Re: The Cloud?
I wrote an MFC wrapper for IBMs via voice for a startup back in the day, and if it was at all trained on a particular speaker (it could handle many but it wanted to be told who was talking) - it was super good. In many opinions better than the Dragon stuff, particularly in the case of custom vocabularies - this was used to transcribe doctor's patient notes, so it needed to know medical jargon and a multitude of ways to say any number (one hundred, a hundred, one zero zero, and on and on for more complex ones). It got so if a doctor often coughed in the middle of saying some weird drug name, it'd sill get it right - due to regular human transcriptionists error-checking and telling the speech engine what was really said.
Adapting how a thing written for unix then to windows then rose some serious eyebrows and won the odd award. It was definitely a complex thunk operation. I've thought about resurecting the codebase, this time just using in linux as I abandoned windows around .net and the VB'ing of visual studio time, as no one was paying me to fix windows anymore - linux ever since.
It's long been known in the speech recog biz that working for one person (or a few known ones) is a metric ton easier than "all ya'll out there". It is in fact easier to tell who is speaking (biometric fashion) than what they are saying for a limited population.
This is one reason the big boys use the cloud. The other is of course, the obvious snooping and slurping.