RELATIVE TRANSFER MATRIX-BASED BINAURAL SIGNAL DENOISING OF HEAD-MOUNTED MICROPHONE ARRAY RECORDINGS

Abstract

Head-mounted microphone arrays are increasingly prevalent in applications ranging from virtual reality to assistive hearing devices. Accurately enhancing binaural signals from these devices is crucial yet challenging in complex acoustic environments characterized by multiple sound sources and significant reverberation. The Relative Transfer Matrix (ReTM) approach, which generalizes relative transfer functions for multiple simultaneously active sources and receivers, has demonstrated success in speech denoising. This paper addresses the problem of binaural signal denoising by utilizing ReTM derived from headmounted microphone array recordings. Our key contribution is adapting the ReTM computation to accommodate the user’s head movements based on head-tracking data, which enhances the fidelity of the denoising process. We demonstrate this application with an augmented reality (AR) glass setup, equipped with four microphones on the frame and two over-ear microphones. The noiseonly ReTM, computed between the on-frame and overear microphones across various head orientations, is employed to estimate and subsequently subtract noise from the binaural signal. The simulation results indicate that a higher resolution of ReTM-Dictionary leads to better speech quality (STOI, PESQ, SegSNR) scores, with improved preservation of binaural cues (ITD and ILD).