This talk explores the hypothesis that syllable divisions are, under specific conditions, not securely discovered and learned. This is not to say that "syllables don't exist". Rather, the argument developed here is that language learners may encounter more uncertainty when they seek to categorize strings in terms of their syllabic division than when they categorize the same strings in purely segmental terms. For instance, the set of English strings {VprV, VbrV, VfrV, VplV, VblV, VflV, VtrV, VdrV, VkrV, VklV, VgrV, VglV} may be described in two ways which are nearly equivalent observationally: as (1),

     (1) V-Complex Onset-V

or as (2):

     (2) V-Non-strident Obstruent-Liquid-V.

The two descriptions diverge only for rare strings like VtlV, whose rarity insures that their phonological patterning may not be immediately accessible to the learner. For this string set and the analytical choice (1) vs. (2), the syllabic description (1) may be disfavored if the learner is in any doubt regarding the syllabic parse of the {VprV, etc.} set: if he cannot discover how VprV is divided, then he cannot tell that [pr] is an onset. In such a case, he may favor the segmental analysis (2) of the entire string class, since there is no comparable ambiguity about the segmental categorization of the cluster. We can anticipate that, under the circumstances described, generalizations which could have been expressed syllabically, will more likely be learned as descriptions of segmental strings, independent of prosodic structure.

The talk outlines the circumstances under which syllable divisions are either insecurely learned or not learned at all and relates such cases to phonological patterns in which a potential syllable-based generalization is in fact learned as a segmental, linear condition.