Method of training artificial neural network and method of evaluating pronunciation using the method
The disclosure relates to a method of training an artificial neural network so that a first artificial neural network is trained based on a plurality of training data including a first feature and a second feature that has a correlation with the first feature and depends on the first feature.
1. A method of training an artificial neural network, the method comprising: generating first output data corresponding to first training input data, by using a first artificial neural network, wherein the first artificial neural network is trained based on a plurality of training data comprising a first feature and a second feature that has a correlation with the first feature and depends on the first feature; and the first artificial neural network is a neural network trained to generate output data corresponding to the first feature from input data; generating third output data corresponding to the first output data and second training output data, by using a second artificial neural network, wherein the second artificial neural network is a neural network trained to output a result of comparison between the first output data and the second training output data, and the second training output data includes data comprising the second feature of the first training input data; generating at least one weight correction value for training the first artificial neural network based on the third output data; and applying the at least one weight correction value to the first artificial neural network, wherein the least one weight correction value is determined by a method in which a scale factor applied to at least one gradient value.
2. The method of claim 1 , further comprising after generating the first output data, generating fourth output data corresponding to the first output data and third training output data, by using a third artificial neural network, wherein the third artificial neural network is a neural network trained to output a result of comparison between the first output data and the third training output data, and the third training output data includes data comprising the first feature of the first training input data.
3. The method of claim 1 , further comprising: generating each of the plurality of training data including: the first training input data, third training output data comprising the first feature of the first training input data, and the second training output data comprising the second feature of the first training input data.
4. The method of claim 3 , further comprising: generating first training data from among the plurality of training data including: uttered voice data of a first language of a first speaker as the first training input data, data comprising a feature of the first language as the third training output data, and data comprising a feature of the first speaker as the second training output data, and generating second training data from among the plurality of training data including: uttered voice data of a second language of a second speaker as the first training input data, data comprising a feature of the second language as the third training output data, and data comprising a feature of the second speaker as the second training output data.
5. The method of claim 4 , wherein: generating the first output data further comprises outputting the first output data from the uttered voice data of the first language of the first speaker, generating the third output data further comprises generating the third output data that is a result of comparison between the first output data and the second training output data comprising the feature of the first speaker, and generating the at least one weight correction value further comprises, by referring to the third output data, generating the at least one weight correction value for reducing the feature of the first speaker, from the first output data.
6. The method of claim 4 , wherein: generating the first output data further comprises generating the first output data from the uttered voice data of the second language of the second speaker, generating the third output data further comprises generating the third output data that is a result of comparison between the first output data and the second training output data comprising the feature of the second speaker, and generating the at least one weight correction value further comprises, by referring to the third output data, generating the at least one weight correction value for reducing the feature of the second speaker, from the first output data.
7. The method of claim 1 , further comprising, after the generating of the first output data, generating the first output data depending on a feature of a first language, from input data comprising uttered voice data of the first language of a third speaker, by using the first artificial neural network.