This was an investigation to determine the feasibility of an autoregressive (AR) model for establishing characteristic parameters from recorded occlusal sounds and develop their classification. Thirty four normal subjects with intact natural dentitions were selected for the study. The subjects' occlusal sounds from both sides of their faces respectively were sampled, and the gnathosonic classification (Class A, B and C) was established by observing the original recorded wave pattern and measuring the duration. Then, a 20 order AR model was calculated with the collected data, and the AR model coefficients were found to be similar to the indices of Bayes' discriminatory analysis. The total conformation rates of the modelled left and right occlusal sounds to the classification, estimated by Bayes' discriminant functions were 97.06% and 88.24% respectively. AR coefficients representing the characteristics of human occlusal sounds can be helpful in their classification and allow computer diagnosis of occlusal disorders.