The biggest contributor to the uncanny valley effect probably lies in the lip syncing itself. Maybe it's just me, but talking NPCs look as though some omnipotent invisible puppet master is just pulling FaceGen sliders a la the character generation process. It's kind of weird to see a Redguard change his brow, eyes, and cheeks whilst speaking, but to notice that his nose has been in the exact same position on the screen the whole time. All we can do at this time (other than scrapping voiced dialog entirely, which I wouldn't mind in the least) is to throw in some additional animations and reduce the zoom a bit. There should probably be some slight threshold to the camera's lock on their head, too - instead of your view jerking with the slightest nod in the NPC's head.

This is even weider:

Honditar to a random guard: YOU'R THE DEVINE CRUSADER!

I didn't really like the NPCs conversing with each other either, although more due to the animations. Seriously - everybody either had their arms folded and their body hunched over like some schizophrenic nutcase or they just would stand idly and wait four seconds to respond.