While other researchers monkey around, a team at the University of California San Francisco (UCSF) is developing a brain implant that decodes full sentences from neural activity. The device, called a “speech neuroprosthesis,” hit its first milestone after it successfully interpreted a paralyzed man’s intended words and sentences.
The UCSF team tested its speech neuroprosthesis device on a man who, for privacy, asks to be referred to as BRAVO1. Now in his late 30s, BRAVO1 suffered a brainstem stroke in his teens that left him paralyzed and unable to speak (though he uses a baseball cap equipped with a laser pointer to spell words and communicate with others).
In all, BRAVO1 spent just 22 hours working with the UCSF team (over a span of several months, of course). They started by surgically implanting a high-density electrode over BRAVO1’s speech motor cortex, the part of the brain that’s most responsible for producing speech.
Once BRAVO1 had recovered, researchers regularly brought him in to go over a vocabulary list of 50 common words. As BRAVO1 tried to “speak” these words, his neural implant fed brain activity to an AI, which eventually learned how to interpret BRAVO1’s brain activity as language.
The highlight of this study came during a question-answer test. When the UCSF team asked BRAVO1 “How are you today?” he used his brain implant and a screen to answer, “I am very good.” This is the first time that scientists have decoded brain activity into full, organic sentences.
Unfortunately, there are still some kinks to work out. While the speech AI is very easy to train, it can only interpret language with 75% activity when users “speak” at 15 words a minute (regular conversation is about 100 words a minute). But even at its most rudimentary stages, speech neuroprosthesis looks like an incredibly useful tool for those who cannot speak due to paralysis or other disabilities.