Why not apply ML techniques and get rid of the listener all together?

It's great work, but it just seems like an intermediary step.

p.s. I have no experience in analyzing EEGs, I am sure they are much more nuanced than I know, but could anyone explain the reasons preventing such a thing?