Speaker
Description
Covariance Matrix Fitting estimates acoustic source strengths on a predefined focus grid by fitting a modeled microphone Cross-Spectral Matrix (CSM) to the sample CSM and is often formulated as a nonnegative LASSO problem. Classical algorithms for the LASSO need model selection and many iterations for accurate reconstructions. Algorithm unfolding reduces this effort by learning selected algorithmic parameters from ground-truth source maps.Unfolded algorithms for CMF use a dictionary from an analytic propagation model in each unfolded iteration. In measurements, the unknown true propagation model may differ from this analytic model. Supervised training creates a second mismatch because ground-truth source maps are unavailable for measured array data, so training must rely on synthetic data generated with an analytic propagation model. This work therefore compares supervised training using the Mean-Squared Error (MSE) objective with unsupervised training, focusing on different objectives including LASSO, Cross-Validation (CV), and a smooth approximation of the Bayesian Information Criterion (BIC).In the matched setting, CV and BIC can reach validation NMSEs similar to supervised training using the MSE objective. Under propagation model mismatch, direct unsupervised training on sample CSMs generated from measured transfer functions from MIRACLE-A1 yields sparser source maps than supervised training on perturbed synthetic data.