An iterative multichannel subspace-based covariance subtraction method for relative transfer function estimation

Multichannel speech enhancement systems are capable of extracting desired signals from noisy microphone signals. Optimal spatial filters used for source extraction often require knowledge of the relative transfer functions (RTFs) which describe the coupling among the microphone signals as a response to a signal originating from the location of the desired source. The performance of the speech enhancement system highly depends on the accuracy of the RTFs which are estimated from noisy microphone signals. While RTF estimation methods that apply generalized eigenvalue decomposition (GEVD) on the power spectral density (PSD) matrices of the noise and microphone signals achieve better performance than the covariance subtraction (CS) method, the computational complexity of the GEVD is a limiting factor in real-time applications. In this work, by exploiting the eigenstructure of the desired signal PSD matrix, an iterative RTF estimator with a low computational complexity is proposed, which offers a performance that is comparable to the iterative implementations of GEVD-based RTF estimators, and a better performance than the CS method.