<p>Studies suggested that learners can use both prosodic and distributional information for word segmentation. Interestingly, recent work on statistical word segmentation has suggested rhythm perception plays a significant role. This is because syllable sequences used to study statistical word segmentation are generated by concatenating uniform-length words, which generate a rhythm percept that is consistent with the word boundaries. In this study, we manipulate high-order constraints on how words are concatenated together, such as whether a word can immediately follow itself, and examine its effect on learning. In addition, we ask participants to rate differently constructed test items, so we can model their behavior computationally. We find that learning was comparable regardless of whether the input sequence contains immediate repetition. Moreover, the ratings of different test items are inconsistent with predictions from second-order statistics. We provide a unified theoretical framework to understand these results, where our model generates predictions on how syllable sequences are segmented and represented. The rhythm model predicts that the higher-order organization of the words does not influence learning as long as the features leading to rhythm perception are preserved. Moreover, we can leverage the outputs of the rhythm model to show that first-order statistics can explain results well, better in fact than second-order statistics. These results reassert the theoretical centrality of first-order statistics, such as pairwise counts and positional counts, which have traditionally been shown to influence artificial language learning.</p>