Adaptive far-field spatial-temporal sound prediction using attentive one-dimensional U-Net

Abstract

Predicting far-field acoustic signals from near-field measurements is crucial for effective acoustic design and sound control. This study presents a method utilizing a one-dimensional U-Net neural network that enables low-latency, three-dimensional spatial-temporal predictions of far-field acoustics, requiring only a limited number of near-field waveform inputs. The proposed approach autonomously adapts to diverse indoor acoustic environments, accounting for variations in source positions, air temperatures, and reverberation times. The integration of a self-attention mechanism further enhances prediction accuracy. Experimental results indicate that the proposed method achieves high signal-to-noise ratios in far-field predictions. Additionally, the arrangement of near-field receivers influences the prediction, and the effectiveness of the self-attention mechanism is illustrated under varying levels of disruptive noise.