AI 中文总结
针对固定参数ANC在声路变化时性能下降的问题，提出反馈引导的DNN控制器融合框架，结合因果WaveNet与MoE模块，在1-8 kHz频段实现显著低频降噪且计算负载均匀。
AI 中文摘要
在主动噪声控制（ANC）系统中，自适应方法可能存在不稳定或发散问题，限制了其实际应用。因此，固定参数控制器被广泛采用，但在噪声特性和声路条件变化时，其性能会下降。本文提出一种反馈引导的基于DNN的控制器融合框架，用于鲁棒固定参数ANC。该方法将因果WaveNet控制器与反馈引导的混合专家（MoE）模块相结合，其中门控网络根据当前声路条件估计多个预训练FIR专家的权重。所提方法无需在线参数更新即可提升对变化声路条件的鲁棒性，且模型完全因果，支持逐样本流式推理，计算成本均匀分布在采样点以降低峰值计算负载。耳机ANC实验结果表明，该方法在1-8 kHz频段实现了显著的低频噪声降低，且噪声放大可忽略不计。
英文摘要
In active noise control (ANC) systems, adaptive approaches may suffer from instability or divergence, limiting their practical deployment. Consequently, fixed-parameter controllers are widely adopted, but their performance degrades under varying noise characteristics and acoustic path conditions. This paper proposes a feedback-guided DNN-based controller fusion framework for robust fixed-parameter ANC. The proposed method combines a causal WaveNet controller with a feedback-guided mixture-of-experts (MoE) module, where a gating network estimates the weights of multiple pre-trained FIR experts according to the current acoustic condition. The proposed approach improves robustness to varying acoustic conditions without online parameter updating. Furthermore, the model is fully causal and supports sample-wise streaming inference, with computational costs evenly distributed across sampling points to reduce peak computational load. Experimental results on headphone ANC demonstrate substantial low-frequency noise reduction with negligible noise amplification over 1-8 kHz.
