OPTIMALLY CLASSIFIED NEURAL NETWORK CONSTRUCTING METHOD AND AUTOMATIC LABELING METHOD AND SYSTEM USING THE SAME

摘要

PURPOSE: An optimally classified neural network and automatic labeling method and system using the method are provided to carry out automatic labeling rapidly and accurately. CONSTITUTION: L phoneme combinations composed of the names of left and right phonemes are acquired using a phoneme boundary obtained by manual labeling(31). K neural network sets having a multi-layer perceptron structure are generated from training data including input parameters. The neural network sets or updated neural network sets are searched for a neural network having the minimum error for the L phoneme combinations(32), and the L phoneme combinations are classified into K phoneme combination groups searched in the same neural network(33). The K neural networks are trained using corresponding training data to update a weight until an individual error of each neural network converges(34). The K neural networks obtained when the sum of errors of the K neural networks converges construct optimally classified neural network sets(36).