General Robust Subband Adaptive Filtering: Algorithms and Applications

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

TL;DR

The authors propose a General Robust Subband Adaptive Filtering (GR-SAF) framework designed to handle "colored" (highly correlated) input signals in the presence of impulsive noise (sudden spikes). By minimizing the mean square deviation under a random-walk model, the algorithm effectively acts as a robust filter with a self-tuning variable regularization parameter, achieving faster convergence and lower steady-state error than existing subband adaptive filters.

What Problem Does This Paper Address?

Adaptive filters are the workhorses of echo cancellation and system identification. However, they face two major hurdles in real-world environments:

  1. Colored Input Signals: Standard algorithms like NLMS (Normalized Least Mean Square) converge very slowly when the input data is highly correlated (e.g., speech). Subband Adaptive Filtering (SAF) solves this by splitting the signal into frequency bands, but standard SAF is fragile.
  2. Impulsive Noise: Real-world signals often contain non-Gaussian "spikes" (impulsive noise) caused by things like double-talk in telephony or acoustic interference. These spikes can cause standard filters to diverge or "freeze" incorrectly.

Existing solutions often force a trade-off: you can have robustness against noise, or you can have fast convergence, but rarely both. This paper seeks a "general" framework that provides both by treating weight uncertainty mathematically.

Methods

The core of the GR-SAF approach is the derivation of a weight update rule based on a random-walk model with individual weight uncertainty.

Key technical components include:

  • Subband Decomposition: The input signal is passed through an analysis filter bank, decimated, and processed in parallel subbands. This decorrelates the signal, speeding up convergence for colored inputs.
  • Probability Update Mechanism: Instead of just using a hard threshold to ignore noisy samples, the authors define a parameter $P_{e,i}(k)$ that models the probability of a subband signal being "clean."
  • Flexible Scaling Factors: The framework is "general" because it allows researchers to plug in different robust criteria, such as the M-estimate (based on the modified Huber function) or the Maximum Correntropy Criterion (MCC).
  • Computational Efficiency: While the theoretical derivation involves a covariance matrix update (similar to a Kalman filter), the authors provide a diagonal approximation that keeps the computational complexity at $O(M)$, making it practical for high-order filters used in echo cancellation.

Datasets

The paper evaluates the algorithm using standard signal processing simulation scenarios rather than static machine learning datasets:

  • System Identification: Using first-order autoregressive (AR) signals to simulate highly correlated input.
  • Network Echo Cancellation: Utilizing impulse responses from the ITU-T G.168 standard (specifically channel-1 and channel-4) to represent dispersive and sparse network environments.
  • Real Speech Signals: Evaluation using actual far-end and near-end speech recordings to test performance in "double-talk" scenarios.
  • Noise Models: Performance was tested against contaminated-Gaussian noise and $\alpha$-stable noise to simulate various levels of impulsiveness.

Key Insights and Ideas

  • Variable Regularization as an Emergent Property: One of the most interesting findings is that the GR-SAF algorithm naturally reduces to a robust NSAF with a variable regularization parameter. It automatically adjusts its "step size" based on the estimated variance of the subband noise and the uncertainty of the filter weights.
  • Individual Weight Uncertainty: By assuming each filter coefficient has its own level of uncertainty (the random-walk model), the filter becomes much more agile at tracking changes in the system (like a sudden change in an echo path) compared to algorithms that assume a global uncertainty.
  • Robustness via Scaling: The use of scaling factors $q(e)$ allows the filter to effectively "ignore" impulsive spikes without the stability issues that plague sign-based subband filters.

Key Results

  • Convergence Speed: In system identification tasks with impulsive noise, GR-SAF reached a steady-state Mean Square Deviation (MSD) of -40 dB significantly faster than the standard M-NSAF and VSS-M-NSAF algorithms.
  • Steady-State Accuracy: The proposed method consistently achieved a lower MSD (often 5-10 dB lower) than its counterparts in both dispersive and sparse channel environments.
  • Echo Cancellation: In double-talk simulations (where both parties speak at once), the GR-SAF maintained a high Echo Return Loss Enhancement (ERLE), whereas non-robust algorithms saw their performance collapse.
  • Robustness to Noise Type: The algorithm showed superior stability in $\alpha$-stable noise environments, where the "spikiness" of the noise is extreme.

Limitations and Future Work

  • Parameter Tuning: While the algorithm is robust, it still requires the selection of certain hyperparameters (like the threshold $\xi$ for M-estimates or the kernel width $\kappa_\sigma$ for MCC). The authors suggest typical values, but optimal performance may require manual tuning for specific applications.
  • Filter Bank Delay: Like all subband adaptive filters, there is an inherent delay introduced by the analysis/synthesis filter banks. While the authors mention "delayless" variants, the primary focus here is on the update logic rather than minimizing group delay.
  • Future Direction: The authors suggest that this framework could be extended to distributed networks or non-linear adaptive filtering scenarios.

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 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 outperforms its counterparts.
BibTeX
@article{yu2022general,
  author = {Yi Yu and Hongsen He and Rodrigo C. de Lamare and Badong Chen},
  title = {{General Robust Subband Adaptive Filtering: Algorithms and Applications}},
  journal = {IEEE ACM Trans. Audio Speech Lang. Process.},
  year = {2022},
  url = {https://openreview.net/forum?id=rK7KUmfc1f},
}

Loading related content...

Figure