Neural MVDR beamformer learns WNG constraints for robust speech enhancement
Deep learning now adapts microphone array robustness in real-time, outperforming fixed thresholds.
The minimum variance distortionless response (MVDR) beamformer is a classic technique for multichannel speech enhancement, but its real-world performance degrades due to microphone self-noise and array mismatches. Traditional approaches rely on fixed, manually tuned white noise gain (WNG) thresholds or diagonal loading, which fail under unknown or time-varying acoustic conditions. In a paper accepted to INTERSPEECH 2026, Deng et al. propose a fully data-driven MVDR framework that replaces manual tuning with a deep neural network. The network jointly predicts two components: a time-frequency noise mask for accurate covariance estimation, and a frequency-dependent WNG threshold that adapts the beamformer's robustness versus directivity trade-off dynamically.
A differentiable robust MVDR layer is integrated, allowing end-to-end optimization. Experiments show consistent gains in both objective speech quality (e.g., PESQ, STOI) and intelligibility over conventional fixed-WNG MVDR methods, without requiring any prior knowledge of microphone noise levels. This work bridges traditional array processing with modern deep learning, offering a practical path to robust speech enhancement in hearing aids, smart speakers, and teleconferencing systems where microphone conditions are unpredictable.
- Jointly learns noise mask for covariance and frequency-dependent WNG threshold via a single neural network.
- Differentiable robust MVDR layer enables end-to-end training, eliminating manual diagonal loading.
- Accepted to INTERSPEECH 2026; outperforms fixed-WNG MVDR in speech quality and intelligibility metrics.
Why It Matters
Adaptive WNG control makes MVDR beamforming practical for noisy, dynamic environments like smart speakers and hearing aids.