Multichannel learning-based spatially extended active noise control via model matching and sensor transfer function interpolation

This paper presents a learning-based multichannel active noise control (L-MANC) method that utilizes a model-matching framework along with a sensor interpolation approach. In a feedforward ANC scenario, L-MANC aims to cancel the primary noise field for an extended control region with a limited number of measurements. To achieve global control, the training data is generated by interpolating the acoustic transfer functions (ATFs) based on a small number of measurements via kernel ridge regression. Noise control filters for the secondary loudspeakers are learned using a convolutional recurrent neural network (CRNN) in an encoder-decoder architecture. In the training phase, the network parameters are obtained using the mean square error criterion in a model-matching paradigm. Simulations are performed for an ANC system consisting of a nine-loudspeaker linear array and eight measured control points. The proposed L-MANC approach is tested for various unseen scenarios with variations in reverberation time and location of the primary noise source. The proposed system has demonstrated superior ANC performance in terms of mean square noise reduction (MSNR) over several baselines.