| 摘要 |
<p>PURPOSE: To make it possible to easily utilize a language model without extremely increasing searching activity. CONSTITUTION: At the time of recognizing a coherent sound, the reliability of recognition is increased by effectively using a language model capable of especially considering the probability of word combinations of word pairs. Thereby a language model value corresponding to the probability is added to a boundary between words. In the case of establishing a vocabulary e.g. from phonemes of a tree shape in various kinds of recognition methods, any word practically follows a terminated word at the time of continuously starting hypotheses after terminating the word and a language model value can not be considered until the termination of a succeeding word. In order to attain the consideration, the device is provided with a means for considering an optimum preceding word or word sequence about the langulage model value without copying a search tree in each end of a preceding word sequence.</p> |
