Compressed Sensing based Speech Compression using Dictionary Learning and IRLS algorithm

摘要

Compressed sensing (CS) has gained much interests in the speech processing community and especially in compression. The success of this appealing paradigm relies, heavily, on the sparsity of speech signals in a given dictionary. Thanks to machine learning, it is possible to go beyond the limits of analytical dictionaries by designing learned dictionaries, which are more able to fit the nature of data. In this paper, we propose speech compression scheme based on CS using K-singular value decomposition (K-SVD) algorithm to learn a dictionary for sparse representation and iteratively reweighted least square (IRLS) algorithm for signal reconstruction from randomly generated measurements. Different sizes of dictionaries are trained and compared with discrete cosine transform (DCT). Signal quality results, measured with perceptual evaluation speech quality (PESQ) revealed that learned dictionary using K-SVD improves the performance of speech compressed sensing comparing to DCT.