if the angles are really really small, you might get away with it (actually pretty sure you will). nlerp is already approximating, so you want to minimize the error. if the angles for your source rotations are almost the same, you also won't notice any additional error. if they are too different, the blend will differ significantly from the expected animation. from personal experience i would say if the renormalization constants differ around 2-3% it already gets noticeable (like a slight awkward twist in a joint). just try it out, if your keyframes are tight enough, it should work out.

however i would suggest you go for normalizing in between, since the nlerp approximation only changes speeds, not angles or axis, unlike interpolating unnormalized quaternions. and a few square roots less are not worth getting crappy animations.

by the way, the problem multiplies the more animations you blend.

also, as a general rule of programming (imho), is get stuff working perfectly and as expected first, then optimize and judge the quality/speed tradeoff. especially for something so easily implemented.