Speech Compression And Reconstruction Based On Multi-scale Compressed Sensing

With the rapid development of information technology, the need to deal with the increasingamount of information is growing, and the inadequacies of traditional information collectionmethods are slowly manifested. How to collect less data, and make the collected data containscomplete information is a challenge.In recent years, a theory called compressed sensing(CS) is proposed, and so to solve theproblems mentioned above becomes possible. The compressed sensing theory and previouslyproposed information collection theory are different essentially. For the compressed sensingtheory, the sampling frequency is not related to the frequency of the signal. But the theory needsthat the signal can be sparse representation in a domain. This theory has a very importantsignificance in reducing sampling rate and saving storage resources.Since the compressedsensing theory is put forward, Researchers in all countries are attracted. The researchers didextensive research related to the theory, and made some achievements.In this paper, compressed sensing theory is applied to speech Signal processing; we putforward the speech signal processing scheme based on wavelet transform and compressedsensing. In order to obtain a better reconstruction of speech signal, we studied some importantparameters in the scheme, such as the length of the frame, the choice of wavelet base and theprocessing mode about the high frequency component.We did a lot of experiments using matlabsimulation tools and didn’t consider the coding rate in the scheme. Though the quality of thereconstruction of speech signal is very good, the coding rate is too high. In the waveletdecomposition process, we just decomposition low frequency coefficient, and the high frequencywavelet coefficient is not further decomposition. Considering this shortcoming, this paper alsoputs forward the speech signal processing scheme based on wavelet packet and compressedsensing, Given attention to the reconstruction quality of speech signal and code rate, weresearched the main parameters in the scheme and did a lot of simulation experiments usingmatlab simulation software. The results are analyzed.After the conversion of the speech signal, no matter the conversion is wavelet transform orwavelet packet transform, the observation points have been increasing. The higher the order ofwavelet bases, the more increase in the number of the observation points. However the lowerorder will make a poor quality of the recovered speech signal. How to make the observationpoints do not increase or increase less after wavelet transform, and also make the quality of the recovered speech signal, this is one problem that need to be solved.In the scheme that we use wavelet packet transform to deal with, the vector quantizationcodebook has a great influence on the quality of the recovered speech signal. Therefore, how toselect the representative training sequence constructed a better codebook is an urgent problem.In the scheme that we use wavelet packet transform to deal with, although we consider thefactor of the encoding rate, the coding rate we get at last is not very low, so the scheme that wedesign does not reflect the advantages of compressed sensing theory.