Motion Texture: A two-level Statistical Model for Character Motion Synthesis 
------------------------
This paper is pretty similar to the dance generator paper. The main differenc is that the beat don't have to be of  
the same size. Some motions may be longer than others. And therefore, the method for identifying the pieces are  
different. In the Dance Generator paper, they used the crossings in the second derivative to find the beats. In this paper  
they use a statistical model to find the textons. Each clip is modelled as a linear dynamic system.

It looks to me that the approach used in this paper is superior to that in the dance generator paper, because they don't assume beats of the same size. So, the solution is more general.