As ever, Daniel Ross’s answer is so thorough and well thought out (Vote #1 Daniel Ross’ answer to Why does reconstructed Proto-Indo-European seem so cumbersome to pronounce?), that it is embarrassing for me to attempt a better answer. In fact, I won’t: I’ll offer a worse answer, but one that is actually hinted at in his “PIE might be, in minor ways, a little bit over-reconstructed”.
And that is that the reconstructions of PIE are not for the purposes of being spoken at all. They are for the purposes of expressing correspondences between cognate languages in a shorthand.
We observe a systematic correspondence between k and w and p, and we put them in the blender, and we call it *gʷ. We could have called it *ʛ. Or *%. Or *Jimmy. We called it *gʷ because that’s an economical articulatory hypothesis for how a single sound can end up as k or w or p. But we don’t know for sure; we weren’t there.
And if you multiply that by a few dozen other hypotheses, and add in the strange algebra of Saussure’s laryngeals, you get a proto-language that internally makes sense, has a consistent root structure and explains the daughter languages—but was never meant to be spoken. It’s a theoretical construct. In practice, we may have missed some smoothing out of the sounds. We may be conflating different stages of the proto-language. We may be reconstructing an abstract phonology of the language, and be completely in the dark about its far more pronouncable allophony.
And maybe Proto-Indo-European did actually sound just like that. But remember: its sound is not what it was reconstructed for. It’s an explanatory tool for linguistic diversity, not a time machine.