Enhancement of VSELP speech coding for North American digitalcellular

摘要

We consider enhancement of the North American IS-54/136 vector sumexcited linear prediction (VSELP) speech coding in the sense of reducingcomplexity or improving speech quality. For complexity reduction, wefind a few techniques which can reduce, respectively, the pitch andexcitation search complexities by approximately 40-50% at a moderateSSNR loss of roughly 0.5 dB. For quality improvement, we considertree-structured pitch and excitation search. Aside from demonstrating aroughly 0.3-1 dB SSNR gain, we present a technique which cansignificantly reduce tree-search complexity. And we analyze thesimulation results to obtain rule-of-thumb formulas for quasi-optimaldetermination of the tree parameters under different complexity andperformance conditions