Research On Speech Enhancement By Combining EEMD And K-SVD

Speech is one of the most important and most common way for communication.Researches on speech signal processing is very popular, such as: speech enhancement, speech coding, speech recognition and speech synthesis, etc. In general, researchers choose pure speech signal for their studies. But real speech signal collected in natural environment is often contaminated the ambient noise. The existence of noise will reduce the quality and intelligibility of speech signal, which make a serious impact on the performance of speech signal processing system. The purpose of speech enhancement is to obtain pure speech signal from noisy speech signal efficiently. The thesis focused on speech enhancement by combining with the K-singular Value Decomposition(K-SVD) algorithm and Ensemble Empirical Mode Decomposition(EEMD) method. The main works and originalities are as follows:Firstly, the thesis proposed a novel speech enhancement algorithm by combining EEMD and K-SVD dictionary training algorithm is proposed. The EEMD algorithm is firstly employed to obtain intrinsic mode function( IMF) components from noisy speech. The cross-correlations and autocorrelations of each IMF are calculated from the IMFs to filter out the noisy IMFs. Meanwhile, the transition IMF components are again decomposed with EEMD to further remove the noisy component. The remained original IMFs alone with the remained transition IMFs are then superimposed to generate the new noisy speech. The new noisy speech is then sparse de-composed by the K-SVD dictionary-training algorithm with an over-complete dictionary trained from clean speech. Enhanced speech is obtained by recovering the speech signal from sparse coefficient vectors. Both subjective experiments and objective results show that the proposed algorithm can effectively remove noise from speech signal. The proposed algorithm achieves significant de-noising results than the spectral subtraction, wavelet threshold de-noising algorithm and K-SVD dictionary-training algorithm in low SNR situation.Secondly, the thesis proposed a double-threshold sparse adaptive matching pursuit(DTSAMP) algorithm for speech enhancement. The proposed algorithm add a new threshold,which is calculated from the energy of pure speech signal, to the traditional sparsity adaptive matching pursuit(SAMP) algorithm that only use single residual threshold. The noise energy is firstly estimated from the noisy speech signal. The estimation value of pure signal is thenobtained. When the energy of reconstructed signal great than 1.2 times of the pure signal energy in each iteration, the iteration is stopped and the speech signal is outputted.Experimental results show that the proposed algorithm can effectively reduce the restoration of the noise component of noisy speech and improve the signal-to-noise ratio of the original noisy speech comparing to the traditional Orthogonal Matching Pursuit(OMP) algorithm and SAMP algorithm. At the same time, the proposed algorithm can also speed up the run time more than two times by reducing the number of iterations.