Construction of automatic evaluation model for timbre impressions

The purpose of this research is to construct and automate an impression evaluation model for timbre using a neural network. The three-layered network, which is based on the structure of auditory system (cochlea), consists of seven input nodes, four hidden layer nodes, and three output nodes. The input is a pattern of tone pressure levels for harmonics from the 1st tone (fundamental tone) to the 7th tone, and the output corresponds to three factor scores, “clear-unclear,” “thick-thin” and “sharp-soft.” Following learning of the neural network model, synaptic weight analysis of all neurons yields results that agree well with existing psychological and physiological knowledge. We conclude that (1) Each harmonic equally influences the impression evaluation value for “thick-thin.” (2) Harmonic components exceeding the 4th harmonic have an inverse effect on the “clear-unclear” impression. (3) Two neurons possessing a lateral inhibition property exist in the hidden layer, and extract the 6th and 7th harmonics. Another neuron in the hidden layer extracts the 5th, 6th and 7th harmonics. These three neurons determine the “sharp-soft” impression.