A comparison framework for sound field reconstruction methods using Bayesian inference

In recent years, the use of data-driven methods in sound field analysis has advanced rapidly. These approaches often require a large number of observations, which are scarce in acoustics. Hybrid approaches offer a promising alternative by integrating computational techniques with sound field models and theoretical considerations, thereby restricting the solution space. This work introduces a Bayesian framework for sound field reconstruction that incorporates general acoustic constraints through prior distributions. The framework presents a modular design, whereby multiple sound field models (plane waves, spherical waves and diffuse kernel) can be used, as well as regularization in both time and frequency domains. Additionally, the Bayesian formulation enables cross-frequency information sharing by selecting the appropriate frequency covariance matrix. The framework’s performance is evaluated in two different real-life scenarios, including interpolation and extrapolation from the measurement array. This comparison highlights that the optimal sound field model is contingent on the characteristics of the observed sound field. The outcome of this study is twofold: first, it presents a modular framework that merges classical sound field reconstruction methods with a Bayesian inference formulation; second, it evaluates the performance of different reconstruction methods in different experimental scenarios.