On the representation/quantisation of variable size magnitude spectral information in prototype interpolation coders

Within the context of low bit rate prototype interpolation and sinusoidal coding, the selection of an accurate technique for the representation of variable size magnitude spectral information is a very important issue. Four possible representation methods have been investigated. These are: (a) linear interpolation, (b) band-limited interpolation, (c) all-pole modelling in the frequency domain, and (d) all-pole modelling in the time domain. The performance of these methods has been evaluated on the basis of recovered speech signals using informal subjective listening tests and subjectively meaningful objective distortion measurements. When combining these results with those presented by Papanastasiou and Xydeas (see Proc. of IEEE Inter. Workshop on Intelligent Signal Proc. and Comm. Systems, Singapore, p.731-5, 1996), it is concluded that a variable size, integrated “transformation and quantisation” approach, in coding the perceptually important magnitude spectral information, is superior to conventional, two step “transformation followed by vector quantisation” schemes and is therefore an essential process in achieving high quality speech at or below 2.4 kb/sec