Calibrating geometric room acoustic models using a gradient descent algorithm

Samuel Bellows (1) , Brian F. G. Katz (1)

1 IJLRDA-LAM - Lutheries - Acoustique - Musique

One aspect of obtaining realistic results in room acoustic simulations for virtual reality environments or room acoustic design is careful material property adjustments to achieve agreement between the simulation output and measured results. Nonetheless, calibrating geometric room acoustic simulations presents several challenges in selecting absorption and scattering coefficient values. This work aims to improve model calibration efforts by developing a gradient-descent algorithm that minimizes differences between simulated and measured results. The algorithm follows by taking the derivative of room acoustic metrics such as reverberation time or clarity with respect to the absorption coefficient values. The resultant gradient thus derives from a single deterministic ray-tracing result represented by a set of ray paths and a set of absorption coefficients values for each surface. This formulation consequently enables efficient computational realization in calibrating absorption coefficients. Incorporating scattering coefficient adjustments would require further repeated simulations, currently not optimized. A reverberation-time optimized initialization and inequality constraints help to maintain realistic absorption coefficient values. Application of the calibration algorithm to an abbey church demonstrates the robustness and efficiency of the method.

Domaines: Son [cs.SD]; Acoustique [physics.class-ph]

Samuel Bellows, Brian F. G. Katz. Calibrating geometric room acoustic models using a gradient descent algorithm. Audio Eng Soc Conv 156, Audio Eng Soc, Jun 2024, Madrid, Spain. ⟨hal-04639133⟩