Here it is described the generation process of the phoneme models. This process uses the information about training samples. It is necessary to calculate the probability parameters of each model and such algorithm of this calculation is described. To choose the better model, the evaluation criterion of model effectiveness is proposed.