Study of General Robust Subband Adaptive Filtering

Yi Yu, Hongsen He, Rodrigo C. de Lamare, Badong Chen

Abstract

In this paper, we propose a general robust subband adaptive filtering (GR-SAF) scheme against impulsive noise by minimizing the mean square deviation under the random-walk model with individual weight uncertainty. Specifically, by choosing different scaling factors such as from the M-estimate and maximum correntropy robust criteria in the GR-SAF scheme, we can easily obtain different GR-SAF algorithms. Importantly, the proposed GR-SAF algorithm can be reduced to a variable regularization robust normalized SAF algorithm, thus having fast convergence rate and low steady-state error. Simulations in the contexts of system identification with impulsive noise and echo cancellation with double-talk have verified that the proposed GR-SAF algorithms outperform its counterparts.

Introduction

Adaptive filters are essential components in modern communication systems, particularly for echo cancellation and system identification. However, their performance is often degraded by two primary factors: colored (highly correlated) input signals and impulsive noise. While standard Subband Adaptive Filtering (SAF) addresses the issue of colored signals by decorrelating the input through frequency band splitting, it remains vulnerable to non-Gaussian impulsive spikes, such as double-talk in telephony. This paper introduces the General Robust Subband Adaptive Filtering (GR-SAF) framework, which achieves both robustness and fast convergence by treating filter weight uncertainty mathematically within a random-walk model.

Proposed Methodology: The GR-SAF Scheme

The core innovation of the GR-SAF approach is the derivation of a weight update rule based on a random-walk model that accounts for individual weight uncertainty. This allows the filter to be more agile in tracking system changes compared to global uncertainty models.

Block diagram of the proposed general robust subband adaptive filtering (GR-SAF) scheme.

As shown in the block diagram above, the input signal $u(n)$ is processed through an analysis filter bank, decimated, and handled in parallel subbands. The framework is "general" because it allows for the integration of various robust criteria, such as the M-estimate (modified Huber function) or the Maximum Correntropy Criterion (MCC), via flexible scaling factors. A key emergent property of this derivation is that it naturally reduces to a robust Normalized SAF (NSAF) with a variable regularization parameter, which automatically adjusts the step size based on subband noise variance and weight uncertainty.

Experimental Evaluation and Results

The performance of the GR-SAF algorithm was evaluated across several simulated scenarios, including system identification with autoregressive signals and network echo cancellation using ITU-T G.168 standard impulse responses for dispersive and sparse channels. Performance was tested against contaminated-Gaussian and $\alpha$-stable noise models to simulate extreme impulsiveness.

MSD performance comparison of GR-SAF against existing algorithms in dispersive and sparse channel environments.

Key results from the simulations include:

  • Convergence and Accuracy: In system identification tasks, GR-SAF reached a steady-state Mean Square Deviation (MSD) of -40 dB significantly faster than standard M-NSAF and VSS-M-NSAF algorithms. It consistently achieved a steady-state MSD 5-10 dB lower than its counterparts.
  • Robustness to Channel Type: The algorithm maintained superior performance in both dispersive and sparse channel environments, as illustrated in the MSD curves above.
  • Echo Cancellation Performance: In double-talk scenarios, GR-SAF maintained high Echo Return Loss Enhancement (ERLE), whereas non-robust algorithms typically fail under such conditions.

Conclusion

The proposed GR-SAF framework provides a robust and efficient solution for adaptive filtering in the presence of impulsive noise and correlated inputs. By incorporating individual weight uncertainty and flexible scaling factors, it achieves a superior balance between convergence speed and steady-state accuracy. Future work may involve extending this framework to distributed networks or non-linear adaptive filtering scenarios. While highly effective, optimal performance still requires careful tuning of hyperparameters like the M-estimate threshold or MCC kernel width for specific applications.

Summary updated

Abstract

In this paper, we propose a general robust subband adaptive filtering (GR-SAF) scheme against impulsive noise by minimizing the mean square deviation under the random-walk model with individual weight uncertainty. Specifically, by choosing different scaling factors such as from the M-estimate and maximum correntropy robust criteria in the GR-SAF scheme, we can easily obtain different GR-SAF algorithms. Importantly, the proposed GR-SAF algorithm can be reduced to a variable regularization robust normalized SAF algorithm, thus having fast convergence rate and low steady-state error. Simulations in the contexts of system identification with impulsive noise and echo cancellation with double-talk have verified that the proposed GR-SAF algorithms outperforms its counterparts.
BibTeX
@article{yu2022study,
  author = {Yi Yu and Hongsen He and Rodrigo C. de Lamare and Badong Chen},
  title = {{Study of General Robust Subband Adaptive Filtering}},
  journal = {IEEE ACM Trans. Audio Speech Lang. Process.},
  year = {2022},
  url = {https://openreview.net/forum?id=rK7KUmfc1f},
}

Loading related content...

Figure