Learning-Based Real-Time Multichannel Dynamic-Zone Active Noise Control Using the Model-Matching Principle

This paper presents a real-time Dynamic-Zone Active Noise Control (DZANC) system based on the modelmatching principle. Our goal is to suppress noise in multiple zones of a room. The proposed system is implemented using the Conv-TasNet to estimate the array filter coefficients for the desired zone. The residual noise power in the controlled zones is employed as the training loss function. In the training phase, the designated zone vector is concatenated with the primary noise signal as the network input. In addition, kernel ridge regression is used to interpolate the Acoustic Transfer Function (ATF) between adjacent measured control points to expand the effective control region with only a limited number of measurements. To validate the proposed multichannel DZANC system, simulations were performed using a six-loudspeaker linear array. The results demonstrated superior noise reduction in the controlled zone compared to several traditional digital signal processing (DSP)-based approaches. In addition, the proposed system effectively creates a quiet zone where the user is located by switching the filter coefficients learned by the Conv-TasNet.