@techreport{oai:ipsj.ixsq.nii.ac.jp:00146174,
 author = {中鹿, 亘 and 滝口, 哲也 and Toru, Nakashika and Tetsuya, Takiguchi},
 issue = {2},
 month = {Nov},
 note = {本研究では，音響特徴量・音韻特徴量・話者特徴量の3つを変数とする Three-Way Restricted Boltzmann Machine(3ＷRBM) を用いて音声モデリングを試みろ．3WRBM はそれぞれの変数のユーナリーポテンシャル，2 変数間のペアワイズポテンシャル，そして 3 変数間の Three-way ポテンシャルを総和したエネルギーに基づく確率密度関数である．本研究では，音響・音韻・話者特徴量の Three-way ポテンシャルを話者正規化学習・話者適応の観点から適切に設計する．一度モデルの学習が終われば 3 変数間の関係性が捉えられ，各特徴量の相互条件付確率を簡単に計算することができる．3WRBM による音声モデリングの性能を評価するために，本稿では声質変換実験と話者認識実験の結果を報告する．話者認識実験における話者特徴量は与えられた音響特徴量から尤度最大下基準により推定することで求めることができ，声質変換は，推定された音韻'情報と，切り替えた話者情報から音響特徴量を推定することで実現される．, In this paper, we argue the way of modelling speech signals using improved three-way restricted Boltz mann machine (3WRBM) where acoustic features, latent phonological features, and speaker-identity features are considered. The 3WRBM is an energy-based probabilistic model that includes three kinds of potentials: unary potentials of each variable, pairwise potentials of every two variables, and three-way potentials of the three variables. In our approach, we design the three-way potentials properly in the speaker-adaptive training (SAT) manner. The optimized model captures the relationships between the variables, enables to compute conditional probabilities of each variables, and is appliable to many tasks in speech signal processing. For example, estimating speaker-identity features given acoustic features is used for speaker recognition. Another example is estimating acoustic features from the phonological features that are estimated given source speaker's acoustic features and the desired speaker-identity features; that is voice conversion. In our experiments, we evaluate the effectiveness of the speech modelling through a voice conversion task and a speaker recognition task.},
 title = {制約付き Three-Way Restricted Boltzmam Machine を用いた音響・音韻・話者情報の同時モデリング},
 year = {2015}
}