The Research Of Chinese Speech Sparse Representation Based On Compressed Sensing

Compressed sensing theory has been a hot research topic of signal processing area inrecent years. It can implement a new compressive sampling method without the restriction ofNyquist sampling theorem. The sparse representation of the signal is an important part ofcompressed sensing theory. Whether the original signal can get a more sparse representationdirectly affects the signal recovery from compressed sensing measurements. This paperaiming at the Chinese speech signal, studies three kinds of sparse representation methods,including the DCT domain residuals domain in linear prediction and wavelet tree model,applied to the speech compressed sensing frame for the purpose to improve the quality ofreconstructed speech, the main research works are listed as follows:1. Speech compressed sensing based on sparse pretreatment of DCT domain is proposed.To get a more accurate construction of speech signal in DCT domain, then based on thespeech signal approximately sparse in DCT domain, two kinds of sparse pretreatmentmethods based on sparsity and threshold are presented. The pretreatment sacrifices the signalaccuracy for the sparsity in the transform domain. At last apply the signal after sparsepretreatment to compressed sensing. Simulation experiments demonstrate the validity of theimproved method, and quality of reconstructed speech is improved.2. Linear predictive speech compressed sensing based on the improvement of circulantmeasure is proposed. The sparse representation of speech signal in the residual domain oflinear prediction, which brings more data of linear predictive coefficients to transmit whilebuilding the sparse transformational matrix; this chapter presents a way to save the linearpredictive coefficients into diagonal matrix in order to build circulant matrix, with which tomeasure speech signal in a circulant way; and then extract linear predictive coefficientsfrom the circulant matrix to build sparse transformation matrix in residual domain of linearprediction. Simulation experiments demonstrate that the circulant matrix with predictivecoefficients included has good construction performance, linear predictive speech compressedsensing has better construction results, and the novel way decreases the data by2.4%at least.3. Speech compressed sensing based on measurement adapted to the sparsity in wavelettree is proposed. As to the mismatch problem between the nodes number in the tree model andmeasurement of speech signal, improve the initial support set in the reconstruction algorithmof wavelet tree model; based on experiments result, get the nodes number of bestreconstruction under a fixed number of measurements,then according to the sparsity inwavelet tree of speech frames, good ones get more measurements and bad ones get less,change the number of tree nodes with different measurements.Simulation results demonstratethat the wavelet tree model can guarantee good sparsity, different sparsity of speech framesusing different measurements, and use the best tree nodes number, reconstructed SNR getimproved. Above the three methods mentioned, the third one keeps a good balance betweentime and accuracy.