Neighbor-Consistent Neural Filters for Robust Personal Sound Zones Under Localization Uncertainty

Abstract: Coordinate-conditioned neural networks can generate head-tracked personal sound zone (PSZ) loudspeaker filters in real time, but they are sensitive to localization uncertainty. Small fluctuations in estimated listener coordinates, caused by optical distortion, temporary occlusions, or tracking jitter, may produce large filter changes even when listeners are physically stationary. This paper proposes neighbor-consistent neural filters that regularize the coordinate-to-filter mapping by penalizing filter differences at randomly perturbed neighboring coordinates during training. To evaluate robustness against tracking noise, we introduce a decoupled protocol that fixes the acoustic transfer functions at a physical anchor while perturbing only the coordinate inputs used for filter generation. Isolation quality and local stability are evaluated using neighborhood median and lower-tail statistics of inter-zone and inter-program isolation, together with spatial variation rates that quantify metric sensitivity within a coordinate neighborhood. In simulation with a split-band woofer-tweeter system and 25 randomly sampled anchor positions, neighbor consistency reduces the root-mean-square (RMS) variation rate by up to 55.9% in the woofer band and 30.3% in the tweeter band while largely preserving isolation quality and improving lower-tail robustness. In in-situ measurements using a 24-driver array and two stationary head-and-torso simulators, the proposed regularization improves worst-case neighborhood isolation by up to 16.9% and reduces spatial variation rates by up to 61.8%. These results demonstrate that neighbor-consistency regularization effectively stabilizes PSZ rendering under localization uncertainty.

• The paper presents neighbor-consistency regularization as its main innovation, reducing sensitivity to coordinate perturbations in neural filter outputs.
• It employs a split-band neural architecture with a decoupled evaluation protocol, ensuring stable audio rendering across different frequency bands.
• Empirical results from simulations and real-room tests reveal significant improvements in isolation quality, with up to 55.9% reduction in spatial variation rates.

Neighbor-Consistent Neural Filters for Robust Personal Sound Zones Under Localization Uncertainty

Introduction

Personal sound zones (PSZ) leverage loudspeaker arrays to deliver independent audio streams to multiple listeners in shared acoustic environments without headphones. State-of-the-art PSZ systems often employ coordinate-conditioned neural networks to rapidly generate loudspeaker filters for head-tracked audio rendering. However, such coordinate-to-filter mappings are typically sensitive to localization noise—small tracking errors or estimation jitter can induce unstable filter outputs, resulting in perceptually objectionable fluctuations in the reproduced sound field even when listeners remain static. The paper "Neighbor-Consistent Neural Filters for Robust Personal Sound Zones Under Localization Uncertainty" (2605.21891) investigates this critical robustness issue and proposes a regularization technique that directly addresses the instability resulting from coordinate uncertainty.

Problem Statement and Motivation

Conventional PSZ evaluation metrics, such as broadband inter-zone isolation (IZI) and inter-program isolation (IPI), do not reflect robustness to small coordinate perturbations in input to the neural filter generator. Practical head-tracking and localization systems often exhibit jitter or systematic errors due to sensor limitations and environmental occlusions, making coordinate-input stability a pragmatic concern. Previous solutions focused primarily on combating acoustic model mismatch and ATF errors, but did not explicitly address the inherent sensitivity of coordinate-conditioned neural mappings.

Methodology

System Model

The proposed system uses a split-band neural architecture: independent neural filter generators, each implemented as a Fourier-encoded MLP, are trained separately for low-frequency (woofer) and high-frequency (tweeter) bands. For listener $k$, the stacked coordinate vector $x$ (typically $d=2$ for horizontal coordinates) serves as input to both generators. The generators map $x$ to filter coefficient vectors for all loudspeaker channels and listener audio programs.

Baseline Objective

Each band-specific model is trained using a multi-term loss, composed of bright-zone reproduction (magnitude fidelity), dark-zone suppression (leakage minimization), and additional regularization on filter compactness and frequency-domain norm. The standard approach provides no explicit constraint on the spatial smoothness of the coordinate-to-filter mapping.

Neighbor-Consistency Regularization

To induce robustness, the core innovation is a neighbor-consistency loss term added during training. For each sampled coordinate input $x$, a random perturbation $\delta \sim U([- \Delta, \Delta])$ is applied, yielding $x' = clip(x+\delta)$. The mean-squared deviation $\| f(x) - f(x') \|^2$ between the neural filters generated for $x$ and $x'$ is penalized, with masking to ensure both coordinates remain in the same spatial regime (overlap vs non-overlap in multi-listener settings). The relative strength of this regularization and the perturbation scale $x$0 control the trade-off between spatial robustness and nominal filter accuracy.

Decoupled Evaluation Protocol

To isolate the impact of coordinate-input perturbations from physical displacement, a novel decoupled evaluation framework is introduced. Acoustic transfer functions (ATFs) are fixed at the true anchor location but the neural filter generator is driven by perturbed estimates of the listener's coordinates. This protocol ensures that all observed variability in PSZ metrics arises solely from the neural mapping's sensitivity to input noise.

Robustness Metrics

Performance is characterized via:

• Quality metrics: Median and lower-tail (CVaR$x$1 or min) IZI and IPI in neighborhoods around the anchor, reflecting typical and worst-case isolation.
• Stability metrics: Mean and RMS normalized spatial variation rates in dB/m, quantifying local sensitivity of metrics to coordinate perturbations.

Empirical Results

Simulations

Simulations are performed using a 24-channel split-band array with two listeners and up to $x$2 perturbed coordinate samples per anchor. Results show:

• In the woofer band, neighbor consistency reduces the RMS spatial variation rate by up to 55.9% (IPI) and minimally affects neighborhood median IZI ($x$3 relative to baseline).
• In the tweeter band, RMS variation rate reductions reach 30.3% (IZI), with median/neighborhood quality preserved and lower-tail robustness slightly improved.
• Step-per-meter variations for both IZI and IPI are substantially reduced, and the metric landscapes become spatially smoother under coordinate noise.

Measurement Validation

In real-room measurements using a 24-transducer array and two static head-and-torso simulators, the proposed regularization yields:

• Worst-case IZI at the perturbed listener improved by up to 16.9% (e.g., from $x$4 dB baseline to $x$5 dB with neighbor consistency at 10 cm spacing).
• RMS spatial variation rates for IZI at the fixed listener are reduced by up to 61.8% at the 2 cm perturbation level, indicating strong stability improvements.
• Both listeners benefit, despite coordinate perturbations affecting only one listener's conditioning input, due to the holistic nature of the neural mapping.

Hyperparameter Sensitivity

Optimal trade-off between stability and nominal isolation is achieved with a perturbation scale $x$6 m and loss weight $x$7 per band. Larger $x$8 or over-regularization leads to quality loss with diminishing stability gains, while too small $x$9 provides limited robustness.

Implications and Outlook

The neighbor-consistency regularization directly mitigates the principal failure mode of dynamic, coordinate-conditioned PSZ systems under localization uncertainty—a regime increasingly relevant as real-time tracking becomes standard. By smoothing the coordinate-to-filter mapping, the method relaxes requirements on tracking accuracy, potentially reducing hardware or calibration costs in practical deployments. In addition, empirical evidence suggests improved generalization under real-world acoustic variabilities, likely due to implicit regularization dynamics.

Theoretically, this approach connects to recent advances in consistency-regularized neural models, demonstrating its direct efficacy in an acoustic control setting. The formal introduction of decoupled evaluation protocols and stability metrics sets a new standard for quantifying robustness in control systems with input-conditioned neural synthesis.

Future Directions

• Extension to continuous head-tracking trajectories and formal perceptual validation of stability metrics.
• Adaptation to high listener density, arbitrary spatial configurations, or additional sound field control objectives.
• Exploration of more advanced regularization paradigms or architectures (e.g., spatial transformers, equivariant networks) for further gains in coordinate-space continuity and model robustness.

Conclusion

Neighbor-consistency regularization for coordinate-conditioned neural PSZ filter generation constitutes a principled and effective strategy for stabilizing audio zones against practical localization noise. Both simulation and experimental evaluation validate significant reductions in spatial variation rates with negligible quality trade-offs, ensuring robust and perceptually stable multi-listener audio rendering in noisy sensor environments. This framework provides a strong foundation for the deployment of learning-based PSZ systems in real-world dynamic audio applications (2605.21891).