Increasing quality of CELP coders by source-filter interrelation using self organising maps

There are various alternatives for secondary excitation formulation for CELP type speech coders. In this paper we present a secondary excitation codebook generation and search algorithm based on the information derived from the linear prediction filter. Source-filter interrelation is extracted using Kohonen Learning algorithm and filter parameters (LSFs) are clustered according to topographic neighbourhood. For each cluster a secondary excitation shape-gain codebook is generated. Using the class information that current LSFs belong, only the associated codebook is searched. Shape codebooks of size 128 and gain codebooks of size 32 lead to statistically indifferent synthetic voice quality according to listening test when compared to FS1016 CELP coder.