From various articles I've read of people doing these network sniffing tests, over an entire day, an Echo that hasn't had any keyword activations sends approximately 2MB for the heartbeats, checking for updates, etc.
And that is a consistent 2MB/day, no matter how much background conversation occur around it.
Since I'll never have one of these, no matter what Amazon or security experts say, I can't test it myself.
Also note that, since the point of any theoretical recording of the conversations would be to use it to assist voice-recognition training and AI training, such a low quality stream for that general training probably wouldn't be very useful. Sure, some low quality stuff you'd want to test that your training can work with low quality audio as well would be wanted, but this'd be a pretty small subset of the overall training data you'd want to collect.