Adjustable Coherent-to-Diffuse Power Estimator for Binaural Speech Enhancement in Multi-Talker Environments

The binaural coherence-to-diffuse power ratio (CDR) estimate in reverberant environments is essential in many speech enhancement algorithms applied within hear-through systems. In this work, we propose a parameterised and adjustable binaural CDR estimator whose formulation is based on a geometrical interpretation of the short-time complex coherence function between binaural microphone signals. Conventional CDR estimators often distort the natural spectro-temporal behaviour of the noise field by relying on theoretical coherence models of the desired signal and/or diffuse noise field. Our proposed CDR estimator relies only on the observed spatial coherence and better preserves the natural characteristics of a binaural noise field. We demonstrate that the proposed CDR estimator can be used effectively for binaural dereverberation and denoising of broadside speech in multi-talker and noisy acoustic conditions and that it often outperforms state-of-the-art coherence-based methods for dereverberation and denoising. Furthermore, the adjustable parameter enables one to minimise the frequency-dependent estimation error of the binaural system in different environments.