If there are no Haar cascades in the OpenCV distribution that currently implement how to segment out lips, consider building your own. Take a look at the guide to building Haar cascades from the OpenCV project for more details:
http://docs.opencv.org/trunk/doc/user_guide/ug_traincascade.html
I am assuming you already have positive and negative examples of what you want to classify, and so you can build your own cascades using your own ground truth data. The above guide will get you started on creating your own Haar cascades.
NB: I am usually against deferring people to external links without some sort of closure in my posts, but the process to do this is quite involved, and I can't invest the effort in repeating that information here.