesized voice, and a talking avatar whose face moved in sync. The text output reached a median of 78 words per minute at 25% error; the synthesized voice, built on a 372-word vocabulary, carried a 28.2% error rate. The avatar added something text cannot: expression. For someone who cannot move a face, seeing a responsive digital face reproduce intent is its own form of communication. ## What this changes Together, the two papers moved speech BCI from "typing by thinking" toward "speaking by thinking." They showed that high-resolution recordings, whether from inside the brain or just above it, can capture the articulatory plan — the shapes and motions of speech — not just a generic "word signal." That distinction is why the decoded output sounds like language rather than random guessing, and why clinicians took notice. For families, the practical promise is quiet rather than cinematic. A person who cannot move a single muscle could, in principle, order food, answer a call, or tell a grandchild a story — through words that appear as they are formed in the mind. The 2023 papers did not deliver that life yet, but they shortened the distance measurably, and they did it in the open, with methods other labs can replicate. ## Honest limitations Both headline results came from a single participant each, and both required invasive surgery — either penetrating microelectrode arrays or an implanted surface grid. The systems were calibrated to one person's neural signature and have not been shown to generalize across users without retraining. Real-time, everyday use outside a research setting has not been demonstrated at scale. Accuracy, while strong, is still well below the near-perfect intelligibility of natural speech, and error rates climb sharply with vocabulary size. These are research milestones, not products, and the BrainGate consortium's long timeline reflects that reality. ## What readers can do now - Families of ALS or stroke patients should ask neurology teams about speech-BCI research programs at academic medical centers rather than assuming the technology is unavailable. - When reading "brain types at X words per minute" claims, check whether the result came from a peer-reviewed study, how many participants were tested, and on what vocabulary size. - Support or follow organizations funding assistive neurotechnology; translational grants, not just consumer hype, are what move these systems from one patient to many. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-brain-braingate-speech-bci Free to quote with attribution and a link to the original.