To perform voice search you have to say "ok, google", so phone needs to listen and to send to server every phrase it hears in order to check whether you said "ok, google" or just ordered a beer in a pub. I'm just trying to theorise.
* Sending all microphone input to a remote server would be really bad for battery life and data usage.
* It still works even in airplane mode.
* On ChromeOS (which isn't Android, but may share the same code) their privacy whitepaper [1] says:
If you opt-in to the feature, Chrome OS will listen for
you to say "Ok Google" and then send the audio of the
next thing you say, plus a few seconds before, to
Google. Detection of the phrase "Ok Google" is performed
locally on your computer, and the audio is only sent to
Google after it detects "Ok Google"."
Those keyword identification algorithms can be implemented at the hardware level. I believe this is done for specialized devices like Echo and Kinect, not sure how phones do it.
It still works even in airplane mode.
I agree that they probably only send speech recorded after the activation words. However, I don't arguments think your arguments hold up. Speech data compresses exceptionally well and generally requires lower sampling rates and/or resolutions. So, it's feasible to compress the speech signal (possibly hardware-assisted), buffer it, and send it to the server at some regular interval or when the user wakes the phone.
Not if you transcribe it first, and upload compressed text.