METHOD OF CORRECTING HMM PROBABILITY FOR REDUCING ERROR IN TRAINING DATA RECOGNITION

摘要

PURPOSE: A method of correcting HMM(Hidden Markov Modeling) probability is provided to reduce a training data recognition error by omitting a process of obtaining an optimum status string and a process of calculating hardness and to improve speech recognition performance by introducing a weight corresponding to HMM. CONSTITUTION: The pattern of input speech is recognized through a classification function(S110). A class for the input speech is determined(S120). A mean loss function representing speech recognition performance based on the determined class is obtained(S130). A model parameter is extracted using the mean loss function(S140). HMM weight training is carried out in response to the model parameter. An HMM weight for each class is applied to a status shift component and a probability density component in the classification function.