Interpolation of Spatial Room Impulse Responses Using Partial Optimal Transport

Abstract

Interpolation between spatial room impulse responses (SRIRs) is necessary for dynamic acoustic rendering in which a listener can move with six degrees-of-freedom. The early part of the SRIR consists of sparse direct and reflected sound events, whose arrival time, direction and level vary with receiver position. Interpolation of the spatio-temporal structure necessitates the non-trivial task of mapping corresponding sound events. Instead of finding an exact map, we propose using partial optimal transport to find a coupling between reflections requiring neither estimation of the room geometry nor explicit knowledge of the source-receiver configuration. Each SRIR is first decomposed into a virtual source space. Then, the interpolated impulse response is calculated based on a partial optimal transport coupling obtained with linear programming. We compare the method against two baseline interpolation methods using simulated SRIR data, and show that it best preserves the temporal fine structure of the omnidirectional response.