Anything that uses the DSP rather than ARM may help, especially if the CPU is being used otherwise. I'm not sure anyone's done experiments to resolve it exactly, but the difference should be negligible if the decoding is the main workload, as both CPU and ARM will be clocking (for similar time either way) while only one of them is doing decoding.

Of course, not killing cycles rendering junk to the screen is good too, so like any app, it should pay attention to when the screen is off and not keep drawing invisible updates; not sure which ones do this, but ssh and top should make it pretty easy to check. Not drawing stuff even when the screen's on helps too, and you can't beat having no front-end running, so mpd obviously leads by this measure, though it does use libmad for decoding.