SPEECH RECOGNITION DEVICE, METHOD FOR LEARNING WEIGHT VECTOR ACCOMPANYING SUBBAND TYPE SPEECH RECOGNITION DECODER, AND SUBBAND TYPE HMM LEARNING METHOD

摘要

PROBLEM TO BE SOLVED: To provide a subband type speech recognition device which has high robustness and high recognition precision.;SOLUTION: The speech recognition device includes an HMM having learned to perform speech recognition by inputting a feature vector of a speech signal, subband feature vector extraction parts 80, 82, 84, and 86 which extract feature vectors of the same dimensions as feature vectors by subbands of the inputted speech signal as the feature vectors, and multiplying circuits 88A to 88K and DCT processing part s 90A to 90K which multiply the feature vectors of the plurality of subbands extracted by the subband feature vector extraction parts by weights previously assigned to the respective subbands and mutually add them together to generate and supply feature vectors to the HMM.;COPYRIGHT: (C)2005,JPO&NCIPI