Patent US 5,737,486

A method for training speech sound predictors, which define a reference pattern representative of a reference category, by using training patterns representing speech sound to adjust parameters characterizing each of the predictors. The training patterns may represent a training category identical with the reference category. The training patterns may alternatively represent a training category different from the reference category. The training method includes steps of initializing the parameters; calculating the prediction error between the reference pattern and one of the training patterns at a time; correcting the parameters so as to either decrease the prediction error (according to the first aspect) or increase the prediction error (according to the second aspect), and iterating the second and third steps for all of the training patterns.