hi,
I hope this is helpfull.
https://hydrogenaud.io/index.php/topic,92825.0.html
A copy & paste of the same.
Audio is normally upsampled before being fed into the DAC so as to make the conversion easier. This requires a low pass filter, and that filter has some kind of roll-off characteristic between the pass band and the stop band.
Some idealise the filter as having a perfect brick-wall cutoff -- an instantaneous transition from pass to stop levels -- and the best approximation of that is the steepest roll-off you can manage. This maximises the possible width of the pass band by allowing you to start rolling off much closer to the start of the stop band (close to the Nyquist frequency). However, this kind of transfer function can ring near transients; both before and after.
If you don't want ringing, then you can use a shallower roll-off. The transition band becomes wider and that necessarily eats into the pass band, so you get a smaller pass band and more high-frequency loss.
These trade-offs are actually the other way around from the more general fact that a wider pass band improves transient response by allowing ringing to settle more quickly. Here I'm saying that you can improve transient response with, effectively, a narrower pass band. Just not too much narrower. It's all about compromise.
You can also mess around with phase, but I'm not going to touch the psychoacoustic debate around that. Presumably both your options have linear phase.
Slower roll-off is also achievable with less computation, and it introduces less delay. I doubt either of these matter here.
For 44kHz content the ringing (when present) will be around 22kHz, and probably not directly audible; but audio hardware will still try to reproduce it and then you might worry about nonlinearities and/or your dog. In any case, it depends on the content, as you'll need content capable of triggering the ringing. I've been told that such content was endemic in the early days of CD.
Conversely, even with a wide transition band the region of activity should still all be ultrasonic, and so the information you save with a steep roll-off is not necessarily more important than the noise you eliminate with a shallow roll-off.
If the drivers allow you to play back 8kHz audio without resampling it then you can bring the artefacts down into the audible band and try both settings. When you are dealing with a perceptible Nyquist frequency the problem is highly subjective and debatable, much like image sharpening.
See also:
http://sox.sourceforge.net/sox.html#EFFECTS (see: rate effect, after "NB")
http://en.wikipedia.org/wiki/Gibbs_phenomenon