Generalization of perceptual adaptation to second-language speech in a single-talker training paradigm: Predictors remain elusive

Listeners often struggle to understand speech from non-native (L2) talkers but can adapt to individual talkers through exposure. While research shows that training on multiple talkers helps listeners generalize their adaptation to new, unseen talkers, the effectiveness of training on a single talker is inconsistent. This study investigated whether two specific factors drive this generalization: the intelligibility of the training talker and the acoustic similarity between the training and test talkers.

The researchers conducted two experiments using a large dataset of L2 English talkers. They pre-selected talker triplets to systematically manipulate specific variables while controlling for others. Experiment 1 tested the intelligibility hypothesis by pairing a test talker with either a high- or low-intelligibility training talker. Experiment 2 tested the acoustic similarity hypothesis by pairing a test talker with training talkers of high or low acoustic similarity (based on self-supervised model representations). All listeners performed a sentence-in-noise transcription task, and performance was compared against a no-training baseline to measure generalization.

The results failed to support the primary hypotheses. In Experiment 1, while training improved overall performance compared to the no-training baseline, training on a high-intelligibility talker provided no greater benefit than training on a low-intelligibility talker. In Experiment 2, neither high- nor low-similarity training talkers led to a significant, predictable advantage in generalizing to the test talker. Furthermore, the expected symmetry of learning—where training on A to understand B should mirror training on B to understand A—was not observed. While many talker pairs showed some evidence of positive generalization, there was no systematic, reliable predictor for which pairs would succeed.