SPEECH RECOGNITION DEVICE, GENERATING METHOD FOR FEATURE VECTOR USED FOR SPEECH RECOGNITION, AND SPEECH RECOGNIZING METHOD

摘要

PROBLEM TO BE SOLVED: To improve the separating performance between categories with high similarity. ;SOLUTION: A coordinate system determination part 2 puts categories with high similarity in the space of a 1st feature vector obtained by analyzing a speech together into a class and one coordinate system is provided for each class by an arbitrary method, preferably, a main component analysis or decision analysis to set plural coordinate systems in the space of the 1st feature vector. A feature vector generation part 4 projects the 1st feature vector converted from an input speech by a speech analysis part 1 on the said set coordinate systems to generate 2nd feature vectors and puts the 2nd feature vectors together into one 3rd feature vector. A model learning part 5 and a recognition part 7 inputs the said 2nd feature vectors or 3rd feature vector, and learn and recognize a reference model respectively.;COPYRIGHT: (C)1997,JPO