A high-performance speech neuroprosthesis
Comprehensive Overview
Published in Nature in 2023, this study asked whether a brain implant could restore fast, accurate communication to someone who has lost the ability to speak. The research centered on one participant, a woman with amyotrophic lateral sclerosis (ALS, a progressive disease that damages the nerve cells controlling movement) who retains limited facial movement but can no longer produce intelligible speech. As part of an ongoing clinical trial, she had four tiny microelectrode arrays surgically implanted in two brain regions involved in planning speech movements. Over dozens of recording sessions, she attempted to say sentences aloud or silently mouth them while researchers recorded the electrical activity of individual neurons and used it to train software, a neural network paired with a language-prediction model, to translate her attempted speech into on-screen text in real time. Training drew on nearly 11,000 attempted sentences collected across many days. As expected from earlier, smaller-scale demonstrations of this technology but far surpassing them, the system decoded her attempted speech with substantially fewer errors and at a much faster rate than any previous speech brain-computer interface, marking a major step toward restoring natural conversational ability to people with severe paralysis.
Outcomes and Implications
The most striking finding is the sheer size of the improvement: on a constrained 50-word vocabulary the system made 2.7 times fewer errors than the previous best speech brain-computer interface, and on an unrestricted, general-English vocabulary of 125,000 words, the first such large-vocabulary demonstration ever achieved, it still decoded correctly about three-quarters of the time. Decoding speed reached 62 words per minute, more than triple the prior record and closing in on the roughly 160-words-per-minute pace of natural conversation. Performance held up whether she vocalized quietly or silently mouthed words, and more implanted electrodes and more days of training data both steadily improved results, pointing to a clear path for future gains. A second, more basic discovery emerged alongside the engineering success: within a fingernail-sized patch of cortex, signals controlling the jaw, lips, tongue and voice box were densely intermixed rather than neatly separated by body part, and this detailed map of speech production remained intact years after paralysis took her voice. Notably, Broca's area, long considered central to producing speech, contributed almost no useful decoding information, challenging a textbook assumption. Because existing communication aids for people who cannot speak, such as letter boards or eye-tracking spellers, are far slower than natural conversation, this speed improvement is clinically meaningful. The authors caution that this remains a single-participant proof of concept, not a finished product, and conclude that higher-channel-count recording devices are a feasible path toward restoring conversational speech to people with paralysis.