I'm assuming the original pronunciation of words in Proto-Indo-European is unknown. How do linguists talk about the reconstructed roots, do they just assume sounds close enough to english or do they spell or ..? What about the oblique roots?
There's a lot of variation. In my experience, many Indo-Europeanists will (try to) pronounce the voiced aspirates as breathy voiced consonants (possibly thanks to having studied Sanskrit, as jlawler points out). A bigger problem is the laryngeals, since the phonetic values of those are far from clear. Some people just pronounce them all as [h] and then specify "...with h₁" or whatever; others try to give each of them whatever phonetic value they think is likely. I've never heard anyone try to pronounce a distinction between the plain and palatalized velars.
Those are the main trouble spots I can think of; vowels are easy because PIE had a small and simple vowel inventory (/a e i o u/ at most).
ETA: as other users point out, some linguists see reconstructed forms as basically an algebraic shorthand for correspondence sets, and are reluctant to treat them as representing sounds with specific phonetic content. To some extent this issue is orthogonal to your question if what you're asking is "How do linguists pronounce PIE words when discussing them with other linguists?" rather than "How was PIE pronounced?", because whatever your views on the status of reconstructions, you still need a way of talking about the forms. (If what you're asking is "How was PIE pronounced?", that's a whole other question.) The Indo-Europeanists I've known mostly happen not to be of the purist school, so I don't know how purists approach this practical question.
Some historical linguists feel about pronouncing reconstructions about the same way as structural linguist grammarians feel about pronouncing deep structures -- it's wrong in principle to attempt it. If you base an analysis on correspondences and patterns, you wind up with the same sort of thing. Not sounds, but rather, patterns. You can't "reconstruct forward". his [Hoenigswald's] conviction that it is not proper to present historical materials "downward, as history" but rather "upward in time, as inference". (from Henry M. Hoenigswald).
(I don't share this view, so I can only hope that I describe it fairly.)
I share Antoine Meillet's skeptical approach to reconstruction. Meillet considered reconstruction as merely a convenient set of formulae to show the pattern of correspondences among forms in different ancient languages. Thus, instead of listing a series of cognate phonemes — for example, Sanskrit bh, Greek ph, Gothic b, Avestan b etc) — we simply write *bh which sums them all. Therefore, *bh is just an abstract formula, not a real sound. In case you have to pronounce it, you just spell it.
Notice that we reconstruct IE forms on the ground of dead languages, of which we hardly know the real pronunciation. The only thing we have at our disposal are some ancient sources written in more or less standardised alphabetic scripts (and sometimes in syllabaries). We are lucky enough to have ancient phonetic treaties describing the pronunciation of Sanskrit, but nothing similar is available for the other sister languages. In many cases we have no certainty about the precise reading of some scripts, even in largely studied languages. For example, the "epichoric" alphabets used by Ancient Greek dialects are far from being understood with all phonetic details. Similarly debated is the pronunciation of the Avestan, Gothic, Cyrillic/Glagolitic alphabets, and many other scripts. The customary reading of such scripts are mere hypotheses, which, moreover, vary from country to country (even Latin and Classical Greek are pronounced differently in Italy, France, Germany and English-speaking countries). Therefore, IE forms are abstractions based on abstractions. Something too far from true linguistic reality.
For example, "sonant coefficients" — later called "laryngeals", which is however quite an arbitrary term introduced by H. Müller by analogy with some Semitic data — were proposed by Saussure as a totally abstract, almost algebraic, means for explaining an apparently deviant phonological correspondences among some ancient IE language.