Predictive Directional Selective Fixed-Filter Active Noise Control for Moving Sources via a Convolutional Recurrent Neural Network

Abstract: Directional Selective Fixed-Filter Active Noise Control (D-SFANC) can effectively attenuate noise from different directions by selecting the suitable pre-trained control filter based on the Direction-of-Arrival (DoA) of the current noise. However, this method is weak at tracking the direction variations of non-stationary noise, such as that from a moving source. Therefore, this work proposes a Predictive Directional SFANC (PD-SFANC) method that uses a Convolutional Recurrent Neural Network (CRNN) to capture the hidden temporal dynamics of the moving noise and predict the control filter to cancel future noise. Accordingly, the proposed method can significantly improve its noise-tracking ability and dynamic noise-reduction performance. Furthermore, numerical simulations confirm the superiority of the proposed method for handling moving sources across various movement scenarios, compared to several representative ANC baselines.

• The paper presents PD-SFANC, a proactive ANC approach using a CRNN to predict next-frame directions for moving noise sources.
• It employs a dual-module architecture that separates real-time control from predictive computation, ensuring rapid, delayless filter switching.
• Experimental results demonstrate stable noise reduction above 15 dB NRL under varying source movements, outperforming traditional ANC methods.

Predictive Directional Selective Fixed-Filter ANC with CRNN for Moving Noise Sources

Introduction

The paper "Predictive Directional Selective Fixed-Filter Active Noise Control for Moving Sources via a Convolutional Recurrent Neural Network" (2604.23144) addresses the critical challenge of real-time active noise control (ANC) for non-stationary noise sources, focusing on scenarios where the direction-of-arrival (DoA) changes dynamically. Traditional ANC algorithms, such as filtered-reference LMS (FxLMS), are hampered by slow adaptation and convergence constraints, resulting in suboptimal suppression for moving noise sources. The directional selective fixed-filter ANC (D-SFANC) approach improves upon this by incorporating directional information, but suffers from lag in filter switching, degrading performance during rapid source movement.

The paper proposes Predictive Directional SFANC (PD-SFANC), leveraging a convolutional recurrent neural network (CRNN) architecture as a predictive filter selector. The CRNN forecasts future DoA using temporally stacked spectral features from multichannel microphone signals and proactively selects the optimal pre-trained control filter for the imminent frame. This predictive strategy reduces latency, aligns filter switching with source movement, and enables robust, delayless suppression even in rapidly fluctuating acoustic environments.

Figure 1: Comparative depiction of (a) directional SFANC with lagged filter updates and (b) PD-SFANC with proactive filter selection enabled by DoA prediction.

System Architecture

The PD-SFANC architecture is comprised of two synergistic modules: a real-time controller and a co-processor. The co-processor executes the CRNN to predict the next-frame DoA from $K$ consecutive frames of $J$-channel reference signals transformed into the STFT domain. The CRNN's output is the index of the DoA class, which dictates filter selection from a pre-trained library covering spatial angles in discrete increments. The real-time controller applies the selected filter at the sampling rate, achieving proactive ANC without dependence on feedback error signals or filter adaptation loops.

All CRNN parameters are automatically learned during training, eliminating manual tuning. The block diagram (Figure 2) illustrates the dual-module architecture with real-time noise control decoupled from CRNN-induced latency.

Figure 2: Block diagram of PD-SFANC featuring the CRNN-based DoA prediction and delayless filter selection for moving-source ANC.

CRNN-Based Next-Frame DoA Prediction

For robust forecasting, the CRNN utilizes multi-frame spectral context of $J$-channel reference signals. Input preprocessing involves STFT transformation and concatenation of magnitude and phase spectra across the channel and temporal dimensions. The CRNN comprises:

• Three convolutional blocks (2D convolutions, group normalization, ReLU, max pooling) to extract time-frequency spatial features.
• Adaptive average pooling to collapse frequency information while preserving temporal evolution.
• A gated recurrent unit (GRU) to encode inter-frame dependencies crucial for trajectory modeling.
• A fully connected layer with softmax activation for class distribution over DoA indices.

This configuration is succinctly shown in Figure 3. The cross-entropy loss is employed for optimization, and the network is trained with the Adam optimizer.

Figure 3: CRNN architecture for next-frame DoA prediction using temporally stacked multichannel spectral inputs.

Pre-Trained Control Filter Design

For each discrete DoA, a filter vector is pre-trained via the FxLMS algorithm with broadband bandlimited white noise. The pre-trained library spans a uniform azimuth grid, enabling rapid, context-aware filter switching. This library circumvents the need for error-driven online adaptation, enhancing system stability and reducing divergence risks.

Proactive and Delayless Noise Control

By forecasting the source trajectory, PD-SFANC updates the control filter with minimal delay, maintaining noise suppression efficacy across transitions. Once the initial $K$ frames provide historical context, subsequent filter updates occur each frame, synchronized with the predicted DoA. This eliminates buffer delays and obviates feedback-driven adaptation.

Numerical Evaluation

Dataset and CRNN Effectiveness

Synthetic and real-world noise datasets are constructed with diverse room impulse responses, array positions, reverberation levels, and SNR values. Training, validation, and test sets include random source motion modes (static, constant-rate, and time-varying-rate) with precise next-frame DoA labels.

The CRNN achieves classification accuracy exceeding 90% even under challenging unseen room-SNR combinations, validating its generalization capacity and suitability for deployment in computationally constrained environments.

Noise Reduction Performance

Experiments are conducted in a simulated enclosure with a four-microphone reference array and one error microphone. PD-SFANC is compared against FxLMS, D-SFANC, and dynamic factor graph-based SFANC (DFG-SFANC) under two vacuum cleaner noise movement scenarios:

1. Constant-rate source movement (linear DoA trajectory): As depicted in Figure 4, PD-SFANC outperforms D-SFANC by eliminating the filter selection lag and consistently aligning filter switching with source trajectory. NRL remains above 15 dB for most of the trajectory; FxLMS fails to converge rapidly, D-SFANC exhibits high fluctuations, and DFG-SFANC struggles with rapid acceleration scenarios.

   Figure 4: Noise reduction and filter selection performance under constant-rate source movement (linear DoA change).

2. Time-varying-rate source movement (sinusoidal DoA trajectory): Figure 5 shows that PD-SFANC robustly tracks the source trajectory, with sustained NRL and minimal performance drops, contrasting the instability observed in DFG-SFANC and limited adaptation in FxLMS and D-SFANC.

   Figure 5: Noise reduction and filter selection performance under time-varying-rate source movement (sinusoidal DoA).

Strong numerical results indicate that PD-SFANC achieves stable, high noise reduction (>15 dB NRL) and rapid adaptation to non-stationary directional trajectories.

Implications and Future Directions

PD-SFANC delivers a robust solution to moving-source ANC, combining CRNN-based predictive spatial tracking with pre-trained filter libraries. The elimination of manual parameter tuning and feedback error signal dependency streamlines system design and deployment, offering real-time efficacy for devices ranging from headphones to vehicle cabins.

Theoretical implications include advancing proactive spatial control strategies in ANC by integrating deep learning–based trajectory prediction. Practically, PD-SFANC can be adapted to multi-source environments with further architectural extensions and higher-dimensional spatial filter banks. Future research may focus on online learning for multi-noise sources, scalable filter libraries, and CRNN-driven adaptive spatial encoding under variable acoustic constraints.

Conclusion

The proposed PD-SFANC method introduces a predictive, delayless noise control paradigm for moving sources. Leveraging a CRNN to forecast DoA and proactively select optimal control filters, the system achieves robust tracking and superior attenuation across diverse motion profiles. The approach advances the state-of-the-art for spatial ANC, with strong numerical validation and broad applicability in dynamic acoustic environments.