The goal of this article is to contribute to our understanding of glides – their properties and distribution in underlying and surface representations, the range and the features of their various phonetic manifestations, and their role in the assignment of syllable structure and stress. Spanish provides a rich opportunity for carrying out this study because of the special properties of high vocoids in this language, which are systematically realised as glides in particular contexts: they can function as both onsets and rhymes; they can occur in prepeak, peak and postpeak position; up to four can occur in a row; and they take on a wide range of surface realisations.

In the pursuit of our goal, we confront a problem in Spanish phonology that has tantalised investigators for the better part of this century, and rightly continues to do so. The conundrum involves the two sets of phonetic segments we transcribe as [i j y [barred dotless j] ž jˇ] and [u w γw gw] (articulatory descriptions and feature characterisations are given below). Classical structuralist studies, and some current analyses as well, see the problem as the taxonomic exercise of assigning each of these segments to a particular ‘phoneme’ or ‘underlying segment’. Our study includes the notion of phonemic inventory, but considers it as only one of many intersecting issues involved in the attempt to elucidate aspects of the mental representations that native Spanish speakers employ in their phonological computations. We undertake to raise the level of discourse concerning the problem at hand not only by redefining it but also by enlarging the set of data and descriptive issues brought to bear on it and by situating the discussion in a rich theoretical context.

A major insight of Optimality Theory (Prince & Smolensky 1993) is that grammatical constraints are ranked and violable. These ranked constraints evaluate an infinite set of candidate output forms, of which the optimal one is the actual output. The winning candidate is a compromise between the potentially conflicting demands that grammatical constraints impose. A question that the OT literature has only rarely addressed is how ungrammaticality arises if constraints are violable and violation does not entail ungrammaticality.

In this paper, we point to some shortcomings of the only existing proposal to deal with ungrammaticality in OT, the special constraint MParse (Prince & Smolensky 1993). We propose a restructuring of the architecture of the OT constraint system that overcomes these shortcomings. We show that one of the great strengths of OT, that of separating well-formedness from the repair strategies to arrive at well-formed structures, is a weakness in dealing with absolute ungrammaticality. MParse forces us to consider what repairs might have been employed to fix up an ill-formed string. However, as we show in several cases, absolute ungrammaticality should be considered separately from the issue of possible repairs. Ungrammaticality results when the optimal form a grammar can produce still fails to satisfy a constraint governing ungrammaticality. MParse, as a component of Eval, requires us to evaluate multiple candidates, hence multiple repairs, simultaneously. We demonstrate that existence of a repair shown by particular alternations in a language (for example to avoid impermissible coda clusters) does not mean that the same repair will be available as a measure of last recourse to save an otherwise ungrammatical form (for example, to augment a subminimal form).

We propose to add a non-optimising constraint component called Control, which contains only those inviolable constraints that cause ungrammaticality rather than repair. If the winning candidate from Eval, the usual ranked and violable constraint component, satisfies all the constraints in Control, it is a grammatical output. If it violates a constraint in Control, no grammatical output is possible. This approach is empirically superior to MParse, and it also makes clearer a crucial distinction between two kinds of inviolable constraints that has not enjoyed much explicit attention in the literature. Inviolable constraints in Eval outrank all potentially conflicting constraints and cause repairs or block otherwise general alternations. Inviolable constraints in Control cause ungrammaticality, never repair.Two new developments in OT might possibly have a bearing on the success or failure of MParse. The first of these is McCarthy's (1998) Sympathy Theory. The second is Sprouse's (1997) Enriched Input Theory. Both of these models are in the early stages of development. There are no published references as yet for either. Furthermore, McCarthy (1999) is a revision of Sympathy Theory designed to reduce its currently excessive formal power. Since the proper form of these theories is as yet unclear, we refrain from discussing them here. To the best of our knowledge, however, our Control proposal is fully compatible with both.

While prosodic systems are typically either tonal or metrical, composite systems including both tone and metrical structure present yet another, albeit less frequent, possibility (Prince 1983, Hayes 1995). The focus of this paper is a prosodic system of the latter, more complex variety, evidenced in the Neo-štokavian dialect of Serbian or Croatian (henceforth NS; see Ivić 1958, 1985, Zec 1993). Although previously analysed as a pitch-accent system with no role allotted to the rhythmic structure (Browne & McCawley 1965, Ivić 1965, 1976, Halle 1971, Kenstowicz 1974, Lehiste & Ivić 1986, Inkelas & Zec 1988), on closer inspection NS discloses important resemblances with stress systems. Metrical structure is an independent agent in NS, as will be demonstrated here, and the entire prosodic system is characterised by a rich interplay between the tonal and metrical components.

Co-presence of tone and foot structure may in principle result in several types of interfaces between these two components. Cases documented in the literature are of two types: those in which the distribution of tone is constrained by metrical constituency, and those in which constraints operate in the opposite direction. third type of case has also been documented. In Japanese, as analysed in Poser (1984, 1990), the accentual system is tonal in nature while various templatic phenomena call for an inventory of iambic feet, yet the two systems are independent of each other and do not interact. The former type is exemplified by a number of languages extensively discussed in the literature: Creek (Haas 1977, Halle & Vergnaud 1987), Kirundi (Goldsmith 1987, Goldsmith & Sabimana 1989, Hayes 1995), Seneca (Prince 1983: 82–86), Winnebago (Susman 1943, Miner 1979, 1981, 1989, Halle & Vergnaud 1987, Hayes 1995 and the references therein) and Ancient Greek (Golston 1989, Sauzet 1989). The other type of unilateral interactions, with the rhythmic structure dependent upon the distribution of tone, is instantiated, for example, by Golin (Bunn & Bunn 1970, Hayes 1995).

In contrast, NS presents a case of bilateral interaction between tone and foot structure: tone exerts influence on the repertory of feet, and foot structure, in turn, constrains the distribution of tone. Although previously analysed as an instantiation of unilateral interaction, on a par with Golin (Inkelas & Zec 1988, Hayes 1995), new evidence to be presented here clearly disputes this position. This case is of immediate theoretical relevance for establishing the range of possible foot inventories. The resulting inventory is richer than in cases generally reported in the theoretical literature (Prince 1990, Mester 1994, Hayes 1995), and as such suggests a possible direction in which foot inventories may expand.

In Optimality Theory (OT), a grammar is a language-particular ranking of universal constraints (Prince & Smolensky 1993).This work was supported by the National Science Foundation under grant SBR- 9420424. I am grateful for comments received from Juliette Blevins, Mike Hammond, Paul de Lacy, Ania Łubowicz, Alan Prince, Doug Pulleyblank, Iggy Roca, Nick Sherrard and Jen Smith. Of course, I alone am responsible for the contents of this review. There are two types of constraints: markedness constraints prohibit (or require) certain output configurations; and faithfulness constraints demand identity in input → output mappings. Phonological generalisations are expressed by the interaction, through ranking, of these constraints.

In classical generative phonology (CGP), a grammar is a language-particular ordering of rules (Chomsky & Halle 1968). Though the rules are also language- particular, they are constructed using universal abbreviatory devices subject to an evaluation metric. Phonological generalisations are expressed by the rules and their ordering.