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

A recurrent neural network (RNN) training method performed by a computer using training data prepared in advance;

• wherein
• the training data includes a sequence of vectors each consisting of a feature of data to be trained and a reference value;
• the training method comprising the steps of:
• initializing the RNN;
• a first designating step of designating, as a training start position, a certain vector in the sequence of vectors;
• and
• a training step of training the RNN by optimizing various parameters of the RNN such that a prescribed error function calculated for the training data is minimized;
• wherein
• the training step includes
• an updating step of updating the parameters of the RNN by Truncated BPTT, using, of the sequence of vectors, consecutive N vectors, where N is an integer not smaller than 3, having the designated vector as a head, and having a reference value of a vector at the tail of the consecutive N vectors as a correct label, and
• a first repetition step of repeating, until a predetermined end condition is satisfied, a process of executing the training step by newly designating a vector at a position satisfying a prescribed relation with the vector at the tail of the N vectors used in the updating step;
• wherein
• the vector at the position satisfying the prescribed relation is a vector positioned at least two or more vectors behind the designated vector designated at the sciccting first designating step.