Yeah this is a really great idea, perhaps you can sample tokens based on some kind of SRS and I+1 sentence system. Kind of like a graded reader but for speaking. Will definitely look into this in the future