Abstract: A novel a PNN model is proposed for class conditional density estimation based on the mixtures of PNN of shared pattern layers and PNN of separated pattern layers. Each class not only has a set of pattern layers belonging to itself, but also has several pattern layers shared for all class, where "shared" means that each kernel may contribute to the estimation of the conditional density of all classes. The training of the novel model utilizes the maximum likelihood criterion and an effective EM algorithms to adjust model parameters .s developed. These results of the closed-set text-independent speaker identification experiments indicate the proposed model and algorithms improve identification accuracy.