On the other hand, it's really not much slower with the state of the art, especially when you mix in stuff like eye tracking at a really core level (we're playing stuff like autocompleting based on a symbol you just looked at!). I'm not far enough along with Talon to be pushing the benchmarking side of things heavily yet, but one early test was about 2/3 the code input performance of a 90wpm typist on the same code, with a lot of obvious places to improve. I think good application of continuous recognition, resumable grammars, and really context-specific helper code can push specific workloads way past what a keyboard/mouse can do (which calls back to the goal of impressing people who aren't injured yet).
Then there's the professional app space - e.g. Photoshop, CADD, and video/audio production tools could really benefit from voice workflows (imagine using a pen tablet augmented with voice + eye tracking instead of complex UI).
But I wonder if we'd be better off enabling voice input sooner, before we're in serious trouble, and using it sporadically. Mixing voice and keyboard input at need looks like a way to substantially reduce typing while still using a keyboard whenever voice work is too inconvenient.