The lack of an efficient technique to encode the LPC excitation functioii impedes the development of high quality speech codecs at very low bit rates, below 4 kbits. In this work we examine the use of singular value decomposition (SVD) to define an orthogonal data-dependent transformation domain into which the associated time domain LPC residual signal is mapped. We seek to understand and evaluate this approach. The main features of the method, detailed by Atal , concern the representation of the LPC excitation function as a linear combination of an orthogonal set of signals, which are the eigenvectors of the LPC filter impulse response matrix. We consider SVD-based CELP coders and the suitability of the basefunctions to describe both the filter and the excitation signal information.
Bibliographic reference. Corney, P. / Mason, J. S. (1991): "A new perspective on LPC excitation using singular value decomposition", In EUROSPEECH-1991, 1315-1318.