Despite the massive variability present in speech, theories of language acquisition often assume for simplicity that children’s language-learning input is equivalent to hyper-clear speech in which words are always pronounced consistently. This dissertation reconsiders such assumptions while investigating four aspects of early language acquisition. Chapter 1 uses computational modeling to investigate speech segmentation (extraction of words from continuous speech), which has traditionally been characterized as a problem of tracking statistics over canonical or “dictionary” pronunciations. Using actual transcribed pronunciations, we found that previous models of speech segmentation performed worse, and face serious challenges for lexicon building, because even correctly extracted wordforms (e.g., er) show a complex many-to-many relationship with word types (e.g., or, her, are). This suggests a need to rethink the traditional theory that infants assume phonological (speech-sound) distinctions imply lexical distinctions and treat all speech as equally valuable input. Chapter 2 uses internet-based eye-tracking to test infants’ early phonolexical representations (representations of how words sound). Even though in real speech phonological distinctions (e.g., want vs. wan’) don’t always imply lexical distinctions, we found that infants are already sensitive to small mispronunciations (e.g., dog vs. mog) by 9-14 months, suggesting phonologically precise representations. Chapter 3 uses corpus analysis to investigate the input to word learning: Does the clarity of parents’ speech align with moments of referential transparency (when the intended referent is easy to guess from the visual scene), such that filtering out some instances would be helpful for learning? Our results suggest yes. Moments when parents said a word more clearly were also moments when the referent was more likely to be apparent—potential “gems” for multimodal word learning. Finally, Chapter 4 uses in-lab eye-tracking to test children’s word recognition given less clear speech. We found that toddlers were unable to recognize naturalistically less clear pronunciations of familiar words, and even 3-year-olds were not adultlike, with significantly better performance in repetition contexts. Altogether, this work suggests that young children may not process all the speech they hear and highlights a need for more research on how speech clarity is distributed relative to other properties of children’s language-learning environments.