Robust Relative Transfer Function Identification on Manifolds for Speech Enhancement

Accurate and reliable identification of the relative transfer function (RTF) between microphones with respect to a desired source is an essential component in the design of microphone array beamformers. In this paper, we present a robust RTF identification method on manifolds, tested and trained with real recordings. This method relies on a manifold learning (ML) approach to infer a representation of typical RTFs in a confined area within an acoustic enclosure. We propose a robust supervised identification method that combines the a priori learned geometric structure and the measured signals. A series of experiments using a recently established database of acoustic responses taken at the Bar-Ilan university acoustic lab, demonstrate the effectiveness of the proposed approach over a standard, non-robust, beamforming design method.