| 摘要 |
PROBLEM TO BE SOLVED: To provide a learning data creation technique for creating learning data which is more suitable for a target domain.SOLUTION: A learning data creation device comprises: a first sound feature amount extraction part 1011 for extracting a sound feature amount series from a voice signal of each speech production of a target set; a component decomposition part 102 for determining a target construction ratio which is a construction ratio of each component in the target set; a second sound feature amount extraction part 1012 for extracting the sound feature amount series from a voice signal of each speech production of a base body set; a component inclusion amount calculation part 103 for calculating an inclusion amount of each component in each speech production included in the base body set; and a speech production selection part 104 for creating learning data, by selecting the speech production so that the construction ratio of the components in the learning set becomes similar to the target construction ratio, out of the speech productions included in the base body set.SELECTED DRAWING: Figure 1 |
