Deep Generative Fixed-Filter Active Noise Control

Zhengding Luo, Dongyuan Shi, Xiaoyi Shen, Junwei Ji, Woon-Seng Gan

TL;DR

GFANC (Generative Fixed-filter Active Noise Control) bridges the gap between slow-to-adapt traditional algorithms and rigid fixed-filter systems. By using a lightweight 1D CNN to dynamically combine sub-filters from a single pre-trained template, it achieves near-instantaneous response times and superior noise suppression across diverse, real-world environments.

What Problem Does This Paper Address?

Active Noise Control (ANC) is the "magic" behind noise-canceling headphones, but it faces a fundamental trade-off.

Standard adaptive algorithms (like FxLMS) are flexible but slow; they take time to "learn" a noise, making them poor at handling rapidly changing sounds. On the other end, Selective Fixed-filter ANC (SFANC) is fast because it picks from a library of pre-trained filters. However, SFANC is limited by the size of its library—if the incoming noise doesn't match a pre-trained profile, the performance tanks.

The authors address this "flexibility vs. speed" bottleneck by proposing a generative approach that can synthesize a custom filter for almost any noise on the fly.

Methods

The researchers introduced GFANC, which relies on three clever architectural pillars:

  1. Filter Decomposition: Instead of training dozens of separate filters, they take one high-quality "broadband" control filter and decompose it into $M$ sub-control filters using perfect-reconstruction filter bank theory. Essentially, they slice the master filter into frequency-specific components that can be perfectly stitched back together.
  2. Lightweight 1D CNN: They designed a 1D Convolutional Neural Network (only 0.21M parameters) that runs on a co-processor (like a smartphone). It analyzes a 1-second window of raw noise and predicts a binary weight vector. This vector tells the system which sub-filters to "turn on" to cancel the current noise.
  3. Adaptive Labeling Mechanism: Training a CNN usually requires massive amounts of labeled data. To avoid manual labeling, the authors used an adaptive LMS algorithm to automatically find the optimal sub-filter combinations for a massive dataset of noises. This "self-labeling" allowed the CNN to learn the relationship between raw audio and the best filter configuration.

Datasets

  • Synthetic Training Data: A massive dataset of 84,000 synthetic noise tracks (1-second each) used for training, validation, and initial testing.
  • Real-World Evaluation: To prove the system works outside the lab, they tested it on real-recorded aircraft noise (50Hz–8,000Hz) and low-frequency traffic noise (40Hz–1,200Hz).

Key Insights and Ideas

  • Binary is Better: By using binary weights (0 or 1) for the sub-filters rather than continuous values, the authors simplified the problem into a multi-label classification task. This significantly reduces the computational load on the mobile co-processor.
  • Eliminating the Feedback Loop: Unlike traditional adaptive ANC, GFANC doesn't require a feedback error signal to update its filter in real-time. This removes the "risk of divergence"—that annoying moment when an ANC system fails and produces a loud screech.
  • The Co-processor Strategy: The heavy lifting (the CNN) happens at a "frame rate" (once per second) on a co-processor, while the actual noise cancellation happens at the "sampling rate" on a dedicated low-power controller. This split-architecture is highly practical for modern consumer electronics.

Key Results

  • High Accuracy: The proposed 1D CNN achieved 97.20% accuracy in predicting the correct filter weights, outperforming much larger 2D networks like MobileNet v2.
  • Superior Suppression: In aircraft noise simulations, GFANC provided 7dB to 12dB more noise reduction than SFANC and FxLMS during the critical transition periods (2s–3s mark).
  • Rapid Response: While FxLMS struggled to converge on non-stationary mixed noises (aircraft + traffic), GFANC adapted almost immediately after the first second of audio analysis.

Limitations and Future Work

  • The "First Second" Problem: Because the CNN requires a 1-second buffer of audio to generate a filter, the system cannot effectively cancel noise during the very first second of a new sound event.
  • Binary Constraints: While binary weights are efficient, future work might explore "soft" or continuous weights to see if even finer-grained noise cancellation is possible without a massive jump in power consumption.
  • Hardware Implementation: The paper relies on numerical simulations. The next logical step is deploying this on physical DSP hardware to measure actual battery drain and real-world latency.

Summary updated

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.
BibTeX
@inproceedings{luo2023deep,
  author = {Zhengding Luo and Dongyuan Shi and Xiaoyi Shen and Junwei Ji and Woon-Seng Gan},
  title = {{Deep Generative Fixed-Filter Active Noise Control}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10095205},
}

Loading related content...

Figure