InterSpeech 2021

A Partitioned-Block Frequency-Domain Adaptive Kalman Filter for Stereophonic Acoustic Echo Cancellation
(Oral presentation)

Rui Zhu (Tencent, China), Feiran Yang (CAS, China), Yuepeng Li (Tencent, China), Shidong Shang (Tencent, China)
The rapid development of online video conferencing systems has caused renewed attention to the multi-channel recording and playback systems. Stereophonic acoustic echo cancellation (SAEC) is the key issue of this systems. This paper proposes an optimally designed partitioned-block frequency-domain Kalman filter (PBFDKF) algorithm for SAEC. We establish the frequency-domain observation equation using the overlap-and-save method and we use the first-order Markov model to describe the state equation. The exact PBFDKF algorithm is derived under the umbrella of Kalman filter theory and two fast implementations are then presented to reduce the complexity. The proposed algorithm is equivalent to the dual-channel partitioned-block frequency-domain gradient-based algorithm with optimum step-size control, and hence it exhibits very good convergence performance and is found to be robust to near-end interference without a double-talk detector. Extensive experiments in different SAEC conditions confirm the effectiveness of the proposed algorithm.