Francis Willett and colleagues published “A high-performance speech neuroprosthesis” in Nature in August 2023 (Vol. 620, pages 1031-1036), reporting work from the BrainGate2 pilot clinical trial at Stanford. The study restored a form of communication to a participant, known as T12, who had lost intelligible speech to amyotrophic lateral sclerosis.
The system recorded the spiking activity of neurons through intracortical microelectrode arrays implanted in speech-related areas of the cortex. As the participant attempted to speak, a recurrent neural network decoded the neural activity into a sequence of phonemes, which a language model then assembled into words. The decoder reached a 9.1 percent word error rate on a 50-word vocabulary and 23.8 percent on a 125,000-word vocabulary, at speeds far beyond earlier brain-to-text systems.
The work is a clear example of modern machine learning meeting clinical neuroscience. The same architectures, recurrent networks and language models, that power consumer AI were the engine that turned faint neural signals into sentences, while the neuroscience supplied the electrodes and the map of where to listen.
For a general reader the result matters because it shows brain-computer interfaces moving from laboratory demonstrations toward practical communication aids for people who have lost the ability to speak, and it illustrates how dependent that progress is on advances in AI.