Learning and discrimination of perceptual vowel distribution by a neural net model
Teruhiko Ohtomo, Katsunori Takahashi, Ken‐ichi Hara · Systems and Computers in Japan · 1989
Abstract The phoneme perception distribution is determined by the psychoacoustic experiment for the synthesized speech. The purpose of this paper is to attempt a learning of this distribution by a neural net model with three‐layered structure. The following observations are made as the result of learning by a neural net model. The weight of the connection from the input layer to the hidden layer has the function to extract the particular frequency range, and the weight from the hidden layer to the output layer has the function to integrate the unit output in the hidden layer, to extract the particular phoneme. It is shown by integrating those results of analysis that the weight sequence from the input layer to the output layer has the function as a filter to extract only the particular phoneme. Finally, the discrimination power is examined when a part of the connections among units is destroyed or the connection strength of some units is varied. It is shown that the discriminating power is not affected by the change of the weight of the connection between the input layer and the hidden layer, with the effect on the output unit being less than several percent. This result indicates that the neural net model has a feature of distributed memory.