Abstract There have been numerous studies investigating the perception of non-native sounds by listeners with different first language (L1) backgrounds. However, research needs to expand to under-researched languages and incorporate predictions conducted under the assumptions of new speech models. This study aimed to investigate the perception of Dutch vowels by Cypriot Greek adult listeners and test the predictions of cross-linguistic acoustic and perceptual similarity. The predictions of acoustic similarity were formed using a machine-learning algorithm. Listeners completed a classification test, which served as the baseline for developing the predictions of perceptual similarity by employing the framework of the Universal Perceptual Model (UPM), and an AXB discrimination test; the latter allowed the evaluation of both acoustic and perceptual predictions. The findings indicated that listeners classified each non-native vowel as one or more L1 vowels, while the discrimination accuracy over the non-native contrasts was moderate. In addition, cross-linguistic acoustic similarity predicted to a large extent the classification of non-native sounds in terms of L1 categories and both the acoustic and perceptual similarity predicted the discrimination accuracy of all contrasts. Being in line with prior findings, these findings demonstrate that acoustic and perceptual cues are reliable predictors of non-native contrast discrimination and that the UPM model can make accurate estimations for the discrimination patterns of non-native listeners.