摘要 |
PROBLEM TO BE SOLVED: To make an apparatus adequate to an unguide-type checking process, and perform a check with a small amount of check data by processing a user inputted voice using a voice model, forming best N sets of recognition interpretations, selecting a series of highly reliable recognition candidates from the best N sets of recognition interpretations, and checking the voice model using the highly reliable recognition candidates. SOLUTION: This recognition apparatus 10 operates in cooperation with a series of voice models 12. These voice models 12 are given as initial form models, namely speaker indefinite models. The recognition device 10 selects the best N sets of interpretations 14, namely a model most corresponding to inputted voice. A module 16, for example in load process method, gives weight depending on the difference of probability for first and n-th interpretations 14 for each of the best N sets of interpretations 14, and removes inaccurate interpretation 14. A model checking module 18 selectively checks the voice models 12 in a conversion-type checking technology, using this highly reliable information. |