Use of quasi-newton methods and delta learning rule for adaptive wavelets in speech signal representations

摘要

Training can be observed as a mapping from an input space to an output space.Adaptive wavelets are an effective tool in speech signal approximations as weighted linear combination of translated and dilated mother wavelets.The objective is to minimize the difference between original and approximated signals by tuning wavelet parameters adaptively.Dilation and translation parameters in wavelet (hidden) layer are adapted by Quasi-Newton methods and coefficients between wavelet and output layer are tuned by Delta learning rule.This algorithm shows better convergence than conjugate gradient algorithm in speech signal approximation.