Speech Reconstruction Based On Compressed Sensing And The Application Of The Speaker Recognition System

In information acquisition, the sampling rate is decided according to the Nyquist Criteria, i.e., the sampling rate must be twice the bandwidth of the signal at least. However, the huge information amount caused that signal’s bandwidth and the sampling rate is getting bigger and bigger. The hardware and complicated calculation are problems which obstacle the information technology’s development. In 2006, Donoho and Candes proposed a novel theory—Compressed Sensing (Compressive Sampling, CS). This theory is different from Nyquist criteria, according to it, just as some conditions satisfied, the signal can be recovered from very few measurements, i.e., it realized low rate sampling. With the increasing of the number of the speakers of speaker recognition, the data quantity increases. This paper focuses on speech reconstruction based on compressed sensing theory, realizes the low rate sampling of speech with little impact of system identification performance which reduces the data storage of the speaker recognition system effectively.The main works as follow:1. A research on basic principle of speaker recognition, classification of the models, extracting characteristic parameters.2. A research on compressed sensing theory, signal reconstruction based on compresses sensing.3. A research on speech reconstruction based on compressed sensing theory,speech signal is projected to a random gaussian measurement matrix and reconstructed by Optimized Matching Pursuit and Simplex base on the approximate sparsity of speech in DCT basis, analysing the influence of reconstruction SNR under different observation ratio. In order to get a better reconstruction result, an improved method was proposed for gaussian random measurement matrix based on approximate QR decomposition and the orthogonalization of upper triangular matrices. The experiments show that the improved gaussian random measurement matrix is better than the original random gaussian measurement matrix when used to reconstruct speech.4. Based on speaker recognition system of vector quantization, reconstructing speech with cs processing,analysing the influence of identification perception properties with different observation ratio. The experiments show that the data storage of speaker recognition system has a large degree and there is little impact on the rate of identification based on compressed sensing processing while there is a appropriate value of observation ratio.