Neural Momentum-Enhanced LMS for Linear Acoustic Echo Cancellation

Acoustic Echo Cancellation (AEC) has been a longstanding research focus in signal processing for voice communications, audio systems and intelligent speech devices. Typically, acoustic echo can be cancelled by subtracting the estimated echo, typically calculated via adaptive filtering algorithms, from the near-end microphone signal. However, in practical applications, adaptive filter is prone to under compensation or over-compensation, particularly in double-talk at high power level. This requires the use of advanced algorithms capable of dynamically adjusting filter parameters to achieve accurate and stable echo cancellation. Hybrid approaches that integrate deep neural networks (DNN) with traditional adaptive filtering algorithms have received increasing attention. Although DNN helps to update the parameters of the adaptive filter, it also results in slower convergence and inaccurate gradient estimation, especially during echo-path changes. This paper proposes a neural momentum LMS algorithm, which employs narrowband complex DNN to capture the dynamic correlations between momentum and gradient terms, leading to more efficient parameter optimization and faster convergence. Extensive experiments were conducted in stationary and non-stationary channels, echo-path change, and real-world recordings. The results demonstrate that the proposed algorithm achieves superior convergence and re-convergence performance across diverse and acoustically challenging environments. Moreover, with only 4.9K model parameters, the proposed algorithm exhibits great potential for deployment in resource-constrained devices.