摘要 |
PROBLEM TO BE SOLVED: To solve the problem that a sound model generating device deletes the whole word from learning data even when abnormal data is present in only a portion of the word. SOLUTION: Only the abnormal data can be deleted from the learning data by the sound model generating device which acquires two or more pieces of second feature vector information from a received second speech, calculates a model parameter based upon an HMM and two or more pieces of feature vector information to be evaluated including a stored first feature vector, calculates the distance to a model, judges whether the second feature vector information is normal or abnormal based upon the distance, and repeats operations to calculate a model parameter, calculate a distance, and make a judgement based upon second feature vector information and first feature vector information, judged to be normal data through recent judgment processing, as feature vector information to be evaluated to output the last calculated model parameter. COPYRIGHT: (C)2007,JPO&INPIT
|