Analysis of a Human Sensibility Model for Musical Chord Motion Using Neural Networks

Abstract

Many psychologists have carried forward the research on “Human Sensibility”. Past techniques and modeling of “Human Sensibility”, however, were inadequate for engineering purposes, so a systematic approach to “Human Sensibility” is currently emerging. Music is a particularly interesting subject on “Human Sensibility” because it has a strong influence on the human mind and arouses various emotions.
This paper presents a model of “Human Sensibility” for music. Music consists of three major factors, namely, rhythm, melody and harmony. Of the three, we choose a human sensibility model for harmony. The model is constructed using a neural network and senses musical chord motion. Each chord consists of three simultaneous tones, and the input to the neural network is a set of two successive chords from a MIDI keyboard. There are two outputs from the model, which characterizes the “cheerfulness-gloominess” and “consonance-dissonance” of the chord motion.
Analysis of the hidden and output units of the network provides the following results.
1. The tone interval between the middle tone and the others is extracted in the first hidden layer and is amplified in the second hidden layer.
2. The “cheerfulness-gloominess” is generated in the output layer by addition of primary and secondary chord values. On the other hand, the “consonance-dissonance” is generated solely from the secondary chord value.
The above results agree well with the knowledge of musical theory.