pretty simple: through direct electromagnetic stimulation of photoreceptors in the retina in patterns corresponding to letter shapes, transduced through layers of processing there, passed through the optic nerve to the lgn of the thalamus and from there to visual cortex, eventually you get a complete neural representation of the text. alternatively, you could use direct stimulation of the chochlea via pressure waves in the ear canal transduced by the bones of the inner ear. the auditory cortex will then do most of the heavy lifting.
seriously, though, that paper's achievement was pretty limited: "With 10 words in the dictionary, 75% of all words are recognized correctly." That is just over 3 bits of information being conveyed. And that is after training the classifier.
Training the brain to classify direct inputs is at about the same level of a few bits. www.nature.com/articles/srep01319
see What research has been done on brain-to-brain interfaces?