BI-CosampSE: Block identification based compressive sampling matching pursuit for speech enhancement

The conventional compressive sampling matching pursuit (CoSaMP) algorithm often produces isolated components (IFC) in the frequency domain, when it is used to recover sparse or compressible signals from their downsampled data. The IFCs in the frequency domain are perceived as musical noise, when CoSaMP is used for speech enhancement (SE). In this paper, we propose a novel method to tackle this problem by using a block based identification strategy (BIS) to seek the most prominent components in the observed data to update the sparse estimate of CoSaMP. The proposed method has been found to be very effective to reduce musical noise in speech enhancement, in combination with some time-frequency smoothing techniques (TFSTs). The perceptual quality of the enhanced signals is significantly improved.