TCN-KAN: A Parameter-Efficient and Causal Neural Network for Nonlinear Active Noise Control

Active Noise Control (ANC) deployment is often hindered by hardware nonlinearity and strict memory constraints. While deep learning approaches offer superior noise suppression, they typically suffer from prohibitive parameter sizes unsuitable for embedded implementation. To address this, we propose TCN-KAN, a parameter-efficient and causal framework. By combining a Causal Dilated TCN with a Kolmogorov-Arnold Network (KAN) decoder utilizing learnable spline activations, the model efficiently captures complex secondary path nonlinearities. Experimental results demonstrate that TCN-KAN achieves state-of-the-art noise reduction and exceptional robustness against severe saturation distortion. Crucially, this performance is achieved with only 0.109 million parameters-a reduction of over two orders of magnitude compared to leading baselines-making TCN-KAN an ideal “compute-for-memory” solution for memory-critical ANC applications.