Abstract
Binaural hearing aids utilize Multichannel Wiener Filters (MWF) to reduce background noise while attempting to preserve essential spatial cues, such as Interaural Time and Level Differences (ITD/ILD). However, traditional fixed weighting parameters in these filters often lead to spatial cue distortion when signal power fluctuates—a common occurrence in real-world environments due to the Lombard effect (where speakers raise their voices in noise). This paper introduces a dynamic weighting strategy that ensures "setpoint invariance" by mathematically aligning the scaling behavior of the noise reduction and spatial preservation terms. Experimental results demonstrate that the proposed method maintains stable spatial cues across a 30dB range of power variations, effectively preventing the perceived "drift" of sound locations typical of standard noise reduction algorithms.
Introduction
Hearing aid users rely on binaural cues—tiny differences in timing, level, and correlation between the two ears—to localize sound sources. While modern hearing aids use MWF to enhance speech in noisy environments, these filters often distort spatial cues, potentially making a noise source on the periphery sound as if it is coming from the front. To mitigate this, researchers add a penalty term to the filter's cost function, controlled by a weighting parameter ($\alpha$).
Historically, $\alpha$ has been a fixed value. However, this study identifies a critical flaw: a fixed weight fails when the absolute power of the signal changes. In dynamic environments where the Lombard effect occurs, the hearing aid "drifts" between prioritizing noise reduction and spatial accuracy, leading to an unstable acoustic experience for the user.
Methodology
The researchers proposed a dynamic weighting parameter designed to achieve mathematical homogeneity across the filter's objective function.
- Homogeneity Analysis: The team proved that while standard noise reduction cost functions scale quadratically with signal power (degree 2), spatial cue penalty terms typically do not scale with power (degree 0). This mismatch causes the filter's performance to shift as volume changes.
- Dynamic Weighting: To resolve this, they redefined the weighting parameter as $\alpha = \beta \cdot g^2$, where $g^2$ represents the average input noise power and $\beta$ is a fixed constant. This ensures the entire objective function scales uniformly, maintaining the optimal filter characteristics regardless of absolute volume.
- Worst-Case Design: A specific algorithm was developed to select the constant $\beta$ based on the lowest expected signal-to-noise ratio (SNR), ensuring spatial cues are preserved even in challenging acoustic environments.
Experimental Setup
The method was evaluated using 32 audio files from the Harvard database and ICRA speech-like noise. The acoustic environment was simulated using Head-Related Impulse Responses (HRIRs) from a KEMAR manikin. Additionally, 16 volunteers participated in psychoacoustic lateralization experiments to verify the real-world perceptual impact of the proposed mathematical model.
Key Results
The proposed method, referred to as MWF-ITF-R, demonstrated significant improvements in both objective metrics and human perception:
- Objective Stability: As shown in Figure 6, the proposed dynamic weighting (blue diamonds) allows the spatial cue errors (ILD and ITD) to stabilize and converge toward the target reference (black squares) as the Lombard gain increases. In contrast, the standard fixed-weight method (red triangles) maintains high error levels across the entire gain range.
- Perceptual Accuracy: In listening tests simulating the Lombard effect, the standard method caused the perceived noise location to drift from -60° toward the front (0°). The proposed method successfully kept the perceived noise at approximately -54°, closely matching the unprocessed perception of -71°.
Conclusion
By identifying the scaling mismatch between noise reduction and spatial awareness goals, this research provides a computationally efficient solution for robust binaural processing. The dynamic weighting strategy ensures that hearing aids remain stable even as speakers and noise levels fluctuate. Future work will explore expanding this robust design to include Interaural Coherence (IC) to improve performance in reverberant, echo-prone environments.
Summary updated