PURPOSE: To improve recognition performance and to enable to express the features of the inclination of a time-series pattern both locally and generally as to the preparation of parameters of HMM (hidden Markov model) adaptive to the recognition of the time-series pattern.
CONSTITUTION: This devices are equipped with a vector conversion part 602 which converts the static feature vector (x) of the time-series pattern into a dynamic vector (y), a linear restriction probability density calculation part 603 which inputs the static feature vector (x) for the probability density function of restriction conditions varying linearly in the parameters in the same continuous state for a frame included in the vector when the same state continues, and a nonrestriction probability density calculation part 601 which inputs restriction condition; and the product of the probability density values from the linear restriction probability density calculation part and nonrestriction probability density calculation part is outputted as the generation probability density of an input signal to represent the inclination both locally and generally.
TSUBOKA HIDEKAZU