Singular value decomposition and its modelling of speech excitation

The desire to be able to design high quality low bit rate speech codecs has caused much recent research to be directed at the efficient representation of the excitation function information of the new established family of LPC codecs. The authors concentrate on the use of the singular value decomposition (SVD) procedure to construct an optimal orthogonal transformation domain in which the speech signal can be analysed. They consider the main characteristics and properties of the singular vectors of the LPC impulse response matrix that form the basis for the new SVD domain. Elementary coding approaches that seek to exploit the SVD domain properties are presented. A particular consideration is made of the characteristics and properties of the LPC excitation function in the transformation domain