Transformer-based End-to-End Control Filter Generation for Active Noise Control

Active Noise Control (ANC) traditionally relies on fixed filters or supervised learning with labeled noise data. A new paper from Nanyang Technological University (Ziyi Yang et al.) introduces a Transformer-based End-to-End Control Filter Generation (E2E-CFG) framework that eliminates these limitations. Unlike prior Generative Fixed-Filter ANC (GFANC) methods that decompose, weigh, and recombine sub-filters, E2E-CFG directly generates control filters in an unsupervised manner. The key innovation is a fully differentiable system where the co-processor and real-time controller are trained end-to-end using the accumulated error signal as the sole objective. The Transformer’s attention mechanism captures global and dynamic noise characteristics, enabling real-time adaptation without manual feature engineering.

Experimental results on real-recorded noises show that E2E-CFG achieves superior noise reduction performance and adaptability compared to the original GFANC framework. By abandoning the decomposition–reconstruction pipeline, the proposed method simplifies the control loop and prevents error accumulation. This work, available on arXiv (2605.00494), represents a step toward more intelligent, data-efficient ANC systems that can be deployed in hearing aids, headphones, and industrial noise cancellation without requiring labeled training data.