Study co-author Maitreyee Wairagkar, a neuroscientist at the University of California, Davis, and her colleagues trained deep-learning algorithms to capture the signals in his brain every 10 milliseconds. Their system decodes, in real time, the sounds the man attempts to produce rather than his intended words or the constituent phonemes — the subunits of speech that form spoken words.
This is a really cool approach. They're not having to determine speech meaning, but instead picking up signals after the person's brain has already done that part and is just trying to vocalize. I'm guessing they can capture nerve impulses that would be moving muscles in the face, mouth, lips, and possibly larynx and then using the AI to quickly determine which sounds that would produce in those few milliseconds those conditions exist. Then the machine to produces the sounds artificially. Because they're able to do this so fast (in 10 milliseconds) it can get close to human body response and reproduction of the specific sounds.