Multimicrophone Signal Parameter Estimation in a Multi-Source Noisy Reverberant Scenario

Estimation of acoustic parameters is of great interest but very challenging in the multichannel microphone signal processing area. Existing methods either assume simple, but less realistic scenarios, or suffer from very high computational costs. In this work, we consider the more general scenario where multiple sources, late reverberation and noise exist concurrently. The parameters of interest include the relative transfer functions (RTFs) of the point sources (both target and interferers) and individual power spectral densities (PSDs) of the sources and the late reverberation. We first propose a robust late reverberation PSD estimator using an iterative compensation scheme. Then, based on an analysis of the variance of the sample covariance matrices, we propose a robust and joint estimator for the sources RTFs and PSDs using multiple time frames that share the same RTFs. We compare the proposed method with the state-of-the-art simultaneously confirmatory factor analysis (SCFA) method and the second order blind identification (SOBI) method. Experiments show that our proposed method reaches the estimation performance of SCFA, which significantly outperforms SOBI, but uses much less computational costs compared to SCFA.