Codebook ordering techniques for address-predictive VQ

The authors have investigated several algorithms aimed at either generating an ordered VQ (vector quantization) codebook or putting in order an already existing one, so that subsequent application of APVQ (address predictive VQ) or L-APVQ (lossy APVQ) will provide good results. In general, it appears that, as long as a reasonably ordered codebook is available, good performance is always obtained and little can be gained by the use of a more sophisticated algorithm. Therefore, a fast technique, like NMA (nearest merger algorithm), FLA (farthest insertion algorithm), or even the trivial MVA (maximum variance algorithm) (limited to lossless APVQ), appears to be a sensible choice. If CPU time is not a concern, however, the BSA (binary switching algorithm), properly tuned for the specific application, can make it possible to squeeze out some more bits in encoding. If a new codebook needs to be generated, the Kohonen algorithm is probably worth using due to its wide range, good performance and robustness.<>