| 摘要 |
PURPOSE: To make it possible to omit the calculation of similarity of a specific part by preparing a 1st word dictionary constituted of words such as nouns and verbs not to be changed and a 2nd word dictionary constituted of changing parts such as particles and conjugation and narrowing down a recognition candidate word. CONSTITUTION: A speech feature parameter is temporarily stored in a feature parameter storing buffer 11. A 1st word dictionary 13 is constituted of stems excluding conjugation. Similarity between the feature parameter and a reference pattern stored in the dictionary 13 is calculated by a 1st word similarity calculation part 12. When the calculated value satisfies a specific value, a 1st word lattice preparing part 14 judges the detection of a recognition candidate word and sends three kinds of prescribed information to a 2nd word similarity calculation part 15. Then a connection dictionary 16 is referred to, and when a connectable word exists, the parameter temporarily stored in the buffer 11 is inputted to the calculation part 15 and a word of a conjugation part is prepared by the similar means. The stem is connected to a conjugation part and a word lattice is outputted. Since it is unnecessary to detect recognition candidates and calculate the similarity of all items, operation quantity can be reduced. |
