Samsung has clarified its terms and conditions to say that it has a contract with a voice recognition company called
Nuance Communications, Inc to handle its voice recognition. This si the "third party" that gets access to the spoken commands for the TV.
The problem with doing the VR on the device is that even if you have only a limited set of spoken commands, you still have to process all the speech you hear in order to figure out whether it contains any of those commands. That, as it turns out, is surprisingly tricky to do. The processing power scales not with the number of commands but rather with how flexible you want to be in recognizing them--it's one thing to recognize a dozen words from a single person in a quiet room with no background noise after training, and quite another to recognize a dozen words spoken by anyone in any pitch or accent in a place with variable background noise. A lot of the processing happens before you even get to the point of recognizing the words--removing noise, normalizing frequency, that sort of thing.
