Statistical analysis and improvement of coherent-to-diffuse power ratio estimators for dereverberation

Dereverberation methods based on coherent-to-diffuse power ratio (CDR) estimates exploit the spatial properties of signals to suppress late reverberation components. Naturally, the quality of the CDR estimate has a major effect on the quality of the dereverberated signals. This paper presents a statistical study of the performance of state-of-the-art CDR estimators under a Gaussian signal model. The study reveals that the variance of the magnitude of the spatial coherence estimate is the major cause for the degradation of known CDR estimators, while the variance of the phase has only limited impact. Moreover, the variance has more impact on direction-of-arrival (DOA)-independent CDR estimators than on DOA-dependent estimators. Accordingly, we propose a cepstrum thresholding technique to reduce the variance of the spatial coherence estimate. Experimental results show that better performance can be achieved for dereverberation when the proposed spatial coherence method is applied to CDR estimators, especially for the DOA-independent estimators.