CRNN-based spatial active noise control in spherical harmonics domain

Abstract

In this work, an active noise control (ANC) in the spherical harmonics (SH) domain, along with a learning model, is developed. The traditional ANCs are built on the adaptive signal processing framework and fail in the presence of nonlinear distortions. Further, these algorithms are limited to the horizontal plane only. In contrast, the proposed work develops the ANC in the 3D space using the SH representation with the spherical microphone array (SMA) as the error microphone. The SH decomposition calculates the SH pressure coefficients from the SMA recordings. Subsequently, a convolutional recurrent neural network (CRNN) is trained to estimate the real and imaginary spectrograms from the SH coefficients as the input features. The output of the CRNN is the canceling signal that eliminates or attenuates the primary noise in the ANC system. A delay-compensated method addresses ANC system latency. According to simulations and experiments, the proposed learning-based spatial ANC reduces wideband noise and generalizes to untrained noises.