| 摘要 |
PROBLEM TO BE SOLVED: To obtain a voice recognition device having a language model which gives zero or a very small value as a bigram probability for an abnormal word chain even though the amount of corpus is not sufficient. SOLUTION: The device is provided with a language model hierarchy determining means 12 which has a statistical language model having a hierarchical structure and determines an optimum hierarchy to compute a language bigram by a preceding word, and a language likelihood computing means 13 which gives a very small value to the language likelihood of a strange word chain while considering the frequency of the appearance of an optimum hierarchy class for the bigram probability of the chain word. Then, zero or a small value close to zero is given as the probability for an abnormal word chain to improve the recognition performance/processing speed. |
