Low data rate adaptive transform coding for parametric representation of speech signals

This paper investigates the potential for data rate reduction through the use of a phonetic segmentation criteria to identify spectral regions exhibiting slowly evolving features. A two dimensional discrete cosine transform (DCT) is then applied to the encoding of the segmented regions represented in this case by line spectral pair (LSP) filter parameters. This adaptive block transform code (ABTC) technique has been integrated into the CELP framework to realise an average worst case reduction of 27% over the standard CELP quantiser for an equivalent average quantisation distortion of 1.3dB. Both objective and subjective performance are compared to the standard CELP quantiser.