| 摘要 |
PROBLEM TO BE SOLVED: To provide a pattern recognition method and device, and a pattern recognition program and its recording medium, capable of efficiently pruning a state hypothesis without reducing a recognition rate. SOLUTION: In a second search section 15, a likelihood calculation section 151 calculates acoustic likelihood by verifying a tree structure dictionary stored in a second database 20, and an acoustic model stored in a third database 21, to a time sequence data of an acoustic feature parameter, to obtain an accumulated likelihood by accumulating the likelihood in a time direction. A self-transition section 152 performs self-transition of each state hypothesis in search process. A left-right (LR) transition section 153 performs LR transition of each state hypothesis in the search process. A reward adding section 154 increases the accumulated likelihood by adding the reward according to the number of words which is possible to reach, to each state hypothesis in the search process. A pruning section 155 removes the state hypothesis in which likelihood is low, from search objects in the search process. COPYRIGHT: (C)2011,JPO&INPIT
|
