Termin – Detailansicht

Master-Vortrag: Combined Beamforming and Acoustic Echo Cancellation for Speech Enhancement

Alexander Bohlender
Mittwoch, 17. Oktober 2018
10:00 Uhr
Hörsaal 4G

In hands-free communication, strong acoustic echoes can occur that can efficiently be suppressed by an acoustic echo cancellation (AEC). However, near-end interference can deteriorate the signal quality as well. When more than one microphone is available, beamforming (BF) techniques can be employed to attenuate all undesired signal components. Although both approaches are designed for an optimization of the speech intelligibility in the transmitted signal, they are still fundamentally different. The AEC only exploits the knowledge of the far-end signal and the BF only the spatial information. Due to their common goal, it is desirable to combine AEC and BF in order to fully exhaust the available information.

A simple approach to do so is the cascade scheme beamformer first (BF-first) in which the independently adapted AEC follows the BF. Despite its low complexity, BF-first is often discarded as the AEC is unable to track frequent or abrupt major changes of the BF. In this thesis, an approach is proposed that aims at mitigating this fundamental deficiency of BF-first. This way, the simple structure and low complexity of the system can be preserved.

The newly introduced change prediction (ChaP) algorithm is initially derived based on the assumption that the room impulse responses do not change within a specified duration of time. Throughout the thesis, ChaP is progressively improved for being able to cope with realistic conditions. These comprise, e.g., the finite AEC filter system distance and the large similarity of the BF weights for specific directions and frequencies. Most importantly, the strict restriction to time-variant room impulse responses is removed. With all extensions of ChaP introduced, it is discussed how ChaP can be used in other combined AEC and BF systems such as the generalized sidelobe canceler with acoustic echo cancellation in the reference path (GSAEC).