Speech Denoising Based on Sparse Representation Algorithm

A new speech denoising method that aims for processing corrupted speech signal which is based on K-SVD sparse representation algorithm is proposed in this paper. Here, the DCT sparse and redundant representation over dictionary is used for the initial redundant dictionary. In order to analyze the time-frequency characteristics of speech signal clearly, the spectrogram patches are applied as training samples for the sparse decomposition in this approach. However, the training samples need to extend their deployment to arbitrary spectrogram sizes because the K-SVD algorithm is limited in handling small size spectrogram. A global spectrogram was defined prior that forces sparsity over patches in every location in the spectrogram. Afterwards, by using the K-SVD algorithm, the greedy algorithm is used for updating which alternates between dictionary and sparse coefficients. Then a dictionary that describes the speech structure effectively can be obtained. Finally, the corrupted speech signal can be sparsely decomposed under the redundant dictionary. Consequently, the sparse coefficients can be obtained and used to reconstruct the noiseless spectrograms. As a result, the purpose of the separation for the signal and noise is reached. The proposed K-SVD algorithm is a simple and effective algorithm, which is suitable for processing corrupted speech signal. Simulation experiments show that the performance of the proposed K-SVD denoising algorithm is stable, and the white noise can be effectively separated. In addition, the algorithm performance surpasses the redundant DCT dictionary method and Gabor dictionary method. In a word, K-SVD algorithm leads to an alternative and novel denoising method for speech signals.