For anyone interested: download cc-shared-culture-20120130.epub from http://code.google.com/p/epub-samples/downloads/list
The epub:trigger is a good way to define the play/pause/mute/unmute controls.
However, still no idea about "confining" an audio element to play only when the XHTML page is actually rendered, and to stop it when the XHTML page is changed. (For example, from my test, iBooks seems to stop the audio when you go back, but not when you go forward)
Also, SMIL does not seem suitable for what I want.