--- title: "From silence to 62 words a minute: inside the speech neuroprosthesis" date: 2026-11-01 category: Brain–Computer Interface site: NeuroAI canonical: https://neuroai.site/a/na-brain-intracortical-speech-decoding language: en --- # From silence to 62 words a minute: inside the speech neuroprosthesis > Stanford and BrainGate showed a paralyzed woman could dictate text at 62 words a minute by decoding attempted speech from the motor cortex. A close look at how intracortical speech BCIs work, what they can and cannot do, and where they are headed. She can still form every word in her mind. What the disease took was the body to say them. For someone with advanced amyotrophic lateral sclerosis (ALS 渐冻症), that silence is the illness itself — not a failure of thought. ## Key takeaways - A 2023 Nature study decoded attempted speech from the brain at 62 words per minute, roughly 3.4 times the previous record for brain-to-text communication. - The system used four tiny intracortical electrode arrays (电极阵列) placed on the brain's speech-planning areas, not on the skin or in a blood vessel. - It is not "mind reading." The person must still try to speak; the implant translates the neural signature of that attempt. - Today's versions are wired and trained per person; wireless, general-purpose speech BCIs remain years away. ## What the study actually did In August 2023, a team led by Francis Willett at Stanford published a paper in Nature describing a speech neuroprosthesis (语音神经假体) built for one participant in the BrainGate2 clinical trial. The participant — a woman with bulbar-onset ALS who could no longer produce intelligible speech — had four small silicon microelectrode arrays implanted in the cortical regions that normally plan speech, areas known as ventral premotor cortex and part of Broca's area. Each array held 64 electrodes in an 8×8 grid, for 256 recording sites in total. The arrays were supplied by Blackrock Neurotech, the same Utah Array (犹他阵列) used across most human BCI research. Thin wires carried the signal to a decoder that watched the patterns of neural firing as she attempted to speak, then predicted which sounds — and ultimately which words — she was trying to make. The participant still had to attempt the words aloud in her mind; nothing was read passively. ## How it reaches 62 words a minute Earlier speech BCIs stumbled on both speed and vocabulary. This one improved both. On a small 50-word vocabulary, the decoded text landed with a 9.1% word error rate, about 2.7 times fewer errors than the prior best. On a large 125,000-word vocabulary, the error rose to 23.8%, but that was still the first demonstration anyone had managed of decoding free-form, large-vocabulary attempted speech. Crucially, the participant's attempted speech was rendered as text at 62 words per minute. Natural conversation runs near 160 words per minute, so this is not yet a casual chat — but it is 3.4 times faster than the earlier record and close enough to be useful for real communication. Two properties of the neural code made this plausible. First, the tuning for speech movements was spatially intermixed, meaning a small patch of cortex carried enough information to decode many articulators. Second, the detailed representation of phonemes, the building blocks of speech, persisted in the brain even years after the muscles stopped obeying. The implant, in effect, eavesdropped on a motor plan the body could no longer execute. ## A companion approach: reading from the surface The same week, a separate Nature paper from Edward Chang's group at the University of California, San Francisco, took a different route. Instead of penetrating electrodes, they placed a thin 253-electrode sheet, known as electrocorticography or ECoG (皮层脑电), on the brain's surface of a person whose brainstem stroke had robbed her of speech. Their system decoded at nearly 78 words per minute with about a 25.5% error rate, and went further — it drove an animated avatar whose face moved in sync with the attempted speech. The two studies are complementary. Penetrating arrays capture single-neuron detail and excel at fine decoding; surface sheets are less invasive and easier to place broadly. Both point the same direction: speech is finally becoming a tractable BCI target. ## Why this matters beyond one patient Restoring communication is the clearest, most urgent use case for invasive BCI. Unlike cursor control, which helps with computers, speech restoration returns something the disease stripped away — the ability to talk to family, to joke, to argue, to be present in a room. For the millions living with paralysis worldwide, a reliable speech neuroprosthesis would be among the highest-value things neurotechnology could deliver. ## Honest limitations This was a single participant, and the result depended on weeks of per-person training. The wired connection is a practical drawback; a person tethered to a decoder is not yet independent. Larger vocabularies still carry meaningful error, and 62 words per minute, while fast for BCI, sits well below natural speech. The work also rests on specialized hardware and surgical implantation, which carries the risks any brain surgery does. None of these studies is a finished product; they are demonstrations of feasibility, not approved therapies, and the gains have not yet been shown to generalize across many users. ## What readers can do now - If you or a family member lives with ALS or another condition that threatens speech, ask a neurologist about speech-preserving care and assistive communication tools available today — BCI is not required to communicate now. - Follow clinical-trial registries such as ClinicalTrials.gov for BrainGate and similar speech-BCI studies if you want to understand enrollment or volunteer pathways. - Treat breathless "mind-reading chip" headlines with skepticism; the real story is slower, harder, and more humane than the hype. ## A trend note According to Li An, chief scientist at BrainNet (脑机网), the bigger unlock will be moving these systems from wired, single-user research rigs to wireless implants that work without a trained operator present. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-brain-intracortical-speech-decoding Free to quote with attribution and a link to the original.