Unfortunately, even YCombinator, just like a vast majority of tech companies, are guilty of supporting this notion. There is no respect or regard for privacy and ownership of your person. if it escapes your flesh and bones, it is deemed fair game to harvest by anyone. You and your actions and your speech and your image and your location are a commodity to be optimized for profit. Say moo, mothertrucker.
EDIT: I'm also one of those guys who's like "no I'm not signing up to X service because they don't need more data about me" but I get the double-edged sword and how more of the right data makes certain decisions easier.
I don't think it's trivial, but I hope to see some regulations coming in the way the data are collected and stored.
Of course command grammars are recognized on the device, otherwise you wouldn't be able to turn your TV on using a command or use basic functions in a speedy manner.
And, of course they send audio data to a server -- because your TV does not have a complete language model for up to date web search or even tv guide installed a priori. i.e. good reasons that have nothing to do with "capturing revenue stream".
Sigh...
This is what I don't get. I used to play with Microsoft's SpeechAPI back in 2007 and it was pretty decent at real-time speech recognition - almost perfect, if you limited your program to a pre-defined command grammar. It was all done completely off-line, real-time on a machine that had much less processing power than your average smartphone of today. I don't see anything inherent in speech recognition that would require so much processing power that the phone can't handle it real-time. The only reason I see is that continuous moving of everything to cloud because it's easier to capture revenue stream that way.