A method to recognize based on a recurrent model and to train the recurrent model and a device are provided. A recognition device using the recurrent model can divide sequence data into window units, and use a state parameter calculated and stored in a previous window and padding data corresponding to padding added to a current window in order to calculate a recognition result for the current window. A recognition result of target data can be calculated more accurately.;COPYRIGHT KIPO 2018
展开▼