Well i was also trying to bring out the fact that although oversampling involves averaging, that's not really averaging. To average out the noise the recommendation is to first oversample (with respective noise and pseudo averaging) and then actually averaging. This would involve first oversampling say 10 bits to get 12 bits using 16 samples, doing that 16 times and summing all the results, then dividing by 16. So in short notation that would be:
First step: (Oversampling+noise+pseudo averaging)
Second step: true averaging
This of course reduces bandwidth still further and requires 16*16=256 samples just to get from 10 to 12 bits.
It applies precise filtering to the signal, as I stated in post #16. The analog filter doesn't need to be a very good one, just needs to attenuate the higher frequencies. The converted signal may still have some undesirable components, which a more precise digital filter can remove. Also, other types of digital filtering/processing can be applied to the signal to achieve other effects, like, oh I dunnno, maybe pre-emphisis.