Active appearance model efficiently aligns objects which are previously modelized in images. We use it for Human Machine Interface (face gesture analysis, lips reading) to modelize mouth on embedded systems (mobiles phones, game console). However those models are not only high memory and time consumer but also not robust in the case of object with high deformations (different pose of a face or different expressions of mouth): this is the manifold problem [3]. We propose a new optimization method based on Nelder Mead Simplex [12] initialized by Gaussian Mixture (GM). The GM is applied to the learning data in the reduced space. This method reduces memory requirement and improves the efficiency of AAM when we modelize high deformable object at the same time. The test, carried out on France Telecom and BioID data bases, shows that our proposition to align mouth outperformed the classical optimization when applied to mouth alignment and give the same results as classical optimization on common face alignment.