Actually it would be very possible to do if you transcribe the eavesdropped audio into text, as the built-in voice assistants already do on the fly. Small packets containing text could be indistinguishable from a variety of other excessive telemetry that the various apps and system services send out to their motherships in the name of providing a better user experience.

I can't help but wonder why you people are always so quick to dismiss the reality of audio data collection when there are so many ways it could be done and so much evidence that it is happening. Are you stuck in a mode of cognitive dissonance that won't let you distrust your devices, because you WANT to trust them with your privacy?