Abstract: Embedded-Hidden Markov model (E-HMM) and artificial neural network (ANN) was combined within the hybrid architecture for face recognition. E-HMM was used to parameterize the face image. Every person's face was represented by a E-HMM, the output of likelihood of the E-HMM was encoded to form the input vector sending to ANN. By taking advantage of the discriminative training of ANN, the weakness in discrimination ability of the Maximum Likelihood training of E-HMM could be overcome, and the recognition performance was enhanced by means of the learning ability of ANN. Experiments with ORL (Olivetti Research Ltd.) face image database show that the discriminative ability and recognition performance of the hybrid architecture is better than normal E-HMM.