Selective fixed-filter active noise control based on convolutional neural network

Abstract

Active noise control (ANC) technology is increasingly ubiquitous in wearable audio devices, or hearables. Owing to its low computational complexity, high robustness, and exemplary performance in dealing with dynamic noise, the fixed-coefficient control filter strategy plays a central role in portable ANC implementation. Unlike its traditional adaptive counterpart, the fixed-filter strategy is unable to attain optimal noise reduction for different types of noise. Hence, we propose a selective fixed-filter ANC method based on a simplified two-dimensional convolution neural network (2D CNN), which is implemented on a co-processor (e.g., in a mobile phone), to derive the most suitable control filter for different noise types. To further reduce classification complexity, we designed a lightweight one-dimensional CNN (1D CNN), which can directly classify noise types in time domain. A numerical simulation based on measured paths in headphones demonstrates the proposed algorithm’s efficacy in attenuating real-world non-stationary noise over conventional adaptive algorithms.