A Kansei-Based Sound Modulation System for Musical Instruments by Using Neural Networks

摘要

This study describes a sound modulation system based on the use of a neural network model. The inputs to the model are a) a basic, original sound wave, and b) the degree of Kansei, while the output of the model is modulated sound depending on the degree of Kansei. The degree of Kansei is the numerical value that expresses the modulation level based on a Kansei linguistic expression, such as hardness or brilliance. In the experiment, the models are constructed for the sounds of piano and Marimba. Three types of training data are used for each sound, and the degree of Kansei is assigned manually for each dataset. By changing the degree of Kansei at the input of the model, we have validated that each model could appropriately modulate the basic sound. In addition, the modulation results are illustrated for one octave of piano sounds. The potential of our proposed model and future work are also discussed.