Synthesizing source directivity using DeepONet room acoustic predictions

Abstract

Incorporating source directivity when simulating spatial room impulse responses for virtual environments is crucial for achieving natural results. Directly injecting the directional source as the initial condition for a wave equation to solve would require re-running a numerical solver for each new directivity pattern, source position or orientation. Here we show that learned surrogate models can lead to large efficiency benefits if such variation shall be simulated. Although training such models is a computationally expensive process, performing inference over them is typically fast. Here we show how to exploit that property by training a deep neural operator network (DeepONet) only using omni-directional sources, and predicting the pressure on a variable grid of points, to which directivity filters are applied to generate arbitrary patterns. We show the fit of directictivity patterns to their re-synthesized versions from the output of a DeepONet.