Neural Network Model of the Impression Evaluation for Button Sounds

The purpose of this paper is to construct an impression evaluation model for button sounds generated when users press the buttons on car audio systems using a neural network. The dynamic characteristics of 11 kinds of button sounds obtained by their wavelet transform frequencies. Their sound pressure values are fed into the neural network model inputs. The model then corresponds with three factor scores, "esthetic", "force" and "metallic", and an evaluation value of "offensive - pleasant" as the outputs. By analyzing the internal functions of the neural network model after learning, we confirmed the model achieved a mechanism that produced four impression evaluation values from the sound characteristics. Thus demonstrating the model could attain automation of button sound design.