That's a great idea! This would really push up input speed. I'm surprised I haven't seen any implementations in the wild yet, although research has been done to explore this [0]
[0] https://pdfs.semanticscholar.org/1170/44feb247a82ab81a30b57f...
[0] https://pdfs.semanticscholar.org/1170/44feb247a82ab81a30b57f...
Long before that, Optikey has had "multi-key selection" which works in the same way, see https://youtu.be/HLkyORh7vKk?t=10
In my experience (as a developer of eye gaze interfaces, not an everyday end user) it is often more efficient to have really good next-word-prediction (such as with the Presage engine) combined with single-letter dwells, rather than using "swipe-like" spelling, where you're committed to tracing out the whole word.