| 摘要 |
PROBLEM TO BE SOLVED: To solve the problem that Baum-Welch algorithm and Viterbi algorithm for speech recognition using variation Bayesian learning are large in calculation quantity since acoustic parameters are obtained through repetitive processing. SOLUTION: In calculation of acoustic model learning of each frame and score calculation in speech recognition by Baum-Welch algorithm and Viterbi algorithm using variation Bayesian learning, common terms which are common to calculation of each frame are previously calculated and recorded, and used for calculation of each frame. Further, a general score calculation function such that formats of the common terms and the function type of the acoustic score calculation are identical is used for acoustic score calculation. COPYRIGHT: (C)2006,JPO&NCIPI
|
