Early and Late Reflections in Acoustic Echo Control: An Experimental Study on (Neural) Kalman Filters and DNN Methods

Inhalt:
In recent years, the field of acoustic echo control (AEC) has seen a variety of approaches. In many cases, these systems rely on a postfilter to remove residual echo alongside background noise, augmenting possibly limited modeling capabilities for late echo reflections in larger rooms. Due to this, an AEC which can effectively remove early reflections, and does so consistently regardless of overall reverberation time or dynamics of late reflections, is highly desirable. We investigate several models—with a focus on neural Kalman filters—regarding their capabilities to remove the critical early reflections. We show that a focus on early reflections in training can lead to a more balanced trade-off between echo suppression and near-end speech preservation. Overall, the classical Kalman filter and its closely related hybrid models NKF and DLAC-Kalman prove effective in removing early reflections without much near-end speech degradation, providing a good starting point for subsequent postfiltering.