音の時系列データに対する聞き手の印象評価結果に基づく音生成システム

抄録

We propose a method of generating a sound which is expected to give a listener a good impression. We aim to use the system to help a robot give a person a good impression in interaction between the robot and the person. In the proposed method, an impression of a sound is estimated based on results of a listener's evaluation to a sound. A spectrogram of a sound is used to learn the relationship between a sound and an evaluation. The evaluation of a sound is estimated based on the spectrogram and not based on the criteria of a sound. Therefore, the proposed method is expected to generate unpredictabile sound. The relationship between the value of an evaluated impression and the data of a sound is estimated by using RNN. We conducted some experiments. It was composed of gathering data of evaluation to some sounds, estimating a listerner's evaluation to sound and generating sounds which was expected to give a listener a good impression.