Measuring phonetic distance between speakers is essential for understanding speech intelligibility in noise, yet methods for quantifying these differences vary widely. Recent advances in AI-based speech representations offer new approaches to capture talker-listener similarity, but their effectiveness relative to traditional acoustic metrics remains unclear. We compared four types of phonetic distance metrics—legacy acoustic measure (ACCDIST), automated edit distances, embedding-based measures from self-supervised models, and novel logit-based measures—to determine which best predicts intelligibility within a homogeneous accent community. Forty standard southern British English (modern RP) speakers participated as both talkers and listeners in a sentence recognition task across four signal-to-noise ratios. Talker typicality—measured as average distance from a talker to all other talkers—strongly predicted intelligibility, with individual talker-listener distance providing additional explanatory power. Embedding-based metrics from an English phoneme recognizer optimally captured the phonetic variability relevant to intelligibility, with logit-based measures offering a computationally efficient alternative. These findings demonstrate that fine-grained phonetic variation within a single accent group systematically affects comprehension in adverse conditions, extending accent-distance theory to the idiolectal level and providing practical guidance for selecting distance metrics in speech intelligibility research.