Codebook Search in LD-CELP Speech Coding Algorithm Based on Multi-SOM Structure
Mansour Sheikhan, Vahid Tabataba Vakili, Sahar Garoucy · 2009
Abstract: In the family of CELP coders, codebook search has high computational complexity. In this paper, the codebook search in low delay-code excited linear prediction (LD-CELP) G.728 coder is performed by a multi-self organizing map (SOM) neural model. A modified-supervised SOM training algorithm is also used in this work. In this algorithm, the codebook vectors are assigned to a class during training and a rejection term for codebook entries is used. The proposed neural search codebook module consists of 48 SOMs, which determine optimum index values of shape codebook. Empirical results show that the proposed model, which has an average classification rate of 97.8%, leads to 28 % reduction in execution time as compared to a traditional implementation of G.728 encoder. However, the degradations in mean opinion score (MOS) and segmental signal to noise ratio (SNRseg) are 0.16 and 0.17 dB, respectively. Key words: Codebook search • LD-CELP speech encoder • self-organizing map