On Multichannel Coherent-to-Diffuse Power Ratio Estimation

The significance of the coherent-to-diffuse-power ratio (CDR) has grown in the fields of speech dereverberation and noise reduction. However, the existing CDR estimators are typically limited to applications with only two microphones. In this article, we investigate CDR estimation in multichannel acoustic systems with more than two microphones. We propose two estimation methods. The first approach involves decomposing the microphone array into several groups of subarrays, where each subarray consists of only two sensors. We estimate the CDR for each group and then fuse these group CDR estimates through weighted averaging to form the multichannel CDR estimate. This weighted-average CDR estimation can be seen as an extension of traditional two-channel CDR estimation methods to the multichannel scenario. The second method is based on array manifold estimation using a joint matrix diagonalization technique, eliminating the need for subarray decomposition. By integrating the CDR estimates with a parametric Wiener-type postfilter, we demonstrate, via simulations, the superior performance of the proposed techniques in terms of CDR estimation accuracy, signal-to-noise ratio (SNR) gain, log-spectral distortion (LSD), and direct-to-reverberant-energy ratio (DRR).