This study investigated how adult second language (L2) learners of Mandarin Chinese use knowledge of phonological and lexical statistical regularities when acoustic information is insufficient for word recognition. A gating task was used to test intermediate L2 learners at two time points across a semester of classroom learning. Native Mandarin speakers (tested once) served as a control group. Mixed-effects modeling revealed that upon hearing truncated speech, L2 learners, like native speakers, identified high token frequency syllable-tone combinations more accurately than low token frequency syllable-tone combinations. Error analysis of correct syllable/incorrect tone responses revealed that native speakers made specific probability-based errors. L2 learners primarily demonstrated more acoustic-based errors but exhibited a trend toward greater probability-based errors during the second test. These findings are interpreted in light of L2 speech learning models that emphasize a statistical learning mechanism.