When a word is embedded in a nonsense matrix such that it belongs to two syllables of the matrix, detection of the word is inhibited when the second syllable is full but not when it is reduced: Thus melt is detected significantly faster in [mɛltə∫] than in [mɛltajv]. The effect is not determined by matrix or syllable length or by implicit morphological structure. Strictly left-to-right models of auditory word recognition predict no such difference because the embedded word should be detected as soon as it ends. Comprehension models which assume an obligatory syllabic level of representation also predict no difference since syllable boundaries occur at the same point in both matrices. The obtained result is, however, predicted by a language-universal model of continuous speech recognition proposed by the present authors, which emphasizes the importance of juncture detection. Applied to English, this model asserts that occurrence of a clear syllable boundary signals the onset of a full (stressed) syllable, which in turn implies an island of acoustic reliability and high phonetic yield, to which perceptual attention is diverted.