Robust dual-channel noise power spectral density estimation

In this paper, we propose a novel noise power spectral density (PSD) estimator which is beneficial for speech enhancement systems with two microphones in diffuse noise environments. The algorithm has a low computational complexity and requires low memory usage. The main advantage is that arbitrary models of the noise field coherence can be employed and a scalable extension of existing single-channel speech enhancement systems to dual channels is also possible. Experiments demonstrate with simulated and measured data that the proposed algorithm outperforms related algorithms in diffuse noise conditions.