Research On The Speech Denoising Based On Sparse Coding Technology

Speech is widely used in our daily life. But the speech is often degraded by interferers in practical applications. To deal with this problem is particularly important.Speech denoising is to recover the speech signal from the speech degraded by realworld interferers, whereas the difficulty is due to the fact that interferers are often non-stationary and potentially similar to speech.The clean speech can be reconstructed by sparse coding of the learned speech dictionary. Partially coherent interferers can be reconstructed by sparse coding of the learned interferer dictionary. The clean speech is recovered from the degraded speech by sparse coding of the mixture in a composite dictionary consisting of the concatenation of a speech and interferer dictionary. But for the partially coherent interferers, in the process of sparse coding, there is the phenomenon that some interferer dictionary atoms coding the speech component, whereas the speech dictionary atoms coding the interferer component. For this problem, we have done the following works:1. Speech and interferer dictionary learningDictionary learning is very important in the speech denoising algorithm based on sparse coding. Usually in the speech denoising algorithm based on sparse coding, we assume that both the speech and partially coherent interferers can be sparse coded by over-complete dictionaries. But in this thesis, we assume that the partially coherent interferers can also be well approximated by a low-rank matrix. We use the K-SVD dictionary learning algorithm to learn the speech dictionary. For the interferer dictionary,based on the experiment, we have proved the rationality of the low-rank assumption,then we use the K-SVD dictionary learning algorithm and the algorithm of principal component analysis to get the interferer dictionary. In this way, we can reduce the number of atoms in the interferer dictionary, so we can also reduce the probability that the atoms of interferer dictionary coding the speech component.2. Sparse codingIn speech signal, between adjacent frames there exists strong correlation. In the process of atom selection, the traditional sparse coding algorithm like LARC only consider the correlation between the residual and atoms. In this thesis, we put forward the improved LARC based on the traditional LARC. In the improved LARC, we notonly consider the correlation between the residual and atoms, but also increase the probability of those atoms that have been used in the adjacent frames to be selected. In this way, we can reduce the probability that the interferer dictionary atoms code the speech component but the speech dictionary atoms code the interferer component.