Abstract: Compressed Sensing（CS） which is a signal sparsity-based sampling method, can effectively extract the information contained in the signal. A new method is designed for noisy speech enhancement based on the grouping separation of compressed sensing. Speech sparse expression is used in discrete Fast Fourier Transform（FFT） domain. The algorithm can implement compression measurement and denoising in noisy speech by the design of the complex domain observation matrix and soft threshold. Sparse Reconstruction by Separable Approximation, SpaRSA algorithm is used to restore the speech signal, to achieve speech enhancement. The experiments show that the denoising signal can be compressed and reconstructed for noise signal compression refactoring. The signal-to-noise ratio can be improved greatly. The background noise can be  more effectively suppressed.

Key words: speech enhancement, compressed sensing, group-separable, soft threshold, denoising

摘要： 压缩感知（Compressive Sensing，CS）是一种基于信号稀疏性的采样方法，可以有效提取信号中所包含的信息。提出了一种分组分离压缩感知语音增强新算法。算法利用语音在离散快速傅里叶变换（Fast Fourier Transform，FFT）域下的稀疏性，设计复域观测矩阵与软阈值对带噪语音进行压缩测量与去噪，通过可分组分离逼近稀疏重建（Sparse Reconstruction by Separable Approximation，SpaRSA）算法恢复语音信号，实现语音增强。实验表明：该算法对含噪信号压缩重构，信噪比幅度较大提高，能更有效地抑制背景噪声。

关键词: 语音增强, 压缩感知, 分组分离, 软阈值, 去噪