Method of setting optimum-partitioned classified neural network and method and apparatus for automatic labeling using optimum-partitioned classified neural network

An apparatus for automatic labeling using an optimum-partitioned classified neural network, comprising:

• a labeling unit to generate a first label file by performing automatic labeling for a manual label file;
• an optimum-partitioned classified neural network composing unit searching neural networks having minimum errors with respect to a number of L phoneme combinations from a number of K neural network combinations generated at an initial stage or updated, updating weights during learning of the K neural networks by K phoneme combination groups searched with the same neural networks, and composing an optimum-partitioned classified neural network combination using the K neural networks of which a total error sum has converged;
• and
• a phoneme boundary tuning unit tuning a phoneme boundary of the first label file by using a phoneme combination group classification result and the optimum-partitioned classified neural network combination supplied from the optimum-partitioned classified neural network composing unit, and generating a final label file reflecting the tuning result,
• wherein the phoneme boundary tuning field of the phoneme boundary tuning unit is set to a predetermined field of a duration time of left and right phonemes of the phoneme combination.

An apparatus for automatic labeling using an optimum-partitioned classified neural network, comprising:

• a labeling unit to perform automatic labeling of a manual label file and generate a first label file;
• an optimum-partitioned classified neural network composing unit to receive input variables, segment phoneme combinations into partitions applicable to neural networks, and compose optimum-partitioned classified neural networks of Multi-Layer Perceptron-type from re-learned partitions;
• and
• a phoneme boundary tuning unit to tune a phoneme boundary of the first label file supplied from the labeling unit and to generate a final label file reflecting the tuning result,
• wherein the phoneme boundary tuning unit tunes the phoneme boundary by using the optimum-partitioned classified neural networks composed after completing learning in the optimum-partitioned classified neural network composing unit and judges the phoneme boundary according to whether an output of a neural network is 1 or 0 after applying the same input variable as the input variable used during the learning.

A method of composing the optimum neural network combination minimizing a total error sum, comprising:

• preparing an initial neural network combination using input variables;
• searching for the optimum neural network having minimum errors with respect to the phoneme combinations from the initial neural network combination and classifying phoneme combinations with optimum neural networks;
• merging phoneme combinations with the same neural network and classifying the merged phoneme combinations into new partitions;
• updating each neural network and learning each neural network according to the partitions generated by the classifying and the merging;
• determining whether all neural networks are converged, wherein if all neural networks are converged, composing the neural network combination as an optimum-partitioned classified neural network combination, and if all neural networks are not converged, then re-learning the optimum neural network having minimum errors by repeating the classifying, the merging, and the updating operations until all neural networks are converged.

A method of learning and updating a neural network, comprising:

• preparing initial neural network combinations composing neural networks of multi-layer perceptron-type;
• searching a multi-layer perceptron index having a minimum error in the initial neural network combinations of all phoneme combinations;
• classifying merged phoneme combinations into new partitions by merging phoneme combinations with the same multi-layer perceptron index if the multi-layer perceptron index having a minimum error for all phoneme combinations is searched;
• and
• re-training neural networks to update weights by learning data applicable to each partition, wherein the re-training procedure of individual neural networks calculates errors using the updated weights and repeating the re-training until a changing rate of errors becomes smaller than a first threshold value.