Speech Enhancement Research Based On Small Dictionary Learning

In actual life,the speech is always affected by various kinds of noise.Speech enhancement extracts the pure speech from the noisy speech as far as possible to improve the quality and intelligibility.In recent years,the deepening of the study of sparse representation and dictionary learning provides an effective solution to the speech enhancement.Different from the conventional methods which enhance the speech quality by estimate and restrain the noise,the original speech is reconstructed according to the characteristics of the speech.On the basis of analyzing the theoretical knowledge and classical algorithms of sparse representation and dictionary learning,an improved speech enhancement algorithm based on small dictionary learning was proposed in this thesis.The main contents of the thesis are as follows:1.Introduce the significance and basic theoretical knowledge of speech enhancement,the research status of speech enhancement algorithms and the quality evaluation methods.Analysis the principle of speech enhancement algorithms based on sparse representation and dictionary learning.2.To solve the problem of speech distortion caused by the traditional speech enhancement method based on small dictionary learning,a speech enhancement algorithm based on adaptive small dictionary in subspace domain was proposed in this thesis.Firstly,a fully completed small dictionary was constructed by using the noisy speech eigenvalues in the subspace domain.Then,the sparse representation and dictionary updating of the noisy speech were performed by K-SVD.In the OMP,the correlation threshold and energy threshold were set to control reconstruction and iterations.Finally,the experimental results show that the new algorithm has higher SNR and PESQ,and can reduce the speech distortion and improve the speech quality.3.To solve the problem of source confusion caused by the traditional joint dictionary learning algorithm in the sparse reconstruction,a speech enhancement algorithm based on discriminative joint small dictionary learning was proposed in this thesis.The pure speech small dictionary and the noise small dictionary were trained by K-SVD,and they were joined into a small dictionary.Then the noisy speech samples were represented sparsely based on the joint small dictionary by LARC,and dictionary discriminative fidelity items were added to reduce the correlation between pure speech dictionary and noise dictionary.Finally,the pure speech was estimated by sparse representation coefficients.The simulation and performance evaluation verify the effectiveness and feasibility of the proposed algorithm.