Speech Enhancement using Adaptive Data-Based Dictionary Learning

In this paper, a speech enhancement method based on sparse representation of data frames has been presented. Speech enhancement is one of the most applicable areas in different signal processing fields. The objective of a speech enhancement system is improvement of either intelligibility or quality of the speech signals. This process is carried out using the speech signal processing techniques to attenuate the background noise without causing any distortion in the speech signal. In this paper, we focus on the single channel speech enhancement corrupted by the additive Gaussian noise. In recent years, there has been an increasing interest in employing sparse representation techniques for speech enhancement. Sparse representation technique makes it possible to show the major information about the speech signal based on a smaller dimension of the original spatial bases. The capability of a sparse decomposition method depends on the learned dictionary and matching between the dictionary atoms and the signal features. An over complete dictionary is yielded based on two main steps: dictionary learning process and sparse coding technique. In dictionary selection step, a pre-defined dictionary such as the Fourier basis, wavelet basis or discrete cosine basis is employed. Also, a redundant dictionary can be constructed after a learning process that is often based on the alternating optimization strategies. In sparse coding step, the dictionary is fixed and a sparse coefficient matrix with the low approximation error has been earned. The goal of this paper is to investigate the role of data-based dictionary learning technique in the speech enhancement process in the presence of white Gaussian noise. The dictionary learning method in this paper is based on the greedy adaptive algorithm as a data-based technique for dictionary learning. The dictionary atoms are learned using the proposed algorithm according to the data frames taken from the speech signals, so the atoms contain the structure of the input frames. The atoms in this approach are learned directly from the training data using the norm-based sparsity measure to earn more matching between the data frames and the dictionary atoms. The proposed sparsity measure in this paper is based on Gini parameter. We present a new sparsity index using Gini coefficients in the greedy adaptive dictionary learning algorithm. These coefficients are set to find the atoms with more sparsity in the comparison with the other sparsity indices defined based on the norm of speech frames. The proposed learning method iteratively extracts the speech frames with minimum sparsity index according to the mentioned measures and adds the extracted atoms to the dictionary matrix. Also, the range of the sparsity parameter is selected based on the initial silent frames of speech signal in order to make a desired dictionary. It means that a speech frame of input data matrix can add to the first columns of the over complete dictionary when it has not a similar structure with the noise frames. The data-based dictionary learning process makes the algorithm faster than the other dictionary learning methods for example K-singular value decomposition (K-SVD), method of optimal directions (MOD) and other optimization-based strategies. The sparsity of an input frame is measured using Gini-based index that includes smaller measured values for speech frames because of their sparse content. On the other hand, high values of this parameter can be yielded for a frame involved the Gaussian noise structure. The performance of the proposed method is evaluated using different measures such as improvement in signal-to-noise ratio (ISNR), the time-frequency representation of atoms and PESQ scores. The proposed approach results in a significant reduction of the background noise in comparison with other dictionary learning methods such as principal component analysis (PCA) and the norm-based learning method that are traditional procedures in this context. We have found good results about the reconstruction error in the signal approximations for the proposed speech enhancement method. Also, the proposed approach leads to the proper computation time that is a prominent factor in dictionary learning methods.