Maybe they have combined an LLM or something with the speech detection convolution layers or whatever they were doing. Like with JSON schemas constraining the set of available tokens that are used for structured outputs. Except the set of tokens comes from the top 3-5 words that their first analysis/network decided are the most likely. So with that smarter system they can get by with fewer electrodes in a smaller area at the base of the skull where cranial nerves for the face and tongue emerge from the brainstem.