I think you hit the nail on the head there, MDB. As mentioned in the Widex paper (the 3rd link that Bob shared), dynamic compression of the input of 16-bit system is an easy/cheap way to prevent clipping/distortion when subjected to loud sound. It’s kinda like somebody holding the input volume knob and if the sound gets too loud, he turns the volume knob down to avoid distortion, and when the sound gets softer, he turns the input volume knob back up. If you hear very expressive music that changes volume very frequently and quickly, the effect of cranking the volume knob down and up and down and up will start making the music sound funny and not natural anymore. That’s what they call the pumping effect. And dynamic compression may very well be set differently at different frequencies as well, making it sound even more weird.
For speech it’s not as bad because speech tends to be more monotonous and doesn’t fluctuate as much in volume. But if you add speech with loud noise then it starts to get worse because the dynamic compression to cut down on the noise level negatively suppress the volume of the speech as well.