Recurrent neural network training method, computer program therefor and speech recognition device

Abstract

[Object] An object is to provide a training method of improving training of a recurrent neural network (RNN) using time-sequential data. [Solution] The training method includes a stepof initializing the RNN, and a training stepof training the RNN by designating a certain vector as a start position and optimizing various parameters to minimize error function. The training stepincludes: an updating stepof updating RNN parameters through Truncated BPTT using consecutive N (N≥3) vectors having a designated vector as a start point and using a reference value of a tail vector as a correct label; and a first repetition stepof repeating the process of executing the training step by newly designating a vector at a position satisfying a prescribed relation with the tail of N vectors used at the updating step until an end condition is satisfied. The vector at a position satisfying the prescribed relation is positioned at least two vectors behind the designated vector.