I wonder if it is correct to slerp the quaternion representation of the matrix rotation part, and lerp translation part of the matrix in animation linear interpolation. Becouse I get disecting result on some certain bones. Could it apear under some circumstances?

Yes, this is conceptually correct though I wouldn't use slerp to interpolate rotation but nlerp. 

Here is a good discussion on this topic:

http://number-none.com/product/Understanding Slerp, Then Not Using It/

29 minutes ago, Dirk Gregorius said:

I wouldn't use slerp to interpolate rotation but nlerp.

I second this, it is also better when you mix multiple animations, not just between two keyframes. (nlerp is order independent)

Does this apply also for widely differring key frames in animation?

For me it applies the way that if I use constant weight 0.0 or 1.0 on played two frames, I get exported animation exactly, but if I use weight in between, I get a messy result,

I cannot detect wheather it is the wide difference in key frames that is the couse of it or my code somewhere, so I seek confirmation