LEARNING APPARATUS, FEATURE VALUE CONVERSION APPARATUS, LEARNING METHOD AND PROGRAM
A learning device includes a teacher model unit that includes a first encoder and a first decoder including a first recurrent neural network, a student model unit that includes a second encoder and a second decoder including a second recurrent neural network, and a learning control unit that performs learning of the teacher model unit by using a sequence-to-sequence machine learning method and performs learning of the student model unit by using a knowledge distillation method from the teacher model unit. A feature conversion device includes an acquisition unit that acquires a student model unit, and a converter that converts an input feature sequence into a target feature sequence
1 . A learning device comprising:
a teacher model that includes a first encoder and a first decoder including a first recurrent neural network;
a student model that includes a second encoder and a second decoder including a second recurrent neural network; and
a learning controller that performs learning of the teacher model by using a sequence-to-sequence machine learning method and performs learning of the student model by using a knowledge distillation method from the teacher model.
2 . The learning device according to claim 1 , wherein
the teacher model generates an attention matrix,
the second encoder generates an encoded input feature sequence based on a first speaker vector and an input feature sequence of the speaker of the first speaker vector,
a model parameter of the second decoder is initialized and fixed by a model parameter of the first decoder, and
the second decoder generates a second target feature sequence of a speaker of a second speaker vector based on the second speaker vector, a first target feature sequence of the speaker of the second speaker vector, the attention matrix, and the encoded input feature sequence.
3 . The learning device according to claim 1 , wherein
the second encoder generates an encoded target feature sequence based on a speaker vector and the target feature sequence of a speaker of the speaker vector,
a model parameter of the second decoder is initialized and fixed by a model parameter of the first decoder, and
the second decoder generates a second target feature sequence of the speaker of the speaker vector based on the speaker vector, a first target feature sequence of the speaker of the speaker vector, and the encoded target feature sequence.
4 . The learning device according to claim 1 , wherein
the teacher model generates an attention matrix,
the second encoder generates an encoded input feature sequence based on a first speaker vector and the input feature sequence of a speaker of the first speaker vector, and
the second decoder generates a second target feature sequence of a speaker of a second speaker vector based on the second speaker vector, a first target feature sequence of the speaker of the second speaker vector, the attention matrix, and the encoded input feature sequence.
5 . The learning device according to claim 1 , wherein
the teacher model generates a first attention matrix,
the second encoder generates an encoded input feature sequence based on a first speaker vector and the input feature sequence of a speaker of the first speaker vector,
the second decoder generates a second target feature sequence of a speaker of a second speaker vector and a second attention matrix based on the second speaker vector, a first target feature sequence of the speaker of the second speaker vector, and the encoded input feature sequence, and
the learning controller obtains a loss based on the first attention matrix and the second attention matrix.
6 . A feature conversion device comprising:
an acquirer that acquires a student model from a learning device including a teacher model that includes a first encoder and a first decoder including a first recurrent neural network, the student model that includes a second encoder and a second decoder including a second recurrent neural network, and a learning controller that performs learning of the teacher model by using a sequence-to-sequence machine learning method and performs learning of the student model by using a knowledge distillation method from the teacher model; and
a converter that converts an input feature sequence into a target feature sequence using the student model.
7 . A learning method performed by a learning device, the learning method comprising:
performing learning of a teacher model that includes a first encoder and a first decoder including a first recurrent neural network by using a sequence-to-sequence machine learning method; and
performing learning of a student model that includes a second encoder and a second decoder including a second recurrent neural network by using a knowledge distillation method from the teacher model.
8 . (canceled)