Introduction
Acoustic Echo Cancellation (AEC) is a critical technology for modern communication, preventing users from hearing their own voices echoed back during calls. However, AEC faces two major hurdles: double-talk, where both parties speak simultaneously, and nonlinearity, caused by the physical limitations of small speakers in mobile devices. While Semi-Blind Source Separation (SBSS) combined with nonlinear expansion models can address these issues, existing methods like Iterative Projection (IP) are computationally expensive due to frequent large-scale matrix inversions, making them unsuitable for low-power embedded chips.
This paper introduces Element-wise Iterative Source Steering (EISS), a highly efficient algorithm that reduces the complexity of SBSS by an order of magnitude while maintaining state-of-the-art performance.
Methodology
The proposed framework utilizes a $P$-th order power series to model loudspeaker distortion and a Convolutive Transfer Function (CTF) to model the echo path in the frequency domain. The core innovation, the EISS algorithm, replaces the global filter solution—which requires $O((PL)^3)$ complexity—with a sequential, element-wise update strategy. By leveraging the Majorization-Minimization (MM) framework, EISS "steers" the source estimate iteratively, bypassing matrix inversions and reducing complexity to $O((PL)^2)$.
Experimental Results
The researchers evaluated EISS using synthetic data and the ICASSP 2022 AEC Challenge dataset. The results demonstrate that EISS achieves significant computational savings without sacrificing quality.
Computational Efficiency
Runtime tests show that as the filter length ($L$) increases, the time saved by EISS grows dramatically. At $L=12$, EISS is approximately three times faster than the IP-based method, as illustrated in Figure 5.
Performance Parity
Despite its reduced complexity, EISS matches the performance of the IP method across all key metrics, including voice quality (PESQ), intelligibility (STOI), and echo reduction (tERLE). Both SBSS-based methods significantly outperform traditional state-space models (SSM), particularly in double-talk scenarios.
Conclusion
The EISS algorithm provides a computationally efficient solution for nonlinear echo cancellation, making advanced SBSS techniques feasible for real-time deployment on resource-constrained hardware. Future work will focus on extending this approach to multi-channel setups and exploring more complex distortion models.
Summary updated