Lund researchers unveil causal sound field reconstruction for real-time audio

In a new preprint on arXiv (2605.20403), David Sundström and colleagues from Lund University tackle a fundamental problem in real-time sound field control: reconstructing a sound field from causal, time-windowed microphone measurements. Traditional methods process each frequency band independently, ignoring correlations introduced by windowing, which degrades performance in latency-sensitive applications like active noise cancellation or immersive audio.

The team formulates a causal finite-window spatio-temporal linear minimum mean-square error (LMMSE) estimator. They model the sound field as the solution to the wave equation driven by a stationary stochastic source distribution, yielding a physically interpretable covariance function closely linked to the classical diffuse-field coherence model. To manage the computational complexity—which grows rapidly with spatio-temporal observations—they introduce a budget-constrained sample selection strategy that minimizes posterior reconstruction variance. Evaluated on both simulated and measured sound fields, the proposed approach significantly outperforms frequency-domain finite-window baselines, especially for short analysis windows.