HMM GENERATING DEVICE, LIKELIHOOD CALCULATING DEVICE USING SAME, AND TIME-SERIES PATTERN RECOGNIZING DEVICE

摘要

PROBLEM TO BE SOLVED: To provide a speech recognizing method with higher recognizing performance than before by increasing affinity with an input speech by making an average vector as a parameter of MMH continuous even between states while varying the average vector linearly or along a quadratic curve in a state. SOLUTION: An HMM generating device is equipped with parameter calculating means 505 to 518 of using as training data an input signal previously known to be in a category R(w) (w=1 to W) and calculating parameters so that the generation likelihood of the input signal from a model becomes maximum on the restriction condition that when the same state continues, a parameter in the state linearly varies with time for a frame included therein and a parameter in an adjacent state is continuous. COPYRIGHT: (C)2006,JPO&NCIPI