I would hope the researchers were competent enough to space out their explicit uploads with varying periods of noise, while also finding a decent approximation

I would hope the researchers were competent enough to space out their explicit uploads with varying periods of noise, while also finding a decent approximation of how much data is uploaded per x seconds of audio. Sure, Amazon could use all sorts of tricks to attempt to throw off reverse engineering, but it would be pretty hard to do so with a large enough set of tests.