Deep-Learning-Based Multi-Point Sound Field Control in Reverberant Environments

In reverberant environments, multipath propagation severely limits the performance of conventional sound field control methods such as active noise control (ANC) and spatial audio reproduction. This paper presents an end-to-end deep learning framework for multi-point sound field control that operates using only a small set of target control points. Unlike traditional model-based approaches, the proposed system directly learns to generate the secondary loudspeaker’s driving signals by minimizing the residual sound pressure at the specified control positions. A temporal– frequency neural architecture incorporating convolutional and attention layers is employed to capture both the spatial correlations and temporal dynamics of reverberant fields. The method requires knowledge of the room impulse responses (RIRs) between loudspeakers and microphones but does not depend on explicit room geometry or boundary conditions. Simulation results in complex acoustic environments demonstrate that the proposed approach achieves up to 10 dB of noise reduction at controlled positions while maintaining strong spatial selectivity compared with robust FxLMS and wave-domain baselines. These findings highlight the potential of deep learning as a scalable and flexible solution for real-time sound field control in reverberant spaces.