Purpose: This study examines online speech processing in typically developing and late-talking 2-year-old children, comparing both groups' word recognition, word prediction, and word learning. Method: English-acquiring U.S. children, from the "When to Worry" study of language and social--emotional development, were identified as typical talkers (n = 67, M[subscript age] = 27.0 months, SD = 1.4; Study 1) or late talkers (n = 30, M[subscript age] = 27.0 months, SD = 2.0; Study 2). Children completed an eye-tracking task assessing their ability to recognize both nouns and verbs, to use verbs to predict an upcoming noun's referent, and to use verbs to infer the meaning of novel nouns. Results: Both typical and late talkers recognized nouns and verbs and used familiar verbs to predict the referents of upcoming nouns, whether the noun was familiar ("You can eat the apple") or novel ("You can eat the dax"). Late talkers were slower in using familiar nouns to orient to the target and were both slower and less accurate in using familiar verbs to identify the upcoming noun's referent. Notably, however, both groups learned and retained novel word meanings with similar success. Conclusions: Late talkers demonstrated slower lexical processing, especially for verbs. Yet, their success in using familiar verbs to learn novel nouns suggests that, as a group, their slower processing did not impair word learning in this task. This sets the foundation for future work investigating whether these measures predict later language outcomes and can differentiate late talkers with transient delays from those with language disorders.