Semi-Blind Source Separation for Nonlinear Acoustic Echo Cancellation

Guoliang Cheng, Lele Liao, Hongsheng Chen, Jing LuIEEE Signal Processing Letters 2021 · DOI

TL;DR

This paper introduces a novel Semi-Blind Source Separation (SBSS) approach for Nonlinear Acoustic Echo Cancellation (NAEC). By cleverly incorporating nonlinear expansion coefficients into the echo path and treating basis functions as reference signals, their method offers improved robustness against mismatches between actual and modeled nonlinearities, outperforming existing NAEC techniques in challenging scenarios.

What Problem Does This Paper Address?

Acoustic Echo Cancellation (AEC) is crucial for clear communication in hands-free devices. While linear AEC is well-established, real-world devices like smartphones and laptops introduce nonlinearities in their loudspeakers. Nonlinear Acoustic Echo Cancellation (NAEC) aims to address this, but existing adaptive filtering methods for NAEC are highly sensitive to inaccuracies in the nonlinear models used. This mismatch between the assumed and actual nonlinear behavior can significantly degrade performance. The paper tackles this challenge by adapting Semi-Blind Source Separation (SBSS) techniques, which are known for their robustness to model mismatches in linear scenarios, to the nonlinear domain.

Methods

The core of the proposed method is an adaptation of SBSS for NAEC. The authors model the nonlinear echo path as a cascade of a memoryless nonlinear function and a linear echo path. They then represent the nonlinear function using a basis-generic expansion, effectively treating each basis function of the far-end input signal as a known "equivalent reference signal." This allows them to merge the nonlinear expansion coefficients into the echo path.

The adaptation of SBSS is achieved by:

  • Basis-Generic Expansion: Representing the nonlinear mapping $f(x(t))$ using a sum of basis functions $\phi_i(x(t))$ with coefficients $a_i$.
  • Merging Coefficients: Incorporating these coefficients $a_i$ into the echo path $h(t)$, creating a new effective echo path $h'(t) = \sum a_i h(t) * \phi_i(x(t))$.
  • SBSS Framework: Applying the SBSS framework, which relies on the independence between the desired near-end signal and the reference signals. The update algorithm follows a constrained scaled natural gradient strategy.

This approach is designed to be less sensitive to the mismatch between the numerical (modeled) and actual nonlinearities because the SBSS algorithm's performance is primarily based on the independence assumption, not the precise modeling of the nonlinearity itself.

Datasets

The paper does not introduce new datasets. Instead, it uses:

  • Simulated Data: Generated using two types of nonlinear mappings: hard clipping and soft saturation, with varying Signal-to-Distortion Ratios (SDR) and Echo-to-Near-End-Signal Power Ratios (ESR). The echo path was a room impulse response.
  • Real Captured Data: A male speech signal emitted by a miniature loudspeaker (inherently nonlinear) was recorded by a microphone.

Key Insights and Ideas

  • SBSS for Nonlinearity: The paper successfully extends the robustness of SBSS from linear to nonlinear echo cancellation, a significant step for practical NAEC.
  • Model Mismatch Resilience: The proposed SBSS method demonstrates superior performance compared to existing NAEC algorithms when there's a mismatch between the modeled and actual nonlinearities. This is a critical advantage for real-world applications.
  • Basis Functions as References: Treating basis functions of the nonlinear mapping as "equivalent reference signals" is an elegant way to integrate nonlinear modeling into the SBSS framework.
  • Constrained Natural Gradient: The use of a constrained scaled natural gradient strategy helps maintain the structure of the demixing matrix, leading to a stable and effective online update algorithm.

Key Results

  • Single-Talk Case (Simulated): In unmatched conditions (where the modeled nonlinearity differs from the actual one), the proposed SBSS algorithm significantly outperformed the state-of-the-art SD-MCSSFDAF and DualStage-MCK algorithms in terms of Echo Return Loss Enhancement (ERLE).
  • Double-Talk Case (Simulated): The SBSS algorithm achieved the best true ERLE (tERLE) and superior near-end speech quality (evaluated by PESQ and STOI) across different Signal-to-Noise Ratios (SNR) in unmatched conditions.
  • Real Experiments (Single-Talk): The proposed SBSS algorithm showed better ERLE performance than other algorithms on real captured data, confirming its effectiveness in practical scenarios with unknown nonlinearities.
  • Real Experiments (Double-Talk): The SBSS algorithm again demonstrated the best echo cancellation performance and near-end speech quality in a double-talk scenario with real data.

Limitations and Future Work

  • Nonlinear Expansion Order: The performance of the proposed method relies on the chosen order of the nonlinear expansion. While $p=3$ was used, higher orders might be necessary for more complex nonlinearities, potentially increasing computational complexity.
  • Computational Complexity: While the paper presents an online algorithm, the computational cost of the nonlinear expansion and the SBSS update might still be a consideration for very low-power devices.
  • Future Work: The authors suggest that further research could explore more complex nonlinear models or investigate adaptive selection of the nonlinear expansion order. The paper also implicitly points towards the potential for applying this SBSS approach to other areas involving nonlinear system identification.

Summary updated

BibTeX
@article{cheng2021semi,
  author = {Guoliang Cheng and Lele Liao and Hongsheng Chen and Jing Lu},
  title = {{Semi-Blind Source Separation for Nonlinear Acoustic Echo Cancellation}},
  journal = {IEEE Signal Processing Letters},
  year = {2021},
  doi = {10.1109/lsp.2021.3060325},
}
Figure