Compressed Sensing-Based Speech Signal Compression

It is well known that speech signal plays an important role in our daily life all the time. The processing technology of speech signal has experienced the change from analog form to digital one. However, the digital processing leads to higher pressure on transmission and storage than its analog counterpart, it is therefore necessary to compress speech signal effec-tively. Due to the redundancy of speech signal and the mechanism of auditory perception, speech compression coding technology becomes feasible and workable. Generally speaking, the speech signal compression system is desirable with low coding rate, high reconstructed speech quality, low computational complexity and delay.In2006, compressed sensing (CS) theory was proposed by Donoho, Candes and Tao, which has a huge impact on signal processing. The main objective in this thesis is to develop a CS-based speech signal compression system based on CS theory and speech signal processing technologies. The main work and contributions are as follows:1. We study the elementary knowledge of CS theory in terms of signal sparse representa-tion, measurement matrix, signal reconstruction algorithms, etc. The key research con-tent is on dictionaries, in which signals can be sparse represented, including orthogonal bases and over-complete dictionaries. On the other hand, we analyze speech signal and know about its relevant knowledge, such as its characteristics, some basic processing technologies, coding schemes, etc.2. A new dictionary is put forward, which is called ALP basis, according to the technol-ogy of linear predictive analysis. Unlike LP basis, ALP basis pays more attention to the correlation of connected speech frames. In the aspect of speech signal sparse approxi-mation, it needs less amount of data and is of lower computational complexity in ALP basis than LP basis while the approximated effects are almost the same. In the case of DCT basis, it has worse approximated performance than those two although it is much simpler in constructing and processing procedures. 3. A CS-based speech signal compression system is proposed, which employs ALP ba-sis. Then through comparatively analyzing the performance of three systems which are ALP-based, LP-based and DCT-based respectively, we know that the proposed system is better than those two in some aspects. In addition, we discuss the corresponding changes of their performance with different compression ratio M/N.