[国内学会] A comparative study on modeling and controlling emotional acoustic parameters in neural networks based Japanese and Spanish speech synthesis

情報処理学会 第114回音声言語情報処理研究発表会

In neural network based speech synthesis, adding a one-hot vector to the input is an easy, intuitive but useful way to model multiple speakers or multiple language. In this paper, we use the one-hot vector for modeling and controlling emotional acoustic parameters. We have used Spanish and Japanese databases having the multiple acted emotional speech uttered by professional speakers in our experiment and will show the performance of the one-hot vector approach.