Speech Neuroprosthetics

The first Bravo participant had been paralyzed for 15 years after a car accident and later brain stem stroke. The basic speech neuroprosthetic strategy records speech-related cortical activity during attempted speech and uses machine-learning algorithms to decode intended speech. The first decoding system used a 50-word vocabulary to build a computational language model of possible sentences. Speech neuroprosthetic decoding relied on analog-to-digital conversion and machine-learning analysis of subtle neural patterns. Autocorrect-like language modeling was necessary because neural decoding was imperfect. The implanted system placed electrode arrays over cortical areas controlling the vocal tract, including the larynx, lips, tongue, and jaw. Training the speech decoder took weeks and involved prompting the participant to attempt specific words. Speech neuroprosthetics aim to restore communication by decoding neural signals involved in attempted speech into expressive outputs. The Bravo clinical trial applies attempted-speech decoding to people with severe paralysis.