@techreport{oai:ipsj.ixsq.nii.ac.jp:00214085,
 author = {向田, 圭汰 and 森, 大毅 and Keita, Mukada and Hiroki, Mori},
 issue = {22},
 month = {Nov},
 note = {感情次元による連続的な感情表記法に基づいた声質変換手法を提案する．一般的な離散的な感情表記法に基づく感情の声質変換は，変換元と変換先の感情状態に対応するコーパスを用意して学習する．しかし，次元に基づく連続的な感情表記法ではデータセットを変換元と変換先に分割することはできない．本報告では，変換元と変換先の感情次元の差分をサンプリングにより生成することで声質変換モデルを学習する方法を提案する．このモデルは声質変換を担う Generator と変換音声の肉声らしさを検査する Discriminator の競合学習に基づく．Generator には音声の特徴量に加え変換先への感情次元の差分を入力する．Discriminator には音声の特徴量に加えその感情次元を入力し，音声の肉声らしさとその感情次元の妥当性を判別する．, We propose an emotional voice conversion method based on the emotion dimensions. Conventional emotional voice conversion assumes a dataset that consists of disjoint subsets of categorical emotion. However, it is impossible to divide the dataset into the source domain and the target domain when adopting the emotion dimensions. In this paper, we propose a method of constructing an emotional voice conversion model by the sampling of the difference to the target emotion over the dimensional space of emotion. The model is based on the competitive learning of the Generator that performs the voice conversion, and the Discriminator that assesses the genuineness of the converted speech. The Generator receives the speech features as well as the difference of emotion dimensions to the target emotional states. The Discriminator receives the speech features as well as the emotion dimensions of the speech, to check whether the emotion is being expressed by the input speech.},
 title = {感情次元の操作を目的とした声質変換手法の提案},
 year = {2021}
}