Siri voice recognition on the new iPad is remarkably accurate and usable. The processing happens on Apple's servers, which means it is only available for WiFi or 3G iPad models, and then only when one of those tools is available, but, increasingly, we do have access to those features.
What is perhaps most surprising in a pleasant way is that it seems to be hard to outpace the recognition. In initial testing, I've found that in dictating text (that is, creating it and not reading prepared text), the recognition keeps up with my thoughts, and, in fact, I'm lagging behind.
The next round of testing will be pushing the buffer to see how much can be spoken in one session. So far, it seems as if a normal paragraph is not a problem. In fact, I haven't yet had a problem, but we'll see if we can push it further.