DP-MWF: Direction-Preserving MIMO Wiener Filter

• DP-MWF is a multichannel filtering technique that augments the traditional Wiener filter with directionality constraints to preserve spatial cues.
• It uses dynamic weighting to balance noise reduction and cue preservation, maintaining consistent performance despite variations in signal gain.
• The method is pivotal for binaural hearing aids and spatial speech enhancement, combining analytic and neural approaches for real-time audio processing.

The Direction-Preserving MIMO Wiener Filter (DP-MWF) is a family of advanced multichannel filtering techniques designed for spatial audio enhancement. Unlike conventional MIMO Wiener filters that focus exclusively on minimizing mean-squared error (MSE) between estimated and reference signals, the DP-MWF explicitly incorporates constraints to preserve the spatial (directional) properties of both target and residual noise components. This makes it highly pertinent in applications such as binaural hearing aids, microphone array speech enhancement, and spatial audio processing, where preserving cues—interaural level difference (ILD), interaural time difference (ITD), interaural coherence (IC), or the full multichannel covariance structure—is essential for downstream tasks like localization, beamforming, and binaural rendering (Carmo et al., 2021, Deppisch, 13 Apr 2026, Itturriet et al., 2018).

1. Signal Model and Classical Multichannel Wiener Filter

Let $M$ microphone signals be collected in the STFT domain as $\mathbf{y} = \mathbf{x} + \mathbf{v}$, with $\mathbf{x}$ the desired speech and $\mathbf{v}$ the additive noise. The classical MIMO Wiener filter seeks filter vectors (or a filter matrix, depending on output configuration) minimizing the expected MSE between the reconstructed target and desired reference signals:

$$J_{\rm MWF}(\mathbf w_L, \mathbf w_R) = \mathbb{E} \left[|q_L^T \mathbf y - \mathbf w_L^H \mathbf y |^2 + |q_R^T \mathbf y - \mathbf w_R^H \mathbf y |^2 \right]$$

where $q_L, q_R$ select reference microphones for left and right ears, respectively. The minimizing weights are constructed via the input and noise covariance matrices $\mathbf{R}_x$ and $\mathbf{R}_v$, yielding a closed-form Wiener solution (Itturriet et al., 2018, Deppisch, 13 Apr 2026).

However, conventional Wiener filtering generally distorts the interaural and spatial cues of the residual noise field—a significant drawback in spatial hearing applications.

2. Augmented Cost Function: Directionality-Preserving Regularization

To preserve spatial cues, the MWF cost is augmented with a penalty term targeting some binaural measure (BM) such as ILD, ITD, ITF, or IC:

$$J_{\rm cue} = \mathbb{E}\left\{ |\mathrm{BM}_{\rm out} - \mathrm{BM}_{\rm in}|^2 \right\}$$

For example, preserving the input-output Interaural Transfer Function (ITF):

$$\mathrm{ITF}_{\rm in} = \frac{q_L^T \mathbf{R}_v q_R}{q_R^T \mathbf{R}_v q_R}, \quad \mathrm{ITF}_{\rm out} = \frac{\mathbf{w}_L^H \mathbf{R}_v \mathbf{w}_R}{\mathbf{w}_R^H \mathbf{R}_v \mathbf{w}_R}$$

The combined cost function is

$\mathbf{y} = \mathbf{x} + \mathbf{v}$0

where $\mathbf{y} = \mathbf{x} + \mathbf{v}$1 is a non-negative scalar tuning the trade-off between noise reduction (MSE minimization) and preservation of the specified directional cue (Carmo et al., 2021, Itturriet et al., 2018).

3. Homogeneity, Dynamic Weighting, and Robustness

A critical insight is the mismatch in homogeneity degree between the MWF (quadratic in input power, i.e., 2-homogeneous) and the directional cue penalty (power-invariant, i.e., 0-homogeneous). Using a fixed $\mathbf{y} = \mathbf{x} + \mathbf{v}$2 leads to a cost function whose minimizer (i.e., filter solution) varies with overall input gain (the “Lombard effect”), causing the noise/cue trade-off to undesirably shift with changes in absolute signal or noise level (Carmo et al., 2021).

To address this, the weighting parameter is made dynamic: $\mathbf{y} = \mathbf{x} + \mathbf{v}$3, where $\mathbf{y} = \mathbf{x} + \mathbf{v}$4 is the estimated average input noise power. This scaling restores 2-homogeneity to the augmented cost:

$\mathbf{y} = \mathbf{x} + \mathbf{v}$5

ensuring that the optimal filter remains invariant to changes in absolute input gain. The practical implication is setpoint “locking” of the noise reduction vs. cue preservation trade-off, independent of input level changes (Carmo et al., 2021).

4. Analytic and Neural Implementations

Analytic (Covariance-Based) DP-MWF

Several analytic DP-MWF forms exist. In binaural applications, the solution is often written as a regularized Wiener-type system (possibly via linearization of a nonlinear directional penalty). For a parametric MWF, one has:

$\mathbf{y} = \mathbf{x} + \mathbf{v}$6

The direction-preserving variant, as in [Herzog '21] and (Deppisch, 13 Apr 2026), is

$\mathbf{y} = \mathbf{x} + \mathbf{v}$7

where $\mathbf{y} = \mathbf{x} + \mathbf{v}$8 is analytically determined and direct mixing with the identity matrix maintains the spatial eigenstructure of the noise covariance after filtering (Deppisch, 13 Apr 2026).

Neural Covariance Estimation

Neural approaches, such as OnlineSpatialNet, are deployed to estimate the frequency-domain spatial covariance $\mathbf{y} = \mathbf{x} + \mathbf{v}$9 in challenging, time-varying environments:

• Inputs are scale-normalized STFT frames.
• The output is a Cholesky factor $\mathbf{x}$0 such that $\mathbf{x}$1, enforcing Hermitian positive-definiteness.
• The loss combines a multichannel SI-SDR term and a Frobenius norm penalty on mismatch to the true noise Cholesky factor.

This neural estimation is integrated with analytic DP-MWF to yield a hybrid real-time system (Deppisch, 13 Apr 2026).

5. Practical Algorithms and Implementation

Implementation typically occurs in the STFT domain with the following core steps (Carmo et al., 2021, Deppisch, 13 Apr 2026, Itturriet et al., 2018):

1. STFT Analysis: Transforming microphone signals to the time-frequency domain.
2. Noise Covariance Estimation: Recursive averaging or neural estimation of $\mathbf{x}$2 and $\mathbf{x}$3.
3. Dynamic Parameter Computation: Estimation of average noise power to compute $\mathbf{x}$4, or analogous mixing parameters in DP-MWF with neural covariance.
4. Filter Solution: Solving for $\mathbf{x}$5 with regularized/augmented cost, possibly in block-matrix form or linearized through Newton-type methods.
5. Signal Reconstruction: Inverse STFT (ISTFT) for time-domain synthesis of output channels.

Complexity is dominated by per-bin matrix inversion (order $\mathbf{x}$6 per frequency), which is feasible for $\mathbf{x}$7 in real time.

6. Objective and Perceptual Performance

Extensive objective and psychoacoustic evaluation demonstrates the robust directional-cue preservation of DP-MWF versus conventional MWF or ITD-only penalty designs:

• Objective Metrics (e.g., DP-MWF vs. fixed-$\mathbf{x}$8 MWF-ITF):
  • Fixed-$\mathbf{x}$9: Noise-field ILD error rises >4 dB and ITD error >0.3 ms as input gain increases
  • Dynamic-$\mathbf{v}$0 DP-MWF: Keeps ILD error <1 dB, ITD error <0.1 ms across all gains, with only 1 dB less noise reduction (Carmo et al., 2021).
  • Binaural SNR, AITD, and AMSC errors confirm minimal trade-off loss versus unconstrained noise reduction (Itturriet et al., 2018).
• Psychoacoustic Localization:
  • DP-MWF maintains perceived noise azimuth tightly close to ground truth, with minimal hemisphere inversions and low interquartile error.
  • Fixed-$\mathbf{v}$1 or ITD-only penalized MWFs induce substantial localization bias or confusion (Carmo et al., 2021, Itturriet et al., 2018).

In streaming neural settings, DP-MWF with neural covariance estimation approaches oracle performance in SI-SDR, noise reduction, and spatial metrics, while requiring significantly lower computational resources than mask-based systems (Deppisch, 13 Apr 2026).

7. Applications, Significance, and Future Directions

DP-MWFs have become pivotal in hearing aid signal processing, spatial speech enhancement, and multichannel front-end processing for beamforming, binaural rendering, and DoA estimation, due to their ability to preserve the spatial integrity of both speech and noise. Dynamic weighting strategies based on homogeneity theory guarantee robustness to SNR and absolute input power variations, eliminating the need for manual tuning under changing acoustic conditions (Carmo et al., 2021).

Recent developments integrate neural estimation of spatial statistics, closing the gap with oracle performance while reducing parameter and compute demands, likely enabling future ultra-low-cost, real-time edge deployment (Deppisch, 13 Apr 2026).

A plausible implication is the extension of DP-MWF principles to more complex spatial hearing scenarios, joint speech and noise field modeling, and non-linear filtering architectures that exploit learned statistics while rigorously enforcing spatial constraints.