2001 年 67 巻 657 号 p. 1315-1321
This paper presents a dynamic model of "Kansei" for musical chord progression using a recurrent neural network. The three-layered network is constructed based on the structure of auditory sense. The input is a time series of musical chord progression and a two-dimensional vector of Kansei information characterized by "cheerfulness-gloominess" and "stability-Instability" is outputted in real-time. After training the network, the analysis of synaptic weights of all the units gives the following results that agree well with the knowledge of psychology and physiology. 1) The "cheerfulness-gloominess" and "stability-instability" are influenced by the units that have acquired the characteristics of critical bandwidth corresponding to musical tone frequencies. Furthermore, the synaptic weights prove that "Tonality" works on the Kansei information. 2) Of six units in the hidden layer, three units are autoregressive, which result agrees well with the knowledge that sustained type and onset type neurons exist behind the cochlear nucleus. 3) In the output layer, the outputted Information is fed back to the non-outputted unit. Such a feedback circuit exists between the medial geniculate body and the auditory cortex.