Linear neural networks for spectral envelope estimation
Takahiro Emoto, Udantha Ranjith Abeyratne, Masatake Akutagawa, Shinsuke Konaka, Yohsuke Kinouchi · IEICE Electronics Express · 2010
The spectral envelope is mainly determined by the shape of vocal tract and it is generally represented by linear estimation methods such as an autoregressive (AR) modeling. Neural network (NN) is general tools for modeling nonlinear function. Nonlinear aspects of NN have been emphasized. However, in this study, a novel approach to spectral estimation, which is based on the use of a linear feed-forward NN is discussed. Example of the linear network for estimating the spectral envelope of vowel sound is presented. The proposed linear NN technique leads to interesting new paradigms which could not have been considered in advance.