VOICE RECOGNITION SYSTEM AND METHOD THEREFOR

摘要

PROBLEM TO BE SOLVED: To assist prevention againt the relation of an ineffective vocalization with a stored voice model and to improve the accurate detection of effective vocalization by variable stop strictness as a function of background noise level which is measured in training and noise signal measurement which is carried out in recognition and deriving a word entrance penalty as a function of the variable stop strictness. SOLUTION: The noise feature of test vocalization is calculated, first (602). In recognition mode, the background noise measurement is performed. Its value is compared with a reference noise value calculated in training mode. The ratio of a recognized background noise estimated value to a training background noise estimated value is found (604). Then a processor calculates a word penalty (606). This ratio is used to calculate the word entrance penalty. The word entrance penalty controls the strictness of an out-of-vocabulary stop. The recognition follows a main retrieval and parallel garbage model.