A Comparative Study on Room Impulse Response Reconstruction Using Pattern-Coupled Sparse Bayesian Learning with Different Coupling Structures

Abstract

Sparse Bayesian learning (SBL) is widely used for sound field reconstruction (SFR). Among various SBL approaches, pattern-coupled SBL has been demonstrated to achieve superior performance. Building on the pattern-coupled SBL framework, this study replaces matrix multiplication with tensor and matrix cross-correlation operations, significantly reducing the algorithm’s spatial and temporal complexity. Furthermore, we compare the performance of different coupling structures within the pattern-coupled SBL method for reconstructing room impulse responses (RIRs) in the time domain. Specifically, we analyze nine coupling structures that incorporate both temporal and spatial coupling terms and validate their performance via experiments using two datasets. The results indicate that the coupling structure known as CPC-ST (Centered Pattern-Coupled with Spatio-Temporal Coupling) achieves the best performance, especially in the extrapolation of the sound field. For lightweight systems, where a slight performance trade-off is acceptable, the coupling structure known as CPC-S (Centered Pattern-Coupled with Spatial Coupling) is also recommended due to its balance between effectiveness and simplicity.

Keywords

• sound field reconstruction
• pattern-coupled hierarchical model
• coupling structures
• variational inference