Is Large Margin HMM what you are looking for ?

http://machinelearning.wustl.edu/mlpapers/paper_files/NIPS2006_143.pdf

This paper might be useful http://ibug.doc.ic.ac.uk/media/uploads/documents/ICCV-HCI07_ValstarPantic-FINAL.pdf

SVM being highly discriminant can be used to train on classes, later to be transformed to a probabilistic form. This might be good if you're looking at time-variant faces - perhaps an age-resistant face recognition system.

Hi sahahla azizi;

can you explain more; what do you mea, by ; "I think you can use log likelihoods to train and test SVM."

Hi sahahla azizi;

think you for response.

it is very helpful.

Think you all for your responses.