COMPRESSED SPEECH SIGNAL SENSING BASED ON THE STRUCTURED BLOCK SPARSITY WITH PARTIAL KNOWLEDGE OF SUPPORT

Structural and statistical characteristics of signals can improve the performance of Compressed Sensing (CS). Two kinds of features of Discrete Cosine Transform (DCT) coefficients of voiced speech signals are discussed in this paper. The first one is the block sparsity of DCT coefficients of voiced speech formulated from two different aspects which are the distribution of the DCT coefficients of voiced speech and the comparison of reconstruction performance between the mixed program and Basis Pursuit (BP). The block sparsity of DCT coefficients of voiced speech means that some algorithms of block-sparse CS can be used to improve the recovery performance of speech signals. It is proved by the simulation results of the mixed program which is an improved version of the mixed program. The second one is the well known large DCT coefficients of voiced speech focus on low frequency. In line with this feature, a special Gaussian and Partial Identity Joint (GPIJ) matrix is constructed as the sensing matrix for voiced speech signals. Simulation results show that the GPIJ matrix outperforms the classical Gaussian matrix for speech signals of male and female adults.

Structural and statistical characteristics of signals can improve the performance of Com- pressed Sensing （CS）. Two kinds of features of Discrete Cosine Transform （DCT） coefficients of voiced speech signals are discussed in this paper. The first one is the block sparsity of DCT coefficients of voiced speech formulated from two different aspects which are the distribution of the DCT coefficients of voiced speech and the comparison of reconstruction performance between the mixedl2/l1 program and Basis Pursuit （BP）. The block sparsity of DCT coefficients of voiced speech means that some algorithms of block-sparse CS can be used to improve the recovery performance of speech signals. It is proved by the simulation results of thel2 / reweightedl1 mixed program which is an improved version of the mixedl2 /l1 program. The second one is the well known large DCT coefficients of voiced speech focus on low frequency. In line with this feature, a special Gaussian and Partial Identity Joint （GPIJ） matrix is constructed as the sensing matrix for voiced speech signals. Simulation results show that the GPIJ matrix outperforms the classical Gaussian matrix for speech signals of male and female adults.