the Iberian Languages, guest-edited by G. Elordieta and M. Vigario. Primary Word Stress in Brazilian Portuguese and the Weight Parameter

In this paper, we develop an analysis of primary word stress in Brazilian Portuguese (BP). We evaluate the typological and language-specific arguments that are presented in the literature against the relevance of syllable weight in Portuguese, and show that none of them appears to be valid when confronted with cross-linguistic evidence or the facts of BP phonology. We then go on to show that stress in BP represents a mixed system, in which verbs receive stress as a function of the morphological categories of tense (past, present, future), whereas stress in non-verbs is prosody-based and sensitive to the distinction between heavy and light syllables. We finally propose a constraint analysis of this system, which we claim functions in the lexical part of a stratified model.


Introduction
In this paper * , we will discuss a number of observations which strongly suggest that syllable weight plays an important role in predicting the main stress of non-verbs in Brazilian Portuguese (henceforth BP), as well as in other parts of the BP phonological grammar. We will evaluate the arguments that are presented in the literature against the relevance of syllable weight and show that none of them appears to be valid when confronted with the facts of BP phonology. Moreover, since the stress patterns of Spanish and BP are to a large extent comparable, a demonstration of the phonological significance of syllable weight in BP must also consider the arguments put forward against a weight-sensitive stress rule for Spanish 1 . * I wish to thank Ben Hermans, Haike Jacobs, Luiz-Carlos Schwindt, Marina Vigario, and an anonymous JPL reviewer for their comments on a prefinal draft of this paper. The writing of this paper was made possible by a grant from the Netherlands Organization of Scientific Research (NWO) grant number 365-70-006.
We will subsequently sketch a constraint analysis that captures the main facts of BP stress by way of two sets of constraints. One is morphology-based and predicts primary stress in verbs on the basis of morphological category. The other is prosody-based and locates stress in nonverbs in function of syllable weight.

Assessing the arguments against the relevance of syllable weight in BP
Phonologists of Portuguese either have formulated serious doubts with respect to the relevance of syllable weight or have dismissed it altogether. Many analyses of Portuguese stress are based upon the observation that, at least in non-verbs, stress falls on the last vowel of the stem, where 'stem' is defined as the lexical word, disregarding the 'desinence', which in most cases is one of the thematic vowels /e,o,a/ 2 . Thus, the thematic noun cabeça 'head' has prefinal stress, whereas in the athematic nariz 'nose' stress is on the word-final syllable. This stembased analysis successfully predicts main stress in a large majority of the Portuguese non-verbs.
However, both the fact that stress could never fall on the antepenultimate vowel in a word like aberto 'open' (with a prefinal heavy syllable), and the fact that athematic non-verbs usually end in a heavy syllable (feliz 'happy'), remain, from this perspective, purely accidental. A mechanism that takes syllable weight to be a conditioning factor for main-stress placement is at least equally successful. If properly worked out, a quantity-based account provides a principled explanation for the systematic absence of proparoxytonic stress in words with a prefinal heavy syllable, and also explains why stress-placement in newly created words is systematically governed by the closed vs. open syllable distinction. Given this situation, one may ask why phonologists working on Portuguese have been so reluctant to admit that syllable weight plays a role in at least part of the phonological grammar of BP. In this section we will present the arguments that have been put forth in the literature to dismiss the relevance of syllable weight and evaluate their validity against the background of phonological typology and the facts of BP phonology.

Kurylowicz's Universal
One argument that is frequently mentioned to dismiss the relevance of a moraic approach to the stress rules of Portuguese, and also of Spanish and Italian, is based on the typological claim that contrastive vowel length is a prerequisite to weight-sensitive stress. The interest of the argument does not only derive from its persistence in the literature (see, among others, Kury owicz, 1948;Newman, 1972;Greenberg and Kaschube, 1976;Hyman, 1977;Ohsiek 1978;Hyman 1985;Roca, 1990;Andrade and Laks, 1991), but also from the fact that Trubetzkoy (1939) is usually referred to as the source for this implicational universal. In his Grundzüge der Phonologie, translated into English as Principles of Phonology, which is how we will henceforth refer to it, Trubetzkoy expressed himself in a way that could be interpreted along the lines suggested by the critics of a moraic approach to stress in Romance. We provide Trubetzkoy's full statement below: 3 (1) The interpretation of long nuclei as geminated, or in terms of multinumber constituency in general, may be regarded as an 'arithmetic conception of quantity'. Languages in which this conception finds expression are 'mora counting' languages since in these languages the smallest prosodic unit does not always coincide with the syllable (1939/69:177).
According to Trubetzkoy, languages may represent phonemic vowel length in several ways, one of which is by way of a geminate vs. single vowel opposition. Languages that choose this option are mora-counting languages. There are also languages that have a vowel length contrast in which the long member is not a geminate vowel, but an 'extended' short vowel. In these languages, which are not mora-counting languages, vowel length is the phonetic realization of the phonological category of 'intensity'.
The phonologists who refer to Trubetzkoy in the context of Romance stress interpret the cited passage in the following way: Spanish, Italian, and Portuguese cannot have a geminate representation of vowel length, because long vowels are not phonemic. They cannot, therefore, be mora-counting and, consequently, their stress system cannot depend on syllable weight. In Wetzels (2002a;2003), it was shown that this interpretation is based on the erroneous assumption that Trubetzkoy's conception of the mora is equivalent to the one underlying present theories of syllable weight. As it turns out, Trubetzkoy's view as expressed in citation (1) represents the end point of a search into the phonological interpretation of vowel quantity, which he had started almost twenty years earlier. The fundamental and only question that Trubetzkoy was concerned with was how to represent contrastive vowel length. Nowhere was his attention directed towards the factors that cause stress to select syllables with specific properties in some languages, and, consequently, nowhere was he explicitly interested in the relation between weight-sensitive stress and phonological length. One illuminating example that demonstrates how different Trubetzkoy's mora concept was from the one that underlies present theories of syllable weight comes from Italian, a language that opposes single consonants to geminate ones. In Trubetzkoy's view, the fact that this language uses the analytical quantity correlation in consonants does not in itself qualify it as a mora-counting language. Still in Principles of Phonology Trubetzkoy is explicit in including Italian in the class of non-moraic languages. Consequently, in Italian, one part of the geminate belongs to the left-hand syllable, the other part to the right-hand syllable. 4 Also the fact that in Principles of Phonology Trubetzkoy still distinguishes two different ways of expressing contrastive vowel length clearly shows that his concept of the mora relates to the representation of vocalic nuclei, not of syllables. To the best of our knowledge, the universal implication defined in (2) below was never formulated and probably never even envisioned by Trubetzkoy 5 .
(2) Weight-sensitive stress implies contrastive vocalic length Regarding the intellectual origin of (2) it appears significant that, of all the linguists mentioned earlier, Kurylowicz (1948) is the only one who does not attribute it to Trubetzkoy -rightly so, as we have just seen 6 . To the best of our knowledge, it was Kurylowicz himself who first claimed explicitly that syllable quantity can only exist in languages with vowel quantity, as he states: In the remainder of this study we will therefore refer to the implicational law (2, 3) as Kurylowicz's Universal. The validity of Kurylowicz's Universal as a cross-linguistic generalization is called into question in this section via the numerous languages that count as counterexamples.
Let us start by observing that in present theories of the mora the question of phonological weight in tautosyllabic VC clusters does not depend on the presence of long vowels in a principled way.
According to Hyman (1985), moras are prosodic 'units of weight' that belong to segments or segment sequences. Lexical associations between segments and weight units are one-to-one, except for phonologically long segments, which are associated with two weight units. Underlying associations between weight units and segments may differ from surface associations. One difference is systematic and is explained as the effect of a universal Onset  (4): Consonants that do not immediately precede a vowel may remain moraic in some languages or lose their mora by language-specific rules of mora deletion and reassociation. In Latin, for example, the coda consonant contributes to the weight of the syllable as is shown by the stress rule of the language, among other rules. In Hyman's view, the Latin coda consonants remain moraic at least until the level at which stress assignment takes place. Since coda consonants in Latin are not syllable peaks phonetically, the Margin Creation Rule (MCR) removes a consonantal weight unit from the representation, after the rules that need to refer to weight have applied. The effect of the language-specific MCR is shown in (5) A slightly different formalization of the function of the mora in the phonology of the world's languages is proposed by Hayes (1989). As in Hyman's theory, vowels are universally moraic and in languages with a vocalic length contrast long vowels are represented as single segments linked to two moras. Consonants are not moraic underlyingly, except for geminate consonants, which are attached to a single mora in their lexical representation. Additional moras are assigned to coda consonants by a language-specific rule of 'Weight by Position' (WbP). As in Hyman's proposal, the prediction is made that long vowels universally count as heavy for the purpose of rules that discriminate between light and heavy syllables. The case of closed syllables constituting a natural class with syllables containing a long vowel is accounted for by a mechanism that adds moras to the representation (WbP) 9 .
An interesting prediction that follows from mora theory and which bears directly upon the weight issue under discussion is that, if VC syllables do indeed count as bi-moraic in languages with a vowel length contrast only, one would not expect C(ompensatory) L(engthening) to be possible in languages without phonemic long vowels 10 . Hayes, who explicitly addresses this question, concludes: "a language that lacks a vowel length contrast, but has a syllable weight contrast, can create surface long vowels through a process essentially equivalent to CL" (1989:290 We conclude that Kurylowicz's Universal does not survive the test of empirical verification.

Mixed systems
In this section we briefly address another typological claim, which was made recently by Andrade and Mateus (2000:117), who state: "…the pure quantity-sensitive hypothesis seems, in principle, to be incompatible with the coexistence of two stress subsystems, one for nouns and one for verbs".
To the extent that the above statement refers to (variants of) Portuguese, we agree with Andrade and Mateus, because BP is not a pure quantity-sensitive system, in the sense that syllable weight is only relevant for non-verbal lexical categories. If the statement is about languages in general, which we assume it is, meaning that there can be no stress system in which a quantity-sensitive subsystem co-exists with a subsystem that is not sensitive to syllable quantity, it is not valid as a cross-linguistic generalization, as was also shown in Wetzels (2003).
Many, perhaps even most, languages that assign stress on the basis of prosodic categories possess at least some suffixes that disrupt the general prosody-based positioning of the primary accent (quantity-sensitive or not), by being pre-accenting or stress-attracting. Among the languages mentioned earlier, Inga and Stoney combine a subsystem that is weight-sensitive with a subsystem that is morphologically conditioned. An example that closely resembles Portuguese in making a distinction between verbs and non-verbs is Archi, a Southern Daghestanian language. 12 In Archi, stress is assigned within a two-syllable window at the left edge of the word. Stress in non-verbs is sensitive to the open vs. closed syllable distinction, as well as to the degree of sonority of the vowels that occur in the initial or post-initial syllable. In Archi, monosyllabic nominal roots are always closed. Suffixes may be consonant-initial or vowelinitial. When a consonant-initial suffix is added to the root, the root syllable remains closed and stress will always be initial: dump 'ball', dúmp-li 'ball-erg sg', dúmp-mul 'ball-pl'. If the addition of a vowel-initial suffix causes the root syllable to become open, stress will be on the second syllable, unless its nuclear vowel is less sonorous than the vowel in the first syllable:

Spondaic Lowering
In the literature on Portuguese stress, the issue of syllable weight was also addressed in Andrade and Laks (1991) who provide the following arguments against weight sensitivity in Portuguese main stress, alongside the ones discussed in sections 2.1 and 2.2: i. Syllable weight plays no role in other parts of Portuguese phonology; ii. The assumption that syllable weight is relevant for main stress makes it impossible to derive main and secondary stress 'in one sweep'.
Here we will turn to the first of these arguments, leaving the second for a later section. Since we now start addressing arguments that are specific to Portuguese and to BP in particular, we must determine what counts as a heavy syllable in BP.
The BP syllable allows for only two positions in the rhyme. Moreover, the non-peak position is reserved exclusively for sonorant segments (high vowels, liquids, the nasal mora) and /s/. The latter phoneme can moreover be adjoined to the syllable as a second coda element.
We may consequently represent the BP syllable rhyme as in (6): BP has both open and closed syllables. With regard to stress, we may use the notion 'heavy rhyme' in its most general interpretation, which is that any syllable that has two filled rhyme positions counts as heavy. The list of possible rhymes in BP is presented in (7)  There is at least one phonological rule in BP that is conditioned by syllable weight. This rule, called 'Spondaic Lowering' in Wetzels (1992;, accounts for the neutralization of stressed mid vowels in prefinal syllables followed by a (final) heavy syllable. Spondaic Lowering, a lexical rule that applies almost without exception in the nonderived vocabulary, is completely productive in paroxytonic words derived by derivational suffixes that consist of a heavy syllable (-vel, -oN, -il). This generalization is stated in (8)

Primary and secondary stress
We have seen that Andrade and Laks (1991) criticize the quantity-sensitive (henceforth QS) analysis for not allowing the derivation of main and secondary stress in one sweep. A similar criticism is formulated by Roca (1990), when he observes that Spanish secondary stress is quantity-insensitive (henceforth QI).
It is true that syllable weight is not relevant for secondary stress in BP. This fact makes it impossible to assign primary and secondary stresses with a single QS mechanism that would apply in an iterative fashion. However, what looks like an analytical weakness at first sight turns out to be an argument in favor of keeping the two phenomena apart, if we consider the different properties of primary and secondary stresses and their status in the phonological grammar of BP. First, secondary stress shows variation, unlike primary stress. It was observed by Collichon (1994) that secondary stress in BP usually follows a binary pattern, but a wordinitial dactyl is also possible, as in abaca where unstressed [¨] shows that /kaf¨¦ / was used as the base for the formation of the diminutive with the suffix -zinho, which has prosodic-word status 21 ). Note, however, that clashes or potential clashes of primary stresses are configurations of accents whose location is fixed categorically at a deeper level of representation.
Primary stress in BP is a lexical phenomenon, in the sense that the constraints that account for its distribution interact with morphology by way of selecting specific lexical categories (verbs, non-verbs). Also, as we will see below, within the class of verbs, the specific distribution of main stress is conditioned by the inflectional categories of tense. As is to be expected, primary stress in BP is intertwined with lexical (and postlexical) phonology. A large part of the lexical constraints that deal with vocalic alternations are directly or indirectly related to main stress. Spondaic Lowering discussed in the previous section was an example of a lexical constraint that refers to the location of primary stress. Many more generalizations in the lexical phonology of BP are stress-dependent, especially in the area of vowel alternations. One of these is unstressed vowel neutralization, which reduces the number of vocalic contrasts by merging the upper and lower mid vowels to a single series of (upper) mid morphology that refer to secondary stress. Our interpretation of this situation is that primary and secondary stress in BP are assigned in different components of the grammar, as was also argued in Wetzels (1992). We will assume that these different components correspond to the levels that are usually referred to as lexical and postlexical phonology. In this sense, we regard the different ways in which primary and secondary stress function in the grammar of BP as a strong argument in favor of a model, whether derivational or constraint-based, that distinguishes these two components.
When considered from a typological perspective, one observes that in many languages culminative and rhythmic stress are governed by different factors. This is obvious in languages where primary stress is unpredictable, or conditioned by morphology. Similarly, while languages are known that assign primary stress on the basis of the degree of sonority of the syllable nucleus, there seem to be no languages in which rhythmic stress is sonoritydriven. In languages where both culminative and rhythmic stress are sensitive to weight, it may happen that the segment classes that count as heavy are not identical for both. According to Hayes (1995:333), Chugach is a language that counts both long vowels and closed syllables as heavy for primary stress, whereas only long vowels are considered heavy for the sake of rhythmic stress. Exactly the opposite situation is found in Sri Lanka Creole Portuguese, analyzed by Smith (1978) (see footnote 8). Goedemans (forthcoming) establishes on the basis of a sample of 97 StressTyp languages in which weight is relevant for the attribution of primary and/or secondary stress, that in 32 languages primary stress is QS while secondary stress is not. Inversely, in 17 languages primary stress is QS and secondary stress QI. Finally, in 8 languages both primary and secondary stresses are QS, but the definition of what counts as heavy is different for primary accent and rhythm. In less than half of the QS languages in the sample, weight is defined uniformly for primary and secondary stresses. These and other asymmetries between primary and secondary stress have been used by Van der Hulst (1996) to defend his Primary Accent First (PAF) theory 23 .

Palatal onsets in final syllables
Most scholars agree that proparoxytonic stress in BP is incompatible with a prefinal heavy syllable. Supplementary evidence for the awareness of BP native speakers of this restriction can be obtained from hypercorrect diphthongization.
The optional monophthongization of the diphthongs /ej/ and /aj/ is a process that occurs throughout Brazil 24 . It happens most frequently before /r/, but also before the palatal fricatives /! , " /, and more frequently in stressed syllables than in unstressed ones. Some examples are given in (11): The variation created by monophthongization leads to hypercorrect diphthongization in words that contain /e/ or /a/ in the contexts in which usually diphthongs are reduced to monophthongs, as in the examples in (12): Just like optional monophthongization, hypercorrect diphthongization occurs in stressed and unstressed syllables alike. However, it is never observed in post-tonic position in proparoxytonic words like câmera (*câmejra) 'camera', cólera (*cólejra) 'cholera', ópera (*ópejra) 'opera', áspero (*áspejro) 'rough', efêmero (*efêmejro) 'ephemeral', etc. One may conjecture that the reason why diphthongization never occurs in these words is because it would create a prosodic pattern that is prohibited in the language 25 .
It was observed by Harris (1983) and Roca (1990) for Spanish that a palatal onset in the final syllable prevents leftward stress shift from a light penultimate. According to Roca the systematic absence of the pattern XV % CVC palatal V# cannot be explained by a weight-based algorithm, because "there is not a shred of evidence to suggest that these palatals can be analyzed as consonant clusters synchronically…(1990:154)." With his remark, Roca refers to the fact that stress in Spanish, like in Portuguese, cannot 'skip' a prefinal heavy syllable.
Since palatal consonants cannot be analysed as hetero-syllabic clusters, some ad hoc stipulation is necessary to express the property that these consonants are pre-accenting when they occur in the onset of the word-final syllable. As in Spanish, in BP antepenultimate syllables cannot be stressed if the final syllable begins with a palatal glide or consonant.
However, it was shown in Wetzels (1997) that, at least in the case of BP, there is sufficient proof for the fact that these sounds add weight to the preceding rhyme 26 . We summarize the discussion in Wetzels (1997) below.
Within the set of words that belong to the exceptional proparoxytonic pattern (médico 'physician', cérebro 'brain', pássaro 'bird', etc.) the presence of the word type V % XVC palatal V# (e.g. *álcunha) is conspicuously absent. This would not be alarming if this word type were rare, because one could argue that as proparoxytonic words are exceptional anyway, the absence of the *álcunha type is an accidental gap. In reality, however, this word-type is not (14) Distribution of weak and strong /R/ coda mar sea carta letter c.
between diphthong and vowel beira edge f. between vowels carro car caro expensive As can be observed in (14), with the exception of the contrast intervocalically, weak and strong /R/ are in complementary distribution. This indicates a common phonological source for both /R/s. Furthermore, the lexical qualitative distinction that seems necessary to account for the intervocalic contrast can be eliminated in favor of a quantitative distinction between a simple and a geminate /R/. This analysis has the advantage of explaining the stress behaviour of words that have strong /R/ in the onset of their final syllable: the syllable preceding strong /R/ would count as heavy and the ungrammaticality of words like *cáchorro 'dog' could be explained on a par with all other words that have a prefinal heavy syllable.
Other facts of BP phonology show that, at least morpheme-finally, both /R/-sounds derive from a single lexical source, as is clear from the following alternations: . The same behavior can be observed in speech errors: Angenot and Vandresen, 1981: 96).
Although the precise formalization of the distribution of /) / and /0 / is not crucial to the point at issue here, we will assume that both the alveolar tap and the velar approximant are derived from a segment that is lexically underspecified, which we will represent, adopting Hayes ' (1989) representation for geminates, as in (16): The different ways in which underlying /R/ is realized phonetically (its intervocalic occurrences included) can now be derived from its distributional and structural properties. In principle, the phonological spell-out of the underspecified segment can, for a given context or for all contexts, Of the word types that are listed in (13) camponês 'countryman'), be derived from an underlying oral vowel followed by a tautosyllabic nasal mora (VN), an analysis which was adopted in Wetzels (1997). Here we will concentrate on allophonic nasalization.
Many speakers of BP, especially in southern Brazil, realize an alternation in the stem /am/ of the verb amar, as in such forms as ámo 'I love', with a nasal /ã/, and amár 'love-inf', in which /a/ is oral (see also Vandresen, 1975 As can be observed in the words given in (17), the coronal and labial nasal stops, on the one hand, and the palatal nasal stop, on the other, behave differently in their choice for the targets of nasal spreading. Quite surprisingly, nasalization is general before / ' /, but not before /n, m/. To be more precise, allophonic nasalization before palatal nasals occurs independently of the position of primary stress, just like contrastive nasalization. The robustness of this phenomenon was confirmed in a striking way by a large-scale survey carried out in Brazil by Abaurre and Paggotto (1995). These scholars found that contrastive nasality was realised in 100% of the cases. Moreover, allophonic nasalization was categorical in (primary) stressed syllables. They then observe: "Nasalization is categorical when a vowel precedes a palatal nasal consonant, independently of whether the vowel is stressed or unstressed" (1995:24) 33 .
If we are allowed to explain the similar behavior of contrastive nasality and allophonic nasalization before / ' / in terms of a similar lexical representation, we must conclude that in the case of / ' /, nasality is located in the syllable coda. Of course, the immediate consequence of this hypothesis is that, whatever precise underlying representation for / ' / is chosen, it will be such that it makes the preceding syllable heavy. A geminate representation of palatal nasals would account for the observed parallel. / and /m,n/ is borne out by the facts. In the example set (18), the question mark stands for the elements that may occupy the second position in the rhyme preceding the nasal onset. The reader will recall from (7)  In the foregoing discussion it was shown that the distribution of branching rhymes, syllabification of vowel sequences before /@ 9 /, allophonic nasalization before /@ /, and the restriction of palatal sonorants to intervocalic position point to a geminate analysis for these sounds, which explains at the same time why stress may not shift to the antepenultimate syllable in words with a palatal sonorant onset in their final syllable.

Exceptional and regular stress in BP non-verbs
In BP, productive stress in non-verbs falls on the word-final syllable if it is heavy, and otherwise on the prefinal syllable: pomaD r 'orchard', sapaD to 'shoe'. Given this formulation of the stress rule, three classes of exceptions must be distinguished. In one class, stress is on the antepenultimate syllable, as in the words below:  (22) and (23) (22) and (23) represent exceptions to both approaches. The question that comes to mind is whether it is possible to show which characterization of the class in (24) is correct, the one that treats it as regular or the one that treats it as exceptional.
The productivity of phonological rules can be tested by observing how they apply to newly created words. To test the stress rules that apply to non-verbs one should only consider underived words, for which no accentual model is available and with regard to which the speaker needs to make a decision as to where main stress will fall. For BP there exist at least two sources for new words. One source is constituted by the very large class of acronyms, of which the following represent only a small sample: The distribution of stress in the words above, which all end in a closed syllable, shows that there is an unmarked stress rule for BP non-verbs, despite the relatively large class of exceptions in the existing vocabulary. Prefinal stress in acronyms that end in a heavy syllable is extremely rare 45   Foreign nouns do not always adapt automatically to the phonotactic requirements of the borrowing language. This is particularly true for geographical names, especially if the pronunciation in the language of origin is known to (some of) the speakers of the borrowing language. For some words it takes time to adapt, while others only adapt partially, for example with regard to their segmental properties only, but not for the accent position. The question why foreign words are adapted more easily in a given language than in another, even closely related, language is an interesting one, but still poorly understood 46 . The interest of the BP words in (26), therefore, arises from their confrontation with the corresponding Spanish words. Roca observes for Spanish that foreign proper nouns do not submit to the generalization that prefinal heavy syllables cannot be 'skipped' by the stress rule.
Interestingly, in BP not only do prefinal heavy syllables in these words resist the proparoxytonic pattern, but also a final heavy syllable is usually stressed. Clearly, Spanish and BP are very different in their treatment of foreign names and at least in the case of BP the general pattern of adaptation is consistent with a weight-sensitive stress algorithm 47 .

Syllable weight and stress in non-verbs
In section 2 we have addressed the arguments that have been given in the literature against the relevance of syllable weight in Portuguese and Spanish and we have shown that none of them can be backed up with either typological evidence or with evidence drawn from the grammar of BP. With regard to the latter, we have seen that words with a prefinal heavy syllable reject antepenultimate stress systematically. The behavior of words with a palatal sonorant onset in their final syllable was shown to be entirely consistent with this generalization, because of the geminate-like behavior of this consonant type. As regards pre-final syllables, heaviness in BP is defined straightforwardly as 'any branching rhyme'. This is equally true for the stressattracting nature of word-final heavy syllables 48 . Also, all and only heavy syllables condition a rule of mid-vowel neutralization in non-verbs. Furthermore, non-verbs that newly enter the language attract stress on a final heavy and, when their prefinal syllable is heavy, they strongly resist antepenultimate stress. It is equally relevant to recall that Frota and Vigário (2000) have shown that some rhythmic measures (%V) 49 group BP with mora-timed languages like Japanese, which reinforces the idea that weight is a functional notion in the phonological grammar of BP. We find it difficult to consider this array of facts as accidental and irrelevant from the point of view of the speaker's structural knowledge of his language's sound structure. We will therefore consider the weight-based patterns discussed in the previous sections as representing generalizations that are relevant from the point of view of the organization of the phonological grammar of BP.

An Analysis of Brazilian Portuguese Stress
We regard stress in BP as representing a mixed system, conditioned by the tense categories 'present', 'past', and 'future' in verbs, but prosody-based and quantity-sensitive in the non-verbal lexical categories.

Stress in verbs
In the verb forms below, representing the past tenses of BP at some intermediate level of representation, stress falls predictably on the vowel immediately following the root, or theme vowel (in (27) the stressed vowel is italicized and boldened). The different conjugations are illustrated with the verbs falar 'to speak' (a-theme, first conjugation), bater 'to beat' (e-theme, second conjugation), and partir 'to leave' (i-theme, third conjugation  (28) is surface-true, we assume, in agreement with the lexical-postlexical distinction we have adopted, that it belongs to the lexical phonology and applies to the sequences provided in (27). In these sequences, some generalizations regarding the phonological structure of BP verb forms are already accounted for. One is the fact that in the indicative imperfect the theme vowels of the second (-e-) and third (-i-) conjugation classes are neutralized to -i-. We thus need a statement in the lexical phonology that accounts for this fact (cf. also Câmara 1971:71). Next, consider the forms in (29) The sequences given in (29) all represent verb forms in which /a/ is raised to /e/ before /i/. We will see below that the same happens in some future forms. Moreover, /a/ is raised to /o/ before /u/ in the indicative perfect, the only verb form that contains this sequence. Some more adjustments are necessary to generate the correct output sequences of the past tenses.
The underlying forms of the first person perfect indicative /bat+e+i/ and /part+i+i/ have as their surface correspondents the forms [batí] and [partí], with stress on the final vowel. Since [ej], stressed as well as unstressed, is a possible surface sequence in BP, we need a very specific lexical operation to account for the height assimilation of /e/ in the 2conj.1si.perf.ind.
The following sequences represent the abstract forms of the BP future tenses. As in the past tenses, the location of primary stress in the future forms is completely transparent: it systematically falls on the first syllable following the theme, which is at the same time the syllable that initiates the sequence of segments representing the inflectional morphemes 53 .  3 We state the constraint that assigns stress to the past tense forms as in (31) In the verb forms of (33), we have already accounted for the fact that theme vowels are deleted before a following vowel-initial suffix, as in /fala+o/ > falo, /bate+o/ > bato, /parti+o/ > parto, and in all forms of the present subjunctive: /fala+e/ > fale, /bate+a/ > bata, /parti+ais/ > partais, etc 54 . Notice also that in the second person plural forms the raising of /a/ before /+i/ does not occur. We may suppose that this is due to a functional constraint, which blocks the raising of /a/ in order to preserve the underlying distinction between indicative and subjunctive. If this hypothesis is correct, we may maintain the generalization in (32) the way it is stated, and formulate an identity constraint for the present tense vocalism ranked above (32).
For the paradigm of the present tenses the main stress cannot be predicted by a single morphological conditioning, as is possible in the paradigms of the past and the future.
Although in most forms of the present stress falls on the last root vowel, in the 1 st and 2 nd person plural it is attributed to the vowel immediately following the root (or theme vowel), as in the past tenses. Compare the scheme (28), repeated here as (34a), with (34b) where Y is realized as the 1 st or 2 nd plural suffix: (34) a.
The elsewhere case for the present tenses is stated in (35): In the present tense, stress falls on the last vowel of the root Although the verb roots given in (33) are all monosyllabic, the generalization expressed in (35) applies to all verbs: cumprimént-o, -as, -a, -ámos, etc. of the verb cumprimentar 'to greet'. Taken together, (34) and (35) are surface-true.
Another way of predicting stress on the present tense forms is by a rule that places stress on the prefinal syllable. Such a rule would be surface-true at a level of representation before falling diphthongs are created. However, this level does not correspond with the output of the lexical phonology if we wish to express the generalization that, at least in non-verbs, stress cannot be on the fourth syllable counting from the right word edge, as would be the case for words like hermenêutico 'hermeneutic', nêurico 'neural', terapêutico 'therapeutic', and some others. In the next section, it will be argued that the three-syllable-window does constrain stress placement for non-verbs in the lexical phonology of BP. Furthermore, stress in the present tense would be assigned on the basis of a prosodic factor -the syllable countand thus be different from the other verb-stress rules, which locate stress by reference to some morphological boundary. To the extent that external evidence is available that could help us choose between the alternatives discussed, it points to the correctness of the more complex analysis proposed in (34b-35), which considers the present tenses as irregular (or nonuniform) form the point of view of stress distribution, instead of regular, as is implicit in the prosody-based analysis. There are several instances of paradigmatic (analogical) changes available in the history of Romance that relate to tense categories and for which the direction of levelling was decided by the categories (singular and third person plural) corresponding to the elsewhere case (35). These are mentioned in Hooper (1976:27, 30). For example, the survival of Latin stress in the indicative imperfect should have given the stress pattern of (34a) in Portuguese. We have seen above that in contemporary Portuguese stress is on the vowel following the root in all past tenses. The paradigm in (36b) corresponds to the Chicano Spanish present subjunctive, in which the last root vowel is stressed in all forms, different from Castilian Spanish. In Andalusian, the Chicano pattern is represented in the second and third conjugations only. In all of these cases a penultimate stress pattern has been abondened in favour of a pattern that is oriented towards the root boundary: first vowel following the root (36a) or last root vowel (36b). It is difficult to imagine the rationale of these changes, if the penultimate pattern represents a regular paradigm from the point of view of the language learner. We therefore keep to the generalizations (34)(35), closely following Hooper's (1976) analysis for Spanish.
In this section we have briefly discussed the verb-stress rules of BP. Stress is distributed in function of the tense (present, past, future) categories in a way that is surface true and, therefore, completely transparent. The different generalizations proposed above can be formulated in the following way 56 : (37) FOOT HEAD PRESENT: a. The rightmost syllable that is fully contained within the root must coincide with the head of a foot, in all present tense forms b. The syllable containing the first nucleus following the root must coincide with the head of a foot, in the 1 st and 2 nd person plural of the present tense Ranking: 37b >> 37a (38) FOOT HEAD PAST The syllable containing the first nucleus following the root must coincide with the head of a foot, in all past tense forms (39) FOOT HEAD FUTURE The first syllable of the future suffix must coincide with the head of a foot All constraints have a defined domain of application, and, the constraints (37a, b) aside, are unordered with respect to each other.
The constraints (37)(38)(39) only identify the location of the foot head and do not specify the foot form or the location of the primary stress foot inside the word. In the following section, where we discuss stress in non-verbs, we will return to this issue, as well as to the role of the three-syllable-window.

Stress in non-verbs
We have argued above that in BP the productive location of primary accent in nonverbs is determined by the structural properties (branching or non-branching) of the two rightmost syllable rhymes in the prosodic word. I will here adopt the definition of weightsensitivity proposed by van der Hulst (1994) according to which "weight-sensitivity arises whenever certain syllables (i.e. those that are 'heavy') refuse to occupy the dependent position in the foot" (1994: 5) 57 . The combination of the weight parameter with the right-edge oriented location of primary accent, and with the requirement that feet must be left-headed will predict main stress for the following words as indicated: For words that show a sequence of two light syllables at the right edge of the word, primary stress is on the prefinal syllable. The bi-syllabic foot that is constructed over the last two syllables respects the requirements of alignment with the right word edge and left-headedness: (44) is due to the specific definition of WEIGHT, which prohibits the occurrence of a heavy syllable in the weak position of the foot. (45) aberto WEIGHT ALIGN-FT-R, PRWD-R TROCHEE BINARITY (`) Y a(bérto) a(bér)to *! * (áber)to *! * a(bertó) *! * aber(tó) *! (áberto) *! * As is clear from (45), the grammar in (42) does not allow for proparoxitonic words, independently of the weight of the prefinal syllable (see tableau (46) also). A perhaps somewhat surprising aspect of tableau (45) is the preference given for the candidate /a(bérto)/ over /a(bér)to/, which goes counter to the tradition that in weight-sensitive systems the foot contains at most two moras. It also implies that feet may be unbalanced, an issue that is still being debated at the theoretical level. Notice that, in order to select [a(bér)to] as the preferred candidate, it would be enough to rank a bimoraicity requirement higher in the constraint hierarchy, as is done in most weight-based analyses of BP. We will assume here that the sequence of syllables constituted by the primary stressed syllable represents a prosodic unit together with all remaining syllables to its right within the prosodic word. We will maintain this position until urgent theoretical considerations or language-specific arguments derived from BP phonology suggest a different solution.
In (46) below, BINARITY (`) selects ga(véta) as the preferred form over other candidates that contain a right-aligned trochee, gave(tá) and (gáveta). In a grammar containing a high-ranked NON-FINALITY(`) requirement, the constraints that relate to foot type disregard the word-final syllable in the process of evaluating the wordprosodic structure, whereas all other constraints, including those that control the edgeorientation of the foot, take the whole sequence into account. As a result, a ternary foot is created at the right word edge. Given the high ranking of NON-FINALITY(`), the evaluation of the candidate set projected from /aberto/ is achieved in exactly the same way as it is for words that end in a heavy syllable in a grammar that has a low-ranked NON-FINALITY(`) constraint, as (48) shows. From this it follows that a constraint ranking as proposed in (48) will never allow antepenultimate stress in words with a pre-final heavy syllable.
In line with the discussion above, it seems natural to account for prefinal stress in words with a final heavy syllable by way of mora extrametricality, which makes the second mora of a word-final branching rhyme invisible for the sake of accent location, as is illustrated below for the words útil 'useful' and órfão 'orphan-masc': *! * As before, the grammar requires the extrametrical element to be incorporated in the primary stress foot. Proof for the prosodic integration of the final mora comes from Spondaic Lowering, a mora-counting constraint discussed above that bans upper mid-vowels from prefinal syllables followed by a heavy syllable ( open syllable, such as jacaré 'alligator'. The most straightforward way to obtain the word-final accent in this class is by way of the constraint IAMB, which requires feet to be right-headed. The high ranking of this constraint in the grammar (46) will result in a monomoraic (degenerate) foot at the right word-edge, a foot type that is independently necessary, given the existence in BP of subminimal words like chá 'tea ', pá 'shovel', cá 'here', já 'already', lá 'there', etc.: (50) jacare IAMB WEIGHT ALIGN-FT-R, PRWD-R TROCHEE BINARITY (e ) Y jaca(ré) * ja(caré) *! ja(cáre) *! *! (jácare) *! * The different accounts of exceptional types of stress illustrated in (48)(49)(50) must, of course, be limited to specific word sets. At the same time, we must exclude the possibility for the grammar to generate impossible primary accent positions, i.e. accents that do not respect the three-syllable window (TSW) requirement, typical for bounded, right-edge oriented stress systems. In BP, this possibility could only arise in the case of irregular accents, because the productive part of the grammar always generates stress on one of the last two syllables 60 . On the other hand, in a language like BP, but with a productive NON-FINALITY constraint, the grammar would always assign accent to the penultimate or antepenultimate syllable. In other words, the TSW is the result of constraint interaction and, ideally, should not itself acquire constraint status.
Moreover, a TSW constraint would be visible only in languages in which the effect of NON-FINALITY is limited to a marked class of lexical items, as in BP. It seems natural, therefore, to use NON-FINALITY also to account for the TSW effect in the case of exceptional stress. In order to limit the effect of NON-FINALITY to a restricted class of morphemes, we may adopt a proposal by Pater (2000;forthcoming 61 ), who argues that constraints should be allowed to be triggered by a specific lexical item or a lexically marked set of items. In a similar vein, we may adopt an IAMB Since the constraints in (51) are triggered by specific sets of lexical items, they will have no effect on the evaluation of candidates projected from the set of morphemes that are lexically unmarked. However, there is one problem with this way of accounting for irregular stress: how can it be avoided that lexical items be marked for triggering some other constraint, one that could generate an impossible accent? Above it was pointed out that the TSW effect on exceptions is typical for bounded, right-edge oriented stress systems. We suggest that the choice of the constraints that are to account for exceptional stress depends on the structure of the productive part of the prosodic (sub)grammar. For the derivation of exceptional stress in a bounded stress system such as that of BP, only NON-FINALITY is allowed to be triggered by marked lexical items and, consequently, be promoted to a high position in the grammar. In the same spirit, in a grammar that allows degenerate feet to account for stress on monosyllabic words, a morpheme-specific IAMB may account for exceptional word-final stress. By making the high ranking of some morpheme-specific constraints dependent on specific properties of the productive part of the grammar, the generation of prosodic structures by other morphemespecific constraints can be avoided, such as a high-ranked constraint that requires the exceptional alignment of a foot with the left word-edge, which should be excluded in a right- based accent that violates the TSW in mixed systems like that of BP, but we expect that such a system could exist. On the other hand, in line with our earlier discussion, we would not expect to find an exceptional verbal accent that does not respect the regular right-alignment of the main foot with the right word edge, to create, for example, an exceptional word class with root-initial accent in the present tense: exceptional cúmprimento 'I greet' next to productive convérso 'I talk'.

Conclusion
In this paper we have shown that all the arguments used in the literature to dismiss the weightsensitivity of BP primary word stress are not compelling when confronted with the facts of language typology and of BP phonology. As for the typological claims, there is no universal implication between between contrastive vowel length and weight-sensitive stress. Also, mixed systems of the kind defended here for BP are commonly found in the world's languages.
Primary stress is a lexical phenomenon in BP, which is distributed in function of the tense system in verb forms. In non-verbs, main stress is assigned on the basis of phonological properties, of which syllable weight is a determining factor. Palatal sonorant consonants function as phonological geminates, a property which explains their non-occurrence after branching rhymes, their behavior in syllabification, and, for the nasal sonorant, why the nasalization of the preceding vowel is obligatory (insensitive to stress). It is moreover correctly predicted that, when sonorant consonants appear in the onset of word-final syllables, their constraining effect on the possible position of primary stress is typical of prefinal heavy syllables in general and as such does not weaken, but strengthen the claim that the weight of the prefinal syllable is a relevant factor in the location of main stress. The rule of Spondaic Lowering was studied to show that the relevance of syllable weight is not restricted to the BP stress system. In particular, this process reveals the speaker's sensitivity to the bimoraic structure of the word-final syllable. We have then tested the productivity of the stress rules for non-verbs with regard to newly created words and loan words and observed that these rules are fully productive, despite the existence of relatively large classes of exceptions.
As was expressed in our formal account of BP main stress, our proposal differs from earlier accounts by the claim that all primary stress feet in BP, regular and exceptional, in verbs as well as in non-verbs, are left-headed and that all stress feet are aligned with the right word-edge. We have furthermore suggested that there exists a relation between the overall properties of the subgrammar that accounts for a language's foot structure and the mechanisms that may be used to account for exceptional stress patterns.
a specific page or to the quoted passage. In any case, to the best of our knowledge, this is the only place where Trubetzkoy makes a statement that is interpretable in the way it was understood by these scholars. 4 For further discussion of Italian, see also Principles (1969/71:201 phonological length distinction in vowels and consonants, rules that treat syllables containing long vowels as heavy should also treat syllables closed by (half of) a geminate consonant as heavy, but not necessarily simplex coda consonants. One language that appears to falsify both predictions is Sri Lanka Creole Portuguese, analysed by Smith (1978). Sri Lanka Creole Portuguese has both long and short consonants and long and short vowels. The language has a primary stress rule that applies lexically and only counts long vowels as heavy. Post-lexical secondary stress is distributed by a rule for which both closed syllables (including syllables closed by geminates) and syllables containing long vowels are counted as heavy.
10 This is exactly the hypothesis defended by De Chene and Anderson (1979). 11 Many thanks go to Rob Goedemans for putting the StressTyp database at our disposal.
12 Based on Kodzasov (1999), from which all the examples are taken. 13 Notice that, since both syllable weight and sonority of the vowel play a role in the Archi stress rule, this may indicate that the stress rule is prominence-sensitive rather than weightsensitive, in the sense of Hayes (1995:273 ]zínho suggests that the elimination of the stress of the base takes place in the postlexical component, where unstressed mid-vowel neutralization no longer applies. 22 As always happens in BP when the mid-vowel opposition is neutralized under stress. Cf. Wetzels (1992). 23 Roca (1986) provides convincing evidence that secondary stress in Spanish is assigned after primary stress, and in a separate process. See also Vigário (2003:119) for arguments in favour of the postlexical status of rhythmic stress and (prosodic word) initial stress in European Portuguese. 24 Ribeiro (1999) discusses the properties of monophthongization of [ej] in the dialect spoken in the city of Caxias, in the State of Maranhão, northeastern Brazil. It also contains a summary of the existing literature on the subject and a discussion of the relevance of the different linguistic and sociolinguistic variables that favor or disfavor the application of the rule in the different dialects for which the process has been studied. 25 If this explanation is correct, these diphthongs represent heavy rhymes, contrary to what is proposed in Bisol (1994). 26 See, however, Baker (2004) for an analysis of Spanish palatal sonorants very similar to the one proposed in Wetzels (1997) for Portuguese. 27 For more details and the distribution of these sounds in the different dialects, see Angenot and Vandresen (1981: 82); Cristófaro (1998: 51); Netto (2001: 103). 28 We will not address here the question of where in the grammar the distinguishing features of /R/ are spelled out for the different positions in which it occurs, which is not crucial for the problem at hand. 29 The proposal to analyze palatal sonorants as phonological geminates as well as the underlying argumentation was first presented at the European Science Foundation Euro-Type meeting in Lund 1993 and subsequently, both formally and informally, at several occasions thereafter, before its publication as Wetzels (1997). The analysis was inspired by Carreira (1988) who posits a bisegmental analysis for Spanish palatals that consists of a coronal consonant followed by the palatal glide /j/. This analysis, which is problematic for Spanish, is impossible for BP, where word-final sequences derived from /-lia, -nio/, etc. behave differently both phonetically and phonologically from /-r a,s o/ etc. See Baker (2004) for discussion of Carreira's proposal for Spanish. 30 Words that have a word-initial ##uV sequence or in which /u/ occurs intervocalically are almost exclusively loans from indigenous languages or African languages. Word-initially, the pronunciation of /u/ in these words shows the same variation as the one observed for /i/:  Cagliari (1981) does not state explicitly whether his transcriptions are based on laboratory evidence. If they are not, although we know Cagliari as a phonetician with a very good ear, it should be possible to confirm these impressionistic judgments by acoustic evidence.
Importantly, the examples given by Cagliari do not exclusively concern palatal glides preceded by a stressed nucleus. It is a well-known cross-linguistic phenomenon that onsets following stressed open syllables tend towards an ambisyllabic pronunciation. 32 See Baker for discussion of the palatal glide in Spanish: "We may therefore expect that it triggers the same high degree of coarticulatory F2 activity in preceding vowels as the complex segments and that the same lengthening effect should be present" (2004: 238). 41 At least, that is our interpretation of the data based on the retroflex pronunciation of /R/ in the syllable coda, which is typical for that region. 42 The fact that we have chosen Netto's examples of the western paulistano dialect does not imply that the reductions that occur in word-final unstressed syllables are limited to this particular region. In reality these changes are much more general. The reader will also recall that in most dialects of Brazil the vowel system occurring in word-final unstressed open syllables is represented by the 'basic triangle ' [i, u, a]. The raising of /e, o/ in the words of the third column as compared to their correspondents in the second column is the result of the loss of the syllable coda (or, in the case of the nasal diphthong, monophthongization). 43 In most words with stress on a final open syllable the prefinal syllable is also open. This, however, is a characteristic of the indigenous languages from which these words were taken, rather than a condition on this stress type. Words borrowed from African and European languages show that the prefinal syllable may be heavy in such words: candomblé 'religion (of African origin)', dendê 'African oil palm', carnê 'booklet'. 44 Which is, in itself, surprising: why are there no athematic nouns with stress on their prefinal syllable?
competence of the speakers of the 'lingua culta' and, when solicited, it is always produced with stress as indicated in the example sets. 51 Following Wetzels (1997), we assume that the 3pl suffix is lexically represented as a high nasal vowel, which is realized as a high coronal vowel after /e/, as in [báte g j] 'they beat ', etc. 52 The diphthong [íj] is postlexically reduced to [i]. Diphthongization itself must be a lexical constraint, because it interacts with the lexical process of Vowel Harmony (cf. Wetzels 1995). 53 We take -r to be the initial segment of the future morphemes -ra (future indicative) andria (conditional) in BP. 54 We follow Câmara's proposal for the underlying morphological composition of verb forms.
{e,a} + ... 55 In Chicano, as in most dialects of BP, the second person plural form is not used. 56 We will sidestep here the technical issue of how to account for this type of morphologically governed accent from a universalist (constraint) perspective. For some relevant discussion, see Alderete (1999). 57 Thanks to Harry van der Hulst for sending me his unpublished paper on foot typology in OT, which is listed in the bibliography as Van der Hulst (1994). 58 We assume that it is the relative unmarkedness of the singular vis-à-vis the plural that determines the singular as the base for the output-to-output correspondence relation established in this constraint. For a possible, more detailed, formalization, see Alderete (1999: 120). We furthermore suggest that a similar constraint accounts for the prosodic identity