It's not so much about looking at the payload than it is about whether it's sending anything at all and if it is, what amount. Correct me if I'm wrong, but from everything I've read and heard these devices don't really have much in the way of internal storage capacity (yet). It either has to be uploaded or discarded.
edit:
Also, voice compresses really well if you don't care about the quality. With a codec that uses <10 kbit/s (e.g. GSM-HR, AMR), the storage requirements are at most 2.4 MB/hour. A simple gate filter that cuts out quiet periods should cut the storage requirements down to only a few MB/week. A typical flash chip used to store the firmware could easily store years worth of typical household speech in the non-firmware space.
But inside the Echo? It's not my area of expertise, but I'm almost certain that there are people with the expertise that are looking for exactly that and if Google or Amazon had those built in, we'd know about it. I think we can at least rule that one out.
As for the compression, that is very interesting indeed. I had no idea you could compress voice that much. If / When they're going for 24/7 recording, this would be most likely the way to go and it would be extremely hard, if not impossible, to detect.