Exploration of alternative algorithms for multi-channel acoustic echo cancellation

Abstract

Multi-Channel Acoustic Echo cancellation (MCAEC) is a vital component of delivering clean speech to a virtual personal assistant through a smart speaker with multi-channel audio (stereophonic, etc). The use of the Kalman filter as an alternative adaptive filter methodology for this MCAEC application is explored in this work. The Normalized Least Mean Squares filter (NLMS) serves as a benchmark for the Kalman filter. Simulations using room recordings and measured room responses are employed in this exploration. Useful metrics such as the Word Error Rate (WER) and Echo Return Loss Enhancement (ERLE) help to distinguish performance among the two adaptive filter algorithms. For the single channel case, simulations confirm the cancellation and convergence rate advantage of the Kalman filter, in full-band, but the NLMS filter gives similar results in the sub-band domain, as measured by WER and ERLE. In the multi-channel case, both solutions achieve similar steady state cancellation, but the NLMS offers slightly faster convergence rates. In experiments where adaptation was not frozen, the Kalman filter effectively maintains high echo cancellation by tracking input signal statistics. In most cases, the Kalman filter does not present an appropriate alternative for the MCAEC application in this work.