I call BS. I can't even get a device - of any power - to recognise my voice beyond the very slow, pronounced basics and I have to train myself to it (not the other way around).
Sorry to break it to you, but you're wrong.
For one, want you can't do, and what today computers and networks certainly can - after being configured and programmed accordingly - is sample bazillions of phonecalls from millions and millions of people at insane speeds and aggregate speech patterns and their written equivalent by searching for the fitting existing transscripts and do a weighted correlation of those. All with the support of speech and language optimized signal processing, sampling of regional habits and the target groups favorite set of vocabulary.
Guess why Apples Siri and Google Now / Voice Search need an uplink to work
Turning speech into easyly searchable transscripts probably is a piece of cake by now for those who have the storage, processing power, access to unlimited phone-taps and north of 20 000 Mathematicians to programm it all.
Like a certain U.S. three letter agency that has been getting so much unwanted attention lately.