Room Impulse Response Reconstruction Using Pattern-Coupled Sparse Bayesian Learning With Spherical Waves

This work presents a pattern-coupled structured sparse Bayesian learning method for reconstructing room impulse responses (RIRs) in the time domain. It exploits the temporal properties of RIRs for improved reconstruction performance. Existing Bayesian methods with time-dependent regularization exploit the physical knowledge that an RIR transitions from a highly-sparse early part to a non-sparse later part. Building on this foundation, the proposed method utilizes a pattern-coupled hierarchical Gaussian prior in both spatial and temporal dimensions, employing a triangular mesh for the spatial dimension. The inverse problem is solved via stochastic variational inference. Performance assessments with experimental measurements illustrate the effectiveness of the proposed method.