SYSTEM AND METHOD FOR NOISE ADAPTATION FOR SPEECH MODEL, AND SPEECH RECOGNITION NOISE ADAPTATION PROGRAM

摘要

PROBLEM TO BE SOLVED: To perform optimum clustering processing for many noise data and to estimate a speech model series for an input speech more accurately.

SOLUTION: A clustering result is obtained by generating a noise-superposed speech (step S1) by superposing noise on a speech under a noise-to-signal ratio condition, performing processing (step S2) wherein a mean value of speech cepstrum is subtracted as to the generated noise-superposed speech, generating a Gauss distribution model of respective noise-superposed speeches (step S3), and calculating likelihood between noise-superposed speeches and generating a likelihood matrix (step S4). An optimum model is selected (step S7) and linear conversion (step S8) is carried out so that the likelihood becomes maximum. A clustering stage and a model learning stage are applied to the noise-superposed speeches, so clustering processing of many noise data and accurate speech model series estimation can be realized.

COPYRIGHT: (C)2005,JPO&NCIPI