Nonlinear stochastic models and new parameters of computer speech recognition

There are some problems that disturb researchers and developers working on multidimensional signal processing as computer senses. One of these problems is to find more reasonable characteristic parameters for speeches, letters, maps and senses. As is known, LPC-CEP coefficients as the main parameters drawing from signals are widely used and, unfortunately, in the parameter space of which some signals cannot be distinguished. Moreover LPC-CEP coefficients are obtained based on the linear AR (auto-regression) model, so assumption of certain stability for these signals is necessary and the order of the AR model cannot help to simplify the model from ARMA(p,q). But we must address the nonlinear signal to deal with the above information. Finally, the space possess too high a multidimensional number to calculate in time. To avoid these troubles and to strengthen the ability of the models, we study a type of nonlinear stochastic models, AR(p)-MA(q).