Proposals Of New Operator For Neural Networks On The Peak Detection And The Interpolation Problems
Hiroyuki Kamata, Nobuharu Matsumoto, Akimasa Ishida · Information Sciences, Signal Processing and their Applications · 1996
We propose a new method for calculating the spectrum envelope of human voice using neural networks. In this paper, we first show a method which extracts the spectrum peaks based on the voiced speech using neural networks, and next try to interpolate the extracted pe;ak data. Usually, all of the input layers of neural networks require to get the data. In this study, inputted data go through the network, the data that have not been given are interpolated by the other inputted data. It is assumed that the placements of input data are given at noneven intervals. As the method, the Neville interpolation is applied. When the Neville interpolation is adopted into the linea neural network, the value of each weight data is decided by only the information which shows the placement of input data. In this study, all of the processes are realized1 by the neural networks including the proposed new nonlinear operators.