Speech Coding Based on Sparse Modeling

Abstract

In this thesis, we introduce three contributions in the field of sparse-based speech compression. The main contribution of this thesis is introducing a new backward technique so-called the Backward Replacement (BRe) that takes into consideration the impact of backward processing on the signal compression. The new technique doesn’t exploit the correlations to eliminate the weights, but it exploits the converged weights to replace the sparse vector with a sparse symmetric matrix which could be encoded efficiently. As for the second contribution, we introduce a lossybased rate saving enhancement on the BRe using an optimized approach that exploits the correlation among the atoms to reduce the replacement errors. Finally, we introduce a losslessbased rate saving enhancements which is based on hiding rows and columns in the obtained sparse matrix during the index encoding process. By comparing our approach with the backward elimination algorithms in the field of speech compression we concluded that, the BRe enhanced the compression capabilities of the forward modeling by 47% and outperforms the other backward elimination techniques whose enhancements reaches at most 15%. Also, from the obtained results we proved that, the proposed algorithm have the ability of encoding the speech signals at different bit rates with a reasonable quality. Not only, the proposed strategy proved its effectiveness on the compression but also it has a low complexity in comparison to the elimination-based backward algorithms.