Speech Signal Processing And Application Based On Compressed Sensing

Digital sampling technology of speech signal is based upon the Nyquist sampling theorem, as we all know, sampling rate is not less than the Twice the highest frequency of the speech signal and there will be a lot of redundant data. So compression encoding is required for saving storage resources or transmission bandwidth before the storage or transport. For information collection technology, compressed sensing opens up new roads and make it possible to sample below the Nyquist rate. Based on this, studying the compressed sensing applying in speech signal processing and designing the speech signal coding system. The main contents are as follows:First, introducing the basic principles of compressed sensing and its key technology. Then studying the speech signal sparse represent and mainly studying sparseness of DCT-based and wavelet based.Second, studying that compressed Sensing uses in voice processing. In this section, Studying how to construct measurement matrix and the role of different measurement matrix in reconstructing speech signal; Studying how to classify reconstruction algorithm and different reconstruction algorithms for voice remodeling speech reconstruction. Using matlab to do the simulation, it can be seen from the experimental results that using the best measurement matrix can obtain reconstructed signal with minimal distortion under the same conditions. But random measurement matrix is the most easy to construct. And it can be seen that using convex optimization can obtain best reconstructed signal under the same conditions. But greed tracking algorithm is with lower computational complexity and with less computing time.Third, applying compression perception theory to speech compression coding and designing a new coding system. Selecting sparse matrix need to combine with the specific requirements of the coding system. Determining the measurement matrix and reconstruction algorithm in the same way. Through the contrast experiments in the original speech and reconstruction speech, making a concrete study of the coding scheme under different parameter Settings for speech signal reconstruction. The method using random Gaussian matrix observing the speech signal on the encoding side, obtaining fewer observations, then further compressing the data using vector quantization coding. In the decoder, decoding by vector quantization, getting observations based on the speech signal sparsity in the discrete cosine domain, then reconstructing speech signal using orthogonal matching pursuit algorithm. The purpose of the algorithm is to reduce the computational complexity and delay on the premise of guarantee the quality of speech signal reconstruction. Experimental results show that the mono audio signal whose sampling rate is 44100 Hz, quantitative is 16 bit and bit rate is 705.6 Kbps could be compressed to around 100 Kbps, the compressed speech signal still has good voice quality, at the same time the algorithm has lower time delay.