Some of the most convincing 'speech' simulations are in sports games, and they are irritating. And if reading is also involved, then the voice might as well be a garbled mess - I can't pay attention to someone speaking while I read. So really I think it's a case of trying to replicate the way you speak, rather than replicating 'all' text.
In my little demo group on the ST, we made a speaking clock, and it was actually a lot of fun - sorting out the phonetics and number logic - there's a definite sense of achievement when you give a machine the power of speech, well you know what I mean I'm sure. Phonetic speech synthesizers have been around for decades, and they haven't improved very much in our realm at least.
Anyhoo, I think that speech should be recorded with foresight, like in speaking clocks, sports games, or even a game like Fable, where you have several possible nicknames, and the audio just gets swapped out for it. A game where it's only necessary to recite a few phrases and change around numbers, well phonetic speech synthesis is too much, all that is needed is some careful planning and it'll sound natural.
As for such a program, that accurately recites text - well it would be pretty nice, but as I said, replicating yourself is much more of a conceivable idea. You can decide what you need it to say, and can engineer it to your own requirements. Give people a virtual you - but a full solution to this is still a long way away IMO - if it's even possible. Probably be quite fun, and impressive to make a talky box based on yourself though.
