Geo-DConv makes speech enhancement work on any microphone array
New framework adapts fixed-array models to any geometry, boosting performance consistently.
Multi-channel speech enhancement (SE) typically outperforms single-channel methods, but standard models are rigidly tied to a fixed microphone array geometry. This severely limits their deployment across devices with different microphone layouts (e.g., smartphones, smart speakers, hearing aids). While recent array-agnostic approaches handle varying microphone counts, they ignore explicit geometry priors, missing critical spatial cues for optimal filtering.
To address this, the authors introduce Geometry-Aware Dynamic Convolution (Geo-DConv), which dynamically generates convolution kernels based on 3D microphone coordinates. By injecting geometry information into the model during inference, Geo-DConv effectively makes any fixed-array SE architecture “array-invariant.” Experiments on the real-recorded RealMAN dataset show that two popular fixed-array models equipped with Geo-DConv achieve consistent performance improvements across diverse array configurations. This breakthrough allows a single model to be deployed on any device without retraining, significantly simplifying system design and boosting robustness in real-world scenarios.
- Exploits explicit 3D microphone coordinates as geometry priors for dynamic convolution generation.
- Tested on the real-recorded RealMAN multi-channel speech dataset with diverse array topologies.
- Enables two standard fixed-array models (e.g., Conv-TasNet, FaSNet) to achieve array-invariant performance.
Why It Matters
Enables robust speech enhancement across diverse devices without hardware-specific model retraining.