摘要 |
PROBLEM TO BE SOLVED: To make the accuracy of the digit string recognition in Japanese substantially equal to the accuracy in the unclustered model and reduce the required number of voice models by clustering the contexts between words/ phrases into few classes only. SOLUTION: The digitized voice input is processed by the recognizing device program executed on a general-purpose computer 15. This program memory finds the best matching between the input voice and a data base 17 and outputs the recognition result. The data base 17, i.e., data structure, is provided with clustered models and the grammar having the clustered contexts between words/ phrases. The clustered model is formed when the contexts between words/ phrases are clustered. The contexts between phrases are formed into models to attain an articulation connection effect, and the contexts are clustered to considerably reduce the number of models without sacrificing the recognition performance. |