Research On De-noising In Speech Compressed Sensing System

Facing the massive amounts of data brought by Age of Big Data, how to store and how toprocess information? Both of the problems bring a big challenge for the current system. The firststep of the current system based on digital processing technology is sampling. The traditionalNyquist sampling theory produces a large number of sampling data in the face of signals ofbroadband. In addition, the following compressed technology of the massive amounts of sampleddata wastes precious resources. In recent years, compressed sensing, a new sampling theoryemerges, which creatively solves the problem of the high sampling rate and high redundancy of thedate sampled in the classic way. The inherent sparse property of speech signal determines thecombination of speech signal processing and compressed sensing is an inevitable choice. This willbring enormous changes in the field of speech signal processing. In this context, this thesis studiesone of the problems that must be solved when applying speech signal processing based oncompressed sensing to engineering practice: the de-noising methods in the speech compressedsensing system.The main work and innovation of this thesis are as follows：Firstly, this thesis analyzes the key points of compressed sensing in detail, including therequirement of signal’s sparsity, the projection process of observation matrix and the theory ofnonlinear reconstruction. Then, for speech signal, the feasibility and quality of processing results ofcompressed speech sensing from the three aspects mentioned above is analyzed. And on this basis,this thesis analyzes the effects of environmental noise on compressed sensing system. All theseworks above lay the foundation for the following research.Then, to reduce the effects of noise on the compressed sensing system, this thesis discusses apre-processing method for noisy speech. Due to the projection process of observation matrix, itbecomes very hard to process the measurement sequences. A method using wavelet decompositionis proposed, which uses sparse decomposition based on the K-SVD algorithm to process the lowfrequency coefficients and wavelet threshold de-noising algorithm for the high frequencycoefficients. This method can guarantee the sparsity of the signal as well as de-noising, which helpsthe following process of compressed sensing. The experimental results show that the proposedmethod can achieve a good noise suppression effect as well as a good guarantee for CS.Thirdly, this thesis studies a de-noising method for measurement sequences under specialobservation matrix. After the analysis the different property of speech and noise under row echelon observation matrix, reference to traditional time-frequency processing technology, a method usingwavelet threshold de-noising algorithm is proposed for speech degraded by white Guass noise. Inaddition, when the noise is small, an improved scheme for noise estimation is also put forward,which improves the reconstruction results greatly for high SNR. It is demonstrated that this methodachieves better performance than the traditional speech de-noising methods with only1/4of theamount of the original date.In the end, a high speed and robust reconstruction algorithm is studied. Traditional algorithmsreconstruct the signal with only one prior information: sparsity. So a new reconstruction algorithmwhich adds the aggregation property in DCT domain and intra frame relativity of speech to theclassic CoSaMP algorithm is proposed. It not only reduces the reconstruction time by20%, but alsoimproves the robustness to noise.