Eight experiments tested the hypothesis that infants' word segmentation abilities are reducible to familiar sound-pattern parsing regardless of actual word boundaries. This hypothesis was disconfirmed in experiments using the headturn preference procedure: 8.5-month-olds did not mis-segment a consonant-vowel-consonant (CVC) word (e.g., dice) from passages containing the corresponding phonemic pattern across a word boundary (C#VC#; "cold ice"), but they segmented it when the word was really present ("roll dice"). However, they did not segment the real vowel-consonant (VC) word (ice in "cold ice") until 16 months. Yet, at that age, they still did not false alarm on the straddling CVC word. Thus, infants do not simply respond to recurring phonemic patterns. Instead, they are sensitive to both acoustic and allophonic cues to word boundaries. Moreover, there is a sizable developmental gap between consonant- and vowel-initial word segmentation.