Abstract
Due to the slow convergence and poor tracking ability, conventional LMS-based adaptive algorithms are less capable of handling dynamic noises. Selective fixed-filter active noise control (SFANC) can significantly reduce response time by selecting appropriate pre-trained control filters for different noises. Nonetheless, the limited number of pre-trained control filters may affect noise reduction performance, especially when the incoming noise differs much from the initial noises during pre-training. Therefore, a generative fixed-filter active noise control (GFANC) method is proposed in this paper to overcome the limitation. Based on deep learning and a perfect-reconstruction filter bank, the GFANC method only requires a few prior data (one pre-trained broadband control filter) to automatically generate suitable control filters for various noises. The efficacy of the GFANC method is demonstrated by numerical simulations on real-recorded noises.
Introduction
Active Noise Control (ANC) is a critical technology for noise reduction in consumer electronics, such as noise-canceling headphones. However, traditional implementations face a fundamental trade-off between flexibility and speed. Standard adaptive algorithms, such as the Filtered-x Least Mean Square (FxLMS), are flexible but suffer from slow convergence, making them ineffective against rapidly changing or non-stationary sounds. Conversely, Selective Fixed-filter ANC (SFANC) offers near-instantaneous response times by selecting from a library of pre-trained filters. The primary drawback of SFANC is its rigidity; if the environmental noise does not closely match a pre-trained profile, performance degrades significantly.
The GFANC Framework
To address the "flexibility vs. speed" bottleneck, we propose Generative Fixed-filter Active Noise Control (GFANC). This method synthesizes custom filters dynamically, combining the rapid response of fixed filters with the adaptability of generative models.
The GFANC architecture is built upon three primary pillars:
- Filter Decomposition: Using perfect-reconstruction filter bank theory, a single high-quality broadband control filter is decomposed into $M$ sub-control filters. These frequency-specific components can be recombined to form a custom filter tailored to the current noise profile.
- Lightweight 1D CNN: A 1D Convolutional Neural Network with approximately 0.21 million parameters analyzes a 1-second window of raw noise. It predicts a binary weight vector that determines which sub-filters to activate.
- Adaptive Labeling: To facilitate training without manual labeling, an adaptive LMS algorithm is employed to identify optimal sub-filter combinations across a large dataset of noises, creating a self-labeled training set for the CNN.
Experimental Results and Discussion
The GFANC system was evaluated using synthetic data and real-world recordings of aircraft and traffic noise. The 1D CNN achieved a high classification accuracy of 97.20%, outperforming more complex architectures like MobileNet v2.
As shown in the performance comparison, GFANC demonstrates a significant advantage in dynamic environments. In simulations involving aircraft noise, GFANC provided 7dB to 12dB more noise reduction than both SFANC and FxLMS during critical transition periods. While FxLMS requires several seconds to converge, GFANC adapts almost immediately following the initial 1-second analysis window. Furthermore, GFANC eliminates the feedback loop required by traditional adaptive ANC, removing the risk of system divergence and the resulting acoustic artifacts.
Conclusion
GFANC represents a significant advancement in ANC technology by enabling the real-time generation of custom control filters. By offloading the CNN processing to a co-processor at a frame rate while maintaining sampling-rate cancellation on a dedicated controller, the system is highly practical for low-power mobile applications. Future work will focus on addressing the 1-second initial latency and exploring continuous weight vectors for even finer noise suppression.
Summary updated