Deep MCANC: A deep learning approach to multi-channel active noise control.

Area of Science:

• Acoustics
• Signal Processing
• Artificial Intelligence

Background:

• Traditional multi-channel active noise control (MCANC) relies on adaptive filtering, often requiring dedicated control units per channel.
• Existing methods can be complex and may struggle with diverse noise conditions.

Purpose of the Study:

• To introduce a deep learning-based approach for MCANC, termed deep MCANC.
• To develop a robust and effective noise cancellation system adaptable to various environments and noise types.

Main Methods:

• A convolutional recurrent network (CRN) is utilized for spectral mapping.
• The system encodes optimal control parameters and jointly computes multiple canceling signals.
• Summated power of error signals serves as the loss function for CRN training.
• Large-scale multi-condition training ensures robustness against diverse noises.

Main Results:

• Deep MCANC demonstrates effectiveness in wideband noise reduction.
• The approach shows good generalization capabilities for untrained noises.
• Experimental results confirm robustness against variations in reference signals and nonlinear distortions.
• Performance is evaluated across different configurations, including microphone and loudspeaker arrangements.

Conclusions:

• Deep MCANC offers a fixed-parameter, robust solution for multi-channel active noise control.
• The deep learning approach provides significant advantages over traditional adaptive filtering methods.
• The proposed method is effective and adaptable for real-world noise cancellation applications.