Research On The Key Issues Of Compressed Speech Sensing

Speech is the most convenient way for human to communicate with each other. In thedigital era, how to model the speech signal in order to obtain fewer samples without affectingthe quality of the speech, is the research focus in the field of speech signal processing.Compressed sensing (CS) based on the sparsity of the signal is a new theory of linear andnon-adaptive sampling. Compressed sensing shows that the sparser of the signal, the better thecompression performance is. The sparsity or compressibility of the signal is a necessarycondition for achieving good performance of compression reconstruction. As the speech signalis compressible, it can be sampled at the lower rate than the Nyquist sampling rate withoutdistortion, which has brought great convenience to the signal sampling, storage, transmissionand processing. The combination of CS theory and speech signal processing means thesubversion to traditional speech analysis methods based on the Nyquist sampling theory. Withobservations in the CS theory to replace traditional speech sample value is bound to the signalcharacteristics fundamental change, thereby affecting the entire speech signal processing theoryand technology systems of a variety of applications. The CS is used in the field of speech signalto explore new ways of the speech signal processing, which has a good practical significance.Research on the key issues of compressed speech sensing is the foundation to practicalapplication. In this thesis, the main work and innovation are as follows:(1) According to the sparsity of speech signals in the discrete cosine transform basis (DCT),the CS framework of speech signal is described and CS random observation matrix andreconstruction algorithm are selected firstly, and then Via basis pursuit (BP) and orthogonalmatching pursuit (OMP), it is demonstrated that the performance of reconstruction is correlatedwith the dimensions of measurement matrix, the length of frames, the shift of frames and theshape of the window.(2) Using the correlation of speech signals and the sparsity on the residual, we propose anew algorithm of compressed sensing of speech signals based on overcomplete linear prediction(OLP) dictionary from codebook constructed by the linear prediction coefficients of trainingsignals. The proposed method not only improves the performance of reconstructed speechsignals based on Gaussian measurement matrix and basis pursuit, but also does not need to know prediction coefficients of testing signals. We compare OLP with discrete cosine transform(DCT) by applying both methods to compressed speech sensing, and the reconstructed speechsignal is evaluated by the objective and subjective evaluations. The experimental results showthat the performance of compressed speech signal sensing based on OLP is3~8dB higher thanDCT with the same number of measurements.(3) To deal with the problem that the reconstruction performance of compressed sensing ofspeech reconstructed by basis pursuit (BP) is degraded by white Gaussian noise interference, anadaptive basis pursuit de-noising (ABPDN) method is proposed for compressed sensing of noisyspeech. This method adaptively selects the prime regularization parameter according to the SNRof original speech signal and enhancement is achieved as it is applied to compressed sensing ofnoisy speech. And then an adaptive basis pursuit denoising tracked post-denoising method is putforward. Finally, pre-denoised tracked reconstruction method is proposed, which employs anadaptive speech enhancement method based on overcompletely sparse representation in adata-driven dictionary and basis pursuit. The performances of the three methods by the objectiveand subjective evaluations are compared. For compressed speech sensing of the measurementwith noise, an adaptive basis pursuit de-noising method is proposed. The reconstructionperformance of speech signal based on ABPDN is superior to BP.(4) For compressed sensing of the speech signal in the wavelet domain, a matrix form ofSym wavelet decomposition and synthesis is deduced, keeping the length of the coefficient nomore than the length of original speech signals, and then we propose a framework of speechmultiscale compressed sensing (MCS) and an adaptive multiscale compressed sensing (AMCS)method by analyzing sparsity of different wavelet levels of speech signals.We compare AMCSwith MCS by applying both methods to speech compression and reconstruction, and thereconstructed speech signal evaluated by the objective and subjective evaluation is applied tospeaker recognition. The experimental results show that the reconstruction performance ofspeech signal based on AMCS is superior to MCS.