Radioshack sold a voice recognition chip in the 80s[1] that was a simple 8-bit microcontroller. If you are willing to slip on the accuracy and false positives you can do recognition with very little computation.
the downside was power usage. Motorola made one that was power efficient, used by nokia in the 90s and its pretty much the same chip in google's phone line today (just even more power efficient).
The Google Now only listens for the trigger phrase when idle - which is done all locally, without needing to talk to the servers.
It has a battery impact but much less than sending all the voice data continuously to a server somewhere. The biggest battery killer would be the wifi or 3G transmitting non-stop in that case.
It wouldn't have to transmit non-stop -- it could do some parsing/cleanup locally, then queue it up and upload it periodically with other, expected FB traffic.
