Reduced candidate mechanism for an algebraic code-excited linear-prediction codebook search

Abstract

Underlain by algebraic codebook search, a reduced candidate mechanism (RCM) is proposed in this work to reduce the search complexity required. To begin with, the contribution made by each pulse pertaining to each track is evaluated, and then the hit probability of the top pulses in terms of contribution against the optimal pulses identified by a full search is found. Based upon such analysis, the RCM approach proposed performs a search procedure after the top N pulses, 1≤N≤8, in contribution sorting are chosen as candidate pulses. It is demonstrated by experiments that an application of RCM for N=2, designated as RCM-2, to G.729A yields a search complexity that is 5% of that required in G.729A, 25% of a global pulse replacement method (iteration=2) and 33% of a iteration-free pulse replacement method at the cost of 1% speech quality. The RCM proposed is validated as a means to cut the computational complexity considerably.