That's fair, but those two are in constant tension. We can do very good speech if we ship gigs of samples. Better quality smaller models make it easier to get better quality speech in more places.