This study investigates various strategies for selecting relevant populations and their impact on forensic automatic speaker recognition in a likelihood ratio framework. Besides random and demographic metadata-based selection, it explores perceptual voice similarity as a potential criterion. Using a database controlled for gender, language, and recording conditions, an automatic approach based on phonetic features was used to select the most and least perceptually similar speakers to questioned speaker recordings in mock cases. It compares the strength of evidence obtained using these strategies for male and female mock cases and across different relevant population sizes. Random and metadata-based selections converge in log-likelihood ratio cost (Cllr) as the selected population size nears 50. While using perceptually similar speakers improves the overall Cllr, in individual cases effects may vary based on the degree of perceptual similarity or dissimilarity between recordings.