▎ 摘　　要

NOVELTY - The method involves inputting (S110) speech information to an encoder, and extracting (S120) a feature vector. A loss function is calculated (S130) by inputting the feature vector extracted from the encoder to a decoder. The speech information is predicted by the decoder, and another loss function and another feature vector are extracted. The latter feature vector is input to another decoder for performing a graphene unit prediction. A final loss function based on the loss functions is calculated, and the decoders are trained (S140) to reduce the calculated final loss functions. A subsequent prediction is performed in a prediction unit of the former decoder based on a result of previous prediction. USE - Method for speech recognition using graphene information. ADVANTAGE - The accuracy of voice recognition is improved using graphene information. DETAILED DESCRIPTION - INDEPENDENT CLAIMS are included for the following: a computing device; and a structure of a neural network model for speech recognition implemented by a computing device. DESCRIPTION OF DRAWING(S) - The drawing shows a flowchart illustrating a method of learning a neural network model for voice recognition. S110Step for inputting speech information to an encoder S120Step for extracting a feature vector S130Step for calculating a loss function by inputting the feature vector extracted from the encoder to a decoder S140Step for training one of the decoder and the second decoder