Subjectively it seems to me that the RNNoise sample doesn't trigger my brain to attempt to fill in the gaps.

With the Speex/raw ones I have all the data so if I listen to it again over and over I can get more out of it eventually.

With the RNNoise one I obviously don't even have enough extra data to even try doing that so all I can do is blame the algorithm.

Perhaps what you really want is an algorithm that lets through a bit more of the 'possible noise' for the human brain to have another go at.