Neural Projection Filter Generation for Multi-Reference Active Noise Control

High-dimensional and correlated reference signals make multi-reference active noise control (ANC) computationally demanding, prone to slow convergence, and difficult to deploy in real-time applications. This letter introduces condition-aware projection filtering (CAPF), implemented by CAPFNet, which generates causal block-wise linear FIR projection filters for reference signal compression. The generated filters reduce 42 reference channels to 4 projected references while preserving the information required for effective ANC and maintaining compatibility with conventional adaptive algorithms. Simulations with measured in-vehicle road-noise recordings show that CAPF-Newton improves the average attenuation over FDFxNLMS by 2.6 dBA and achieves performance comparable to the neural reference projection-based filtered-x affine projection algorithm (NRP-FxAP), with a 48× reduction in online computational complexity.