Dual channel delay speech denoising based on underdetermined BSS

The mixtures of speech and Gaussian noise is studied in this paper. A delay speech denoising method, which considers two received speeches with time delay, is proposed. The algorithm is based on underdetermined BSS method and applies time frequency sparsity. The noisy speeches are firstly divided the two parts using a threshold and then one part less than the threshold can be regard as noise. Speech signal will be separated from another part using time frequency mask. The separated speech is finally also smoothed by a filter. We demonstrate it effectiveness in several simulations.