Noise PSD Estimation Using Blind Source Separation in a Diffuse Noise Field

This paper considers the problem of two-channel speech enhancement with a single target speaker in a diffuse noise field and proposes a novel algorithm for estimating the power spectral density (PSD) of the background noise. First, a reference noise signal is computed as the output of a blocking matrix based on blind source separation (BSS). Then the ratio between the target noise PSD and the reference noise PSD is calculated by exploiting the separation matrix obtained from BSS and the spatial coherence of the noise field which is assumed to be known. The proposed algorithm can also operate during speech active periods, which is advantageous especially in fast-changing non-stationary noise fields. Experimental results show that the proposed algorithm outperforms existing methods, both in terms of noise PSD estimation and noise suppression.