Immune K-SVD algorithm for dictionary learning in speech denoising

• DOI: 10.1016/j.neucom.2013.02.045
• Publication Date: 2013-08-10T23:02:11Z
• Authors: Yan Zhou; Heming Zhao; Li Shang; Tao Liu
• Keywords: K-SVD

Abstract A speech denoising method based on the sparse representation over a redundant dictionary is proposed in this paper. This redundant dictionary is trained by a corpus of corrupted spectrogram patches using immune K-singular value decomposition (IK-SVD) algorithm, which is an efficient sparse representation method for speech signals. Since the traditional K-SVD algorithm is an iterative method that alternates between sparse representation with respect to a given dictionary and a process of updating the dictionary atoms to better fit the data, it has the disadvantages of large calculation quantity, low approximation speed and accuracy. However, the immune algorithm has the characteristics of local fast convergence and global optimal, so it can optimize the K-SVD learning by introducing immune mechanism to modify the pursuit algorithm. This immune optimized K-SVD algorithm not only can accelerate the learning speed of K-SVD, but also increases its approximation accuracy. In test, the University of Victoria speech database is exploited to test the speech denoising approach proposed here. And compared with original K-SVD algorithm and Wavelet Transform (WT) denoising algorithm, experimental results show that the property of our algorithm behaves more stable and behaves better improvement signal-to-noise ratio (ISNR).