Development and Real-Time Implementation of an Adjustable Coherence Mask for Robust Binaural Speech Enhancement in Noisy and Multi-Talker Environments Using Hearable Devices

Abstract

Hearable technology has rapidly evolved, offering improved communication and listening experiences, especially in noisy environments. However, achieving acoustic transparency and preserving binaural cues such as Interaural Time Difference (ITD) and Interaural Level Difference (ILD)—which are essential for spatial hearing—remain challenging. These difficulties are heightened in real-world, multi-talker settings, where existing noise reduction techniques often compromise spatial fidelity and speech intelligibility. This thesis addresses these challenges by developing novel coherence-based postfiltering algorithms aimed at enhancing noise reduction and speech intelligibility while preserving binaural spatial cues. It includes a comprehensive analysis of binaural coherence vectors in the short-time-frequency domain, leading to a novel geometrical interpretation of the coherence disc in a binaural context. The limitations of existing Coherent-to-Diffuse Ratio (CDR) post-filters, particularly their sensitivity to phase and magnitude fluctuations, are examined. To address these shortcomings, a new parameterized binaural CDR post-filter optimized for narrow frequency bands is proposed. The real-time implementation of these algorithms is demonstrated in a low-latency digital signal processing (DSP) system using a generic hearable device. Objective evaluations in various environments, coupled with subjective assessments using MUSHRA, Matrix Sentence, and localization tests, reveal significant improvements in both speech intelligibility and spatial perception preservation. This research advances hearable technology by providing robust, real-time solutions for enhancing listening experiences in noisy, reverberant, and multi-talker environments, with a particular focus on preserving critical binaural cues.