The Vector Sum Excited Linear Predictive(VSELP) coding has the potential for producing high quality synthetic speech at bit rates as low as 4.8kbps. Unfortunately, the computational complexity of VSELP coding is prohibitive for real-time applications.
In this paper, three techniques to reduce the computational complexity of the VSELP coding are proposed. Firstly, the computational load to calculate the weighted synthesis filter responses to every basis vectors is significantly reduced by using overlapped sparse basis vectors. Secondly, the proposed preprocessing step reduces the search space efficiently before applying every linear combinations of the basis vectors to the VSELP codebook search procedure. It determines whether the combination coefficient of each basis vector can be fixed using heuristics so that the number of combinations decreases. Finally, to determine the pitch period only pre-selected candidates are applied to the pitch period search procedure. The candidates are selected by comparing them with the ideal excitation sequence which is produced by inverse filtering.
Then, all the proposed techniques for complexity reduction are combined to develop a VSELP coder with reduced complexity. It has been shown that the proposed coder retains good quality of synthetic speech and it is more than twice as fast as the original coder.
