关键词: 语音增强, K奇异值分解, 稀疏表示, 信号子空间

Abstract:

In the case of a low SNR, it is difficult that the clean speech is separated completely by sparse representation from the noisy speech. To solve the above problem, a speech enhancement method using the improved K-SVD algorithm by subspace is proposed. First, the noise is tracked by the optimal estimator of the subspace, and a noise dictionary is trained by using the K-SVD. Then, the speech dictionary is trained by the K-SVD algorithm. In the process of training, if an atom whose sparse coefficient is lower than the set threshold and could also be found in the noise dictionary, the sparse coefficient is set to zero, which achieves the goal of de-noising. Simulation results show that the algorithm can remove white noise and babble noise obviously, so that the SNR is improved and distortion is reduced greatly. Simultaneously, this improved algorithm can also be applied to eliminate the random noise very well. And the improved algorithm verified by SPSS19.0 software is superior to the K-SVD algorithm and subspace algorithm under a low SNR.

Key words: speech enhancement, K-SVD, sparse representation, subspace