Applied researchers are often interested in linking individuals between two datasets that lack unique identifiers. Accuracy and computational feasibility are a challenge, particularly when linking large datasets. We develop a Bayesian method for automated probabilistic record linkage and show it recovers 40% more true matches, holding accuracy constant, than comparable methods in a matching of Union Army recruitment data to the 1900 US Census for which expert-labelled true matches are known. Our approach, which builds on a recent state-of-the-art Bayesian method, refines the modelling of comparison data, allowing disagreement probability parameters conditional on non-match status to be record-specific. To make this refinement computationally feasible, we implement a Gibbs sampler that achieves significant improvement in speed over comparable recent implementations. We also generalize the notion of comparison data to allow for treatment of very common first names that spuriously produce exact matches in record pairs and show how to estimate true positive rate and positive predictive value when ground truth is unavailable.