A revised reconstruction of the Proto-Tupian vowel system O sistema vocálico do Proto-Tupi: uma nova proposta reconstrutiva

: This contribution is concerned with the reconstruction of the vowel qualities of Proto-Tupian, the ancestral language of the Tupian language family. The study is grounded in a bottom-up application of the comparative method and seeks to offer a more balanced reconstruction that avoids an overreliance on the Tupí-Guaraní branch. It is first shown that the height opposition traditionally reconstructed for the rounded vowel series ( *o vs. *u ) is best interpreted as an opposition between an unrounded vowel and a rounded one ( *ə vs. *o ). It is also argued that multiple instances of *e in the traditional reconstruction should be rather attributed to *ə . Finally, it is shown that two vowels (symbolized as *ɨ and *ɯ ) must be reconstructed in lieu of the traditional *ɨ . The resulting proposal has consequences for the subgrouping of the Tupian family.


INTRODUCTION
contrast and are thus represented as ɔ and o. The vowels of the extinct language Arikém, as well as of its direct ancestor Proto-Arikém, are represented as /i e ae ɒ ʉ/. The choice of the less common characters /ae ɒ ʉ/ is suggested by the variable representation of these vowels as ‹a› ~ ‹e›, ‹o› ~ ‹ḁ› ~ ‹a›, ‹u› ~ ‹u̥ › ~ ‹i› in our only sources on the language .
In most Tupian languages, consonants in the coda position do not contrast for features other than the place of articulation. This is true for Sateré-Mawé, Awetí, Proto-Tupí-Guaraní (and most contemporary Tupí-Guaraní languages), Proto-Tuparian (and all contemporary Tuparian languages), Proto-Arikém, and Puruborá. We write such codas in small caps: p (labial), t (dental/alveolar), c (palatal), k (velar). Their precise surface realizations vary depending on the language, on the point of articulation, and on the phonological environment, and include [p̚ m β] for p, [t̚ n ɾ] for t, [(ʲ)c̚ ɲ j] for c, [k̚ ŋ ɰ g] for k.
The phonetic inventory of many Tupian languages has two-phase stops that start out with a lower velum which raises during the occlusion. These are variably analyzed as postoralized allophones of underlying nasals or prenasalized allophones of underlying voiced stops (cf. Wetzels & Nevins, 2018), or (more rarely) as underlying prenasalized stops (cf. González, 2008). In this paper, such segments are always spelt as mb, nd, nʤ, ŋg, regardless of their phonological status in a given language.
Nasal vowels are always explicitly marked with a tilde, even if our sources leave nasality unmarked in some environments (usually following or preceding nasal consonants).
In languages where [s] and [ts] are free, idiolectal, dialectal, or chronological variants of one phoneme, we have normalized the data from our sources in order to warrant consistency across examples (e.g., only s is used in Tuparí, and only ts in Sakurabiat), following the current transcription practices in the most recent expert source in each case.
Proto-Mawé-Awetí-Tupí-Guaraní (= Proto-Mawetí-Guaraní = PMG) reconstructions are mostly those by , with the following modifications.  *tʲ is rewritten as *c, except when it follows an *i or a *j in their reconstruction, in which case we employ an ad hoc character *ć in order to capture the fact that the reflexes of *ć in all daughter languages are different from those of *c (this also allows us to give up the reconstruction of 'phantom' instances of PMG *i and *j, which are in  proposal hypothesized to have been lost in all daughter languages). For example,  reconstructions such as *itʲet 'his/her name' and *tʲajtʲu 'armadillo' are replaced with *ćet and *caću, given that no daughter language preserves any segmental trace of the alleged PMG segments *i and *j (Sateré-Mawé het, sahu; Awetí tet, tatu [pep]; PTG *tset, *tatu). We also posit PMG *tʲ alongside *c and *ć in order to account for the sound correspondence involving Sateré-Mawé t (in unstressed syllables) / ɾ(j) (in stressed syllables), Awetí ʐ, and PTG *t, found in the stem for 'fire' (reconstructed as *atia, *atja in atʲa in our proposal) as well as before *i (in complementary distribution with *t) 1 . We also diverge from  in reconstructing *w instead of their *kʷ, whereas the segment they reconstruct as *w is considered to have been independently epenthesized in the environment *u_V after the split of PMG (that way,  reconstructions such as *tʲuwaj 'tail' and *tʲuwɨ(k) 'blood' are replaced here with *cuac, *cuɨ > Sateré-Mawé suwac-po, suː; Awetí -uwac, -uwɨ[k]; PTG *tuwac/*-ruwac, *tuwɨ/*-ruwɨ) 2 . We accept Schleicher's (1998, pp. 18-24) suggestion, reinforced by Meira and Drude (2015, pp. 278-279), whereby only one PTG affricate is reconstructed instead of the traditional *c (*ts) and *č (*ʧ); we symbolize it as *ts 3 .
Proto-Mundurukú reconstructions are taken from Picanço (2019). For relational stems, whose leftmost consonant often alternates depending on the left context, Picanço (2019) lists all possible allomorphs. In this article, only the allomorphs with *p-, *ð-, and *ʧ-(rather than *b-, *t-, *ɟ-) are given for such stems. We also omit the hyphen, used by Picanço (2019) in order to indicate that the relational stems are bound.
For Proto-Juruna and Proto-Tuparian, a variety of proposals exist; the reconstructions in this article are mostly extracted from the more recent ones (Carvalho, 2019 for Proto-Juruna;Nikulin & Andrade, 2020 for Proto-Tuparian) or adapted from earlier proposals  for Proto-Juruna;  for Proto-Tuparian) so as to match the phonological reconstruction of the most recent works.
The reconstruction of the Proto-Tupian consonants adopted here differs from previous proposals and is based on the correspondence sets summarized in Appendix 2. For reasons of space, it is impossible to discuss this problem in detail in this contribution. Note, however, that nothing in our reconstruction of the Proto-Tupian vowels hinges on our interpretation of the PT consonants, and the validity of our proposal would remain intact if one adopted a different interpretation (such as that of Rodrigues, 2007).
The acute accent symbolizes the high tone in tonal languages (including the languages of the Mundurukú, Juruna, and Mondé branches, Karo, and maybe also Makurap); the low tone is left unmarked. Tones other than high and low are found only in the Mondé languages, and the transcription of our sources is retained in such cases. In the Tuparian languages Tuparí and Akuntsú, contrastive stress has been described, which is also symbolized by means of an acute accent (when its position is known).
Data quoted from premodern sources, which are not expected to faithfully represent all the relevant phonological oppositions, are given 'verbatim' enclosed in chevrons. Subscript letters after such forms indicate the ultimate source of the data: ‹› S and ‹› L refer to Emilie Snethlage's and Lopes' data on Kuruaya (Snethlage, 1932); ‹› B and ‹› N refer to Barbosa's and Nimuendajú's data on Arikém . Forms followed by ‹›kg come from Koch-Grünberg (1932) on Puruborá, and those with ‹›es from Snethlage (1934), again on Puruborá. Forms in chevrons without subscript letters are from Steinen (1886) for Manitsawá, Nimuendajú (1923Nimuendajú ( -1924 for Xipaya, and Sekelj (1948) for Aruá and Makurap.
Much of the discussion in this paper is based on analyzing cognate sets. In some cases, a given form is not synchronically segmentable, but only a part of it is cognate with the material of other languages. The part which is deemed non-cognate is then given in brackets. In premodern attestations (enclosed in chevrons), the cognate part is given in boldface. 2 The epenthetic nature of the w in these stems is likewise confirmed by the fact that no corresponding consonant is found in branches such as Mundurukú (*ðoj 'blood', *t-oaj-bɨ 'tail'; see Picanço, 2019) or Tuparian (*jeɨ 'blood', *joac 'tail'; see . Our amendment to  reconstruction spares us from the necessity of positing a typologically improbable 'zigzag' development in Sateré-Mawé, whereby PT *w > PMG *kʷ > Sateré-Mawé w (h before u in stressed syllables; ∅ before u in unstressed syllables). In our account, Sateré-Mawé w (h/∅ before u) simply continues PT *w > PMG *w. Only the ancestral language of Awetí and PTG would thus have innovated by transforming PT *w > PMG *w into a stop. 3 The diverging reflexes in the Guaraní varieties that were earlier seen as warranting the reconstruction of two affricates *c and *č for PTG are now explained as late developments involving diffusion or dialect borrowing among Guaraní dialects.

EARLIER SCHOLARSHIP
The vocalic system of Proto-Tupian has been reconstructed by Rodrigues and Dietrich (1997, p. 268) and Rodrigues (1999Rodrigues ( , p. 110, 2005 as comprising six vowel qualities (*a, *ɨ, *o, *u, *e, *i), each of which would also have a nasal counterpart (*ã, *ɨ, *õ, *ũ, *ẽ, *ĩ; see Rodrigues & Cabral, 2012, p. 502). In Table 1, we list the reflexes of the Proto-Tupian oral vowels according to the proposal by . It can be easily seen from the table above that, according to , the evolution of Proto-Tupian *e and *ɨ in the constituent families involved some phoneme splits. In the case of PT *e,  posits a split in the so called 'Eastern' Tupian languages (Mawé-Awetí-Tupí-Guaraní, Mundurukú, and Juruna) allegedly conditioned by an adjacent consonant. The examples 1-2 illustrate the default development of PT *e, whereas 3 instantiates the development of PT *e affected, according to Rodrigues' (2005, p. 40) proposal, by a following labialized stop (the labialization is reconstructed here exclusively in order to account for what are thought to be the divergent reflexes of PT *e) 4 .
(1) PT *kʲet 'to sleep' (Rodrigues, 2005, p PT *epʷ 'leaf' (Rodrigues, 2005, p. 40;Rodrigues & Cabral, 2012, p. 505 Rodrigues (2005, pp. 40-41) posits a split in Juruna and in the so called 'Western' Tupian languages of the Arikém, Tuparian, Karo (Ramarama), and Puruborá groups. This time, however, he does not identify a phonological environment which could have conditioned the alleged split (beyond a generic reference to the 'immediate consonantal context'), nor is he explicit about whether the alleged split proceeded in the same way in all the aforementioned languages. The examples 4-6 illustrate.
(4) PT *kˀɨp 'tree, wood' (Rodrigues, 2005, p. 41;Rodrigues & Cabral, 2012, p. 506 In the subsequent sections, we will argue against the proposal by , suggesting instead that the observed sound correspondences are best accounted for by reconstructing a phonemic inventory of seven (rather than six) vowel qualities for Proto-Tupian and positing a number of mergers in the daughter languages, in addition to one conditioned split. That way, the examples in 1-6 are reconstructed in our proposal as *kʲet 'to sleep', *mẽt 'husband', *jəp 'leaf', *ḳɯp 'tree; stick-like', *mbɨ/*pɨ 'foot', *pətɨc 'heavy'. Tables 2 and 3 show the oral vowel inventories of Proto-Tupian in  and our proposals, respectively.

PT *ə
This section deals with the reconstruction of a vowel we chose to represent as *ə. We start by stating its proposed reflexes in the daughter branches and listing the relevant cognate sets. In subsequent sections, we discuss how our findings relate to  reconstruction of the PT vowels and consonants. We conclude that the recognition of *ə as a contrastive unit allows reducing the phonological inventory of Proto-Tupian by three phonemes (*pʷ, *kʷ, *kˀʷ), to account for the sound correspondences in a number of cognate sets which are unexplainable in  proposal, and to account for the limited distribution of the sound correspondence which underlies  reconstruction of PT *o (which occurs exclusively following labial consonants). Our proposal also entails that the vowels traditionally reconstructed as *o and *u should be reinterpreted as PT *ə, *o.

PROPOSAL
The vowel we reconstruct as PT *ə has evolved in the following way in the daughter languages 5 . In the Mawé-Guaraní branch, it has acquired rounding and changed to PMG *o (in fact, in our proposal PT *ə is the only source of PMG *o). In addition, it has been raised to *u before a vowel (as in 'blood', 'sun') or before a glottal stop and a vowel (as in 'arrow').

ADVANTAGES WITH RESPECT TO RODRIGUES (2005)
In what follows, we discuss four correspondence sets derived by  from three Proto-Tupian vowels. Two of them, identified by  with PT *e and *u, show no overlap at all; these correspondence sets appear as (a) and (d), respectively, in Table 4 below. The remaining two correspondences, given as (b) and (c) in Table 4, are attributed in our account to PT *ə, yet an entirely different account is proposed by . The correspondence (c) has the same reflexes as (d) in Karitiana and Tuparí, but other languages show distinct reflexes, which are typically lower than those of (d); it is associated by  with PT *o. The correspondence (b) shows significant overlaps with (a) and (c): in Tupí-Guaraní, Awetí, Sateré-Mawé, Mundurukú, and Yudjá, the observed reflexes are identical to those of (c), whereas in all other branches the correspondence set in question -in Rodrigues' (2005) account -coincides completely with (a).
According to Rodrigues ( , pp. 40, 42, 2007, the overlapping pattern that involves the correspondence (b) -which coincides with (c) in Tupí-Guaraní, Awetí, Sateré-Mawé, Mundurukú, and Yudjá, but with (a) in the remaining branches -can be explained by reconstructing a secondary rounding feature for the consonant that immediately follows the vowel (the available options in  reconstruction are *pʷ, *kʷ, *kʷˀ). In the proto-language of Mawé-Guaraní, Mundurukú, and Yudjá ('Eastern' Tupian languages in Rodrigues', 2005 terms), this contextual factor would have induced the merger of the correspondence set in (b) with the *o series, Yudjá being later subject to *o > a and Mundurukú undergoing *o > ə. The remaining branches -that is, Arikém, Tuparian, Mondé, Ramarama, and Puruborá -would have not been subject to any contextual coloring and show reflexes identical to those of PT *e, which leads Rodrigues ( , 2007 to reconstruct *e for the correspondence set in question and to posit a conditioned split in his 'Eastern' languages. In synthesis, then,  proposal for the PT segments whose reflexes appear in the correspondences in Table 4 above is as follows: *e for the correspondences (a) and (b), *o for (c), and *u for (d). Moreover, the context-dependent merger of the series (b) and (c) in Mawé-Guaraní, Mundurukú, and Juruna is attributed to the influence of a secondary rounding feature hosted on the following consonant.
Under closer scrutiny, however, it appears that the available evidence does not support either the identification of the correspondence (b) as a context-dependent offshoot of (a) or the reconstruction of a labialized stop series for Proto-Tupian. Although we concur with  in reconstructing PT *e for the correspondence (a), we disagree with his diachronic interpretation of the remaining three correspondences in that: − we consider (b) to be the default development of PT *ə (rather than a positional development of *e); − we consider (c) to be a positional development of PT *ə (rather than the default reflex of a PT phoneme of its own, symbolized as *o by Rodrigues (2005); − we derive (d) from PT *o, which in our account is the only rounded vowel of PT (as opposed to  reconstruction, whereby PT had both *u and *o).  account is seriously undermined by the following facts. 1. First of all, the consonants reconstructed as labialized by  in etyma that instantiate the correspondence set (b) appear not to have reflexes distinct from those of non-labialized consonants. 2. Furthermore, the correspondence in (b) may also occur morpheme-finally or morpheme-internally before vowels, making it impossible to attribute the emergence of the correspondence to a following consonantal segment. 3. The correspondence set in (c) may be explained away as a conditioned offshoot of (b). 4. The reflexes listed by  for (b) and (c) in Suruí-Paiter, Karo, and Puruborá are partially based on non-cognate material and are thus incorrect. 5. Finally, there is typological evidence that renders  hypothesis implausible. Each of these five points is discussed in the subsequent sections.

PURPORTED LABIALIZED CONSONANTS HAVE THE SAME REFLEXES AS PLAIN CONSONANTS
Let us consider the reflexes of the PT segments that Rodrigues (2007) reconstructs as labialized consonants. As will become clear, their reflexes do not differ from those of their plain (non-labialized) counterparts, and the only reason for positing such phonemes in Rodrigues (2007) reconstruction is to account for the correspondence set (b). Once it is recognized that the (b) series does not result from a conditioned split of *e, it is no longer necessary to reconstruct labialized consonants for Proto-Tupian.
We will start by examining the occurrences of *pʷ that are supposed to account for the alleged rounding of PT *e in 'Eastern' Tupian. Rodrigues (2007) reconstructs it for two roots, *epʷ 'leaf' and *epʷa 'face' (as well as in its derivative *epʷa-pokˀ 'to appear'). In the former case, all Tupian branches have a reflex with a plain labial stop p (in some languages, which lack an opposition between oral and nasal codas, it is symbolized as p). In most Tupí-Guaraní languages, as well as in the Arikém language before the suffix -ɒ, the stop is lenited to β or a similar sound (cf. Schleicher, 1998, pp. 29-32;Storto & Baldi, 1994); this development regularly targets word-final stops of any origin in these languages 8 . In our reconstruction, the vowel correspondence between PMG *o, PMu *ɨ, PTpr *e, Kt a, Ari ae, Pu ə, and Mo e is derived from PT *ə, whereas the correspondence between the wordfinal consonants straightforwardly continues PT *-p. That way, there is no need to reconstruct a labialized stop for 'leaf' in our proposal. (34) PT *jəp 'leaf' (Rodrigues (2007) An identical rhyme is found in the word for 'bitter'. Rodrigues (2007, p. 196) reconstructs its PT etymon as *rʲop and lists its reflexes in Sateré-Mawé, Awetí, PTG, and Mundurukú. Had he considered the Tuparí and Karitiana cognates, he would have likely reconstructed *rʲepʷ. (Rodrigues, 2007: *rʲop) In Rodrigues' (2007) account, *pʷ is claimed to have a divergent reflex between vowels in the languages of the Mawé-Guaraní (PT *pʷ > PMG *β, as opposed to PT *p > PMG *p) and Mundurukú (PT *pʷ > PMG *p, as opposed to PT *p > PMG *b) branches. Rodrigues (2007) gives only two cognate sets that instantiate the sequence *epʷ: PT *epʷa 'face' and *epʷapokˀ 'to appear'; Corrêa da Silva (2010, p. 128) claims the latter to be a derivative of the former. Rodrigues (2007, p. 186) lists the following reconstructions and reflexes (quoted verbatim). (36) PT *epʷa 'face' (Rodrigues', 2007 reconstruction PT *epʷapokˀ 'to appear' (Rodrigues', 2007 reconstruction) > PTG *oβapo || Mu ǰ-ebapək || Tu epapok 'to arrive' Note, however, that even within Rodrigues' own framework the proposed etymology for 'face' presents serious irregularities. In his account, PT *e before a labialized consonant would be expected to yield Mw o (rather than e), Mu ə (rather than o), Ku ɨ (rather than o -note that u in the datum cited by Rodrigues (2007) is a phonetic variant of /o/), and Kt a (rather than ɨ). We surmise that in this case Rodrigues (2007) has failed to distinguish between two unrelated cognate sets, which we derive from PT *jəβa 'forehead' and *jopʔa 'face' (in addition, Mw -ewa, or sewa in our notation, appears to be unrelated to either etymon). (38) PT *jəβa 'forehead' (Rodrigues, 2007: *epʷa (Rodrigues, 2007: *epʷa) PT *β in *jəβa 'forehead' is reconstructed based on the correspondence PMG *β ~ PTpr *β, otherwise found in the cognate set for 'wind', PMG *ɨβɨću ~ PTpr *ɨβijo (cf. Nikulin & Andrade, 2020, p. 292). In *jopʔa, the consonant cluster is reconstructed in order to account for the voiceless intervocalic stop *p in PMu, also found in etymologies such as PMu *óropo ~ PTpr *oropʔo ~ PMG *uruβu 'vulture' 9 . In our account, the correspondence between PMG *o and PTpr *e need not be conditioned by any feature hosted on the following consonant.
As for the cognate set for 'to appear', we believe that PTG *oβapo should be excluded from it (no other examples are known where a velar coda in Mundurukú or Tuparian would correspond to zero in PTG) 10 . Moreover, the Mundurukú cognate does not actually contain an initial vowel (the root is papəḱ 'to be visible', with the allomorph bapəḱ occuring after vowels; cf, Picanço, 2005, p. 17). That way, this cognate set does not instantiate the vowel correspondence which interests us in this section, nor does it back up the reconstruction of *pʷ.
The only example which, according to Rodrigues (2007, p. 182;2008, p. 6), instantiates a labialized reflex of PT *kʷ in a Tupian language is PT *ekʷ-at 'plaza' > Xi koað-á, Sk ekʷat, Tu ekoat-pe 'area around the house' 11 . We reconstruct the etymon in question as PT *ək-at and reject the appurtenance of the cited words to this cognate set. First of all, the ultimate sources of the Xipaya and Sakurabiat words provide glosses which are quite distant from '(village) plaza': ‹ku̥ aẓá› 'village of foreigners' (Nimuendajú, 1928, p. 827), ‹hekʷat› 'field' (Hanke et al., 1958, p. 205). Second, Sakurabiat kʷ does not regularly correspond to Tuparí ko (cf. . The third purported labialized stop of Proto-Tupian is reconstructed by Rodrigues (2007) as *kʷˀ for one single etymon. (47) PT *əḳɯp/*jəḳɯp 'arrow' (Rodrigues, 2007: *ekʷˀɨp) Rodrigues (2007, p. 186) states explicitly that *kʷˀ is reflected in the daughter languages precisely in the same way as *kˀ (we prefer to symbolize the segment in question with the ad hoc character *ḳ), and the labialization in the etymology for 'arrow' is reconstructed only in order to account for the correspondence between a front vowel in the Tuparian languages and rounded vowels in the 'Eastern' Tupian languages. Also note that the PT stem for 'arrow' almost certainly contains the formative for tree-or stick-like objects *-ḳɯp (*-kˀɨp in Rodrigues', 2007 reconstruction), with reflexes in all Tupian languages, none of which shows any trace of labialization. In this sense, our proposal is superior to Rodrigues' (2007) in that no need arises to reconstruct an extra consonant found in only one stem.

REFLEXES OF (B) AND (C) IN KARO, PURUBORÁ, AND MONDÉ
According to , the correspondence sets (b) and (c) have different reflexes not only in Tuparí and Karitiana (as we have shown in the previous section, they are in fact in a complementary distribution in these languages), but also in Karo, Puruborá, and Mondé (represented by Suruí-Paiter in Rodrigues', 2005 study). The (b) series (derived from PT *e before a labialized consonant in Rodrigues', 2005 proposal) is supposed to be reflected as e in Karo, Puruborá, and Suruí-Paiter, whereas the correspondence set (c) (< PT *o according to  is expected to yield a in the three languages. In reality, however, the Karo and Puruborá reflexes listed by Rodrigues (2005, pp. 39-40) for the correspondence sets (b) and (c)  The reflex e, listed by  for Karo and Puruborá, simply does not occur in the available data. In Karo, one finds o, ɨ, i, ə, ə̃, a, and u (with no obvious distribution), and in Puruborá, ə is found in all cases. Currently we have no explanation for the reflexes in Karo (but note that Rodrigues' proposal also fails to account for them).
In the Mondé languages, one does indeed find a and e in accordance with Rodrigues' (2005) predictions (PT *e > e; PT *o > a). However, it is also possible to account for the Mondé reflexes if one recognizes that the etyma of all the aforementioned cognate sets contained one and the same vowel, PT *ə. Note that all instances of e occur following a coronal consonant ('leaf', 'larva') or word-initially ('house/village'), whereas all instances of a ('hand', 'snake', 'cultivated field', 'sun', 'heavy') occur following a peripheral (labial or velar) consonant. We propose, therefore, that PT *ə was fronted to Proto-Mondé *e following coronal consonants or word-initially and yielded Proto-Mondé *a elsewhere, and that the Mondé languages lend no support to the Proto-Tupian age of the distinction between the correspondence sets (b) and (c). We parenthetically note that the fronting of the type *ə > e following coronal consonants is also known from the history of Djeoromitxí, a Macro-Jê language of the Jabutian branch (Voort, 2007, p. 147), which is, like the Mondé languages, spoken in the Rondonian East. Typologically, the functioning of labial and velar consonants as a natural class in processes triggering vowel backing (as in *ə > a) is amply documented (Hyman, 1973;Vago, 1976) 13 .

GENERAL PHONETIC CONSIDERATIONS
The development PT *eCʷ > *oC, posited by , inter alia for his 'Eastern' Tupian languages, conjoined with the reconstruction of such consonants with labial off-glides -the factor that accounts for these environmentally restricted vocalic outcomes -raise two issues of phonetic plausibility of historical reconstructions: one related to directionality considerations of the presumed coloring effect of the PT consonants, the other related to the distribution of the labialized consonants. First of all, a baffling aspect of the labializing effect exerted by these consonants is that it always affects the preceding, not the following vowel: a pre-vocalic labialized stop has no effect on a following vowel. From a phonetic point of view this is extremely counterintuitive. If labialization, or, more precisely, a labial release feature, is to play the role of contrastive feature distinguishing between these consonants (i.e. *pʷ and *kʷ) and their plain counterparts (i.e. *p and *k), one would expect its 'coloring' influence upon adjacent vowels to be realized more strongly (if not exclusively) on a following rather than a preceding vowel (that is, in a C w -to-V transition, as opposed to the V-to-C w boundary).
As to their distribution, Rodrigues' (2007) PT labialized stops tend to occur, or are found quite frequently, in word-final position. In fact, the most significant phonotactic gap in their distribution in Rodrigues' (2007) proposal is the absence of *pʷ from word-initial position. The expectation, commented on above, that a consonant with a secondary articulation will exert a stronger coarticulatory effect on a following rather than a preceding vowel derives from the fact that such segments depend, for their realization, on a following resonant element. As a consequence, we also expect such consonants to be less-optimally realized (qua contrastive segments) in word-final or pre-consonantal position, with no vocoid to work as a base for its contrastive release features to be imposed. As a matter of fact, plenty of evidence suggests that this is the case (see, e.g., Blevins, 2004, p. 116). In the words of Ladefoged and Maddieson (1996, p. 357): Thus we can say that labialization is typically concentrated on the release phase of the primary articulation it accompanies. This observation has both phonetic and phonological significance. Many more languages have a restriction between the presence of labialization and the choice of the following vowel, than between its presence and the choice of the preceding vowel, and in many languages with labialized consonants the set of syllable-final consonants, if any, does not include labialized ones.
Aside from general considerations stemming from principles of acoustics and perceptual phonetics, it is not hard to find cross-linguistic evidence supporting the contention that such secondary articulations found in stop consonants behave phonologically as if 'looking for' a supporting vowel. Thus, in Khwarshi, an Eastern Caucasian language, labialization is found as a secondary articulation feature, mostly in velar and uvular consonants (Khalilova, 2009, pp. 17-18). The contrast is restricted basically to word-initial and word-medial position preceding a vowel, as in etʷa 'fly' vs. eta 'touch', lakʷa 'see' vs. laka 'lick'. The dynamic phonology of the language also demonstrates a preference for such labialized release 13 In principle, it is still possible that the regular Mondé reflex of PT *ə is a even after t/d. This would allow us to propose two new Tupian etymologies for Mondé roots at the expense of the etymology for 'larva' shown above: PT *tək 'to pound, to grind' > Pa -tagá 'to smash' (as in ɬo-dagá 'to pound'), Gv tágá 'to beat'; PT *ðəp 'bitter' > Pa [pe]ʧáp, Ar ‹petab›, Gv [pe]tɨɨp (note that Gv ɨ is usually derived from /a/ in diminutives). We thank an anonymous reviewer for bringing the Paiter form [pe]ʧáp 'bitter' to our attention.
consonants to occur preceding a vowel. Labialization is either lost (89a) or transferred to another consonant, one that precedes a vowel (89b), whenever a -C(V) suffix is added to a root containing a final labialized stop.

IV-see-inf IV-see-prs IV-see-inf IV-see-caus-inf
The joint effect of these generalizations, both static phonotactic patterns and processes in the dynamic phonology of Khwarshi, is to suggest that having consonants with a labial release preceding something other than a vowel is a highly undesirable or marked configuration in this language. Similar regularities are found in the phonologies of many unrelated languages, and can be understood more broadly in terms of the acoustic and phonetic constraints mentioned above, the same that make Rodrigues' (2007) proposal of stop consonants with secondary offglides that are almost always realized in contexts other than that of a following vowel very implausible.

INTERIM SUMMARY
Above we have presented evidence against positing a sound change whereby PT *e would have acquired rounding (> *o) preceding labialized consonants as well as against reconstructing these labialized consonants for PT. Instead, we have proposed that the suspect correspondence should be derived from PT *ə. Moreover, in our reconstruction, PT *ə accounts for some sound correspondences deemed irregular in  proposal as well as for the correspondences which underlie his reconstruction of *o. Our proposal is summarized in Table 5, where we list the reflexes of PT *ə as well of two other vowels which do not present split reflexes: PT *e and *o (in Rodrigues', , 2007 interpretation, *e and *u). Table 5. PT *ə, *e, and *o and their reflexes. A = before PT *ɨ or *ɯ in the next syllable; 14 B = before a vowel; C = next to a labial in a stem-final syllable; D = after a labial; E = after a coronal. The notational change in the reconstruction of the sole rounded vowel of Proto-Tupian (*o) as opposed to (*u) does not affect the correctness of the sound correspondences identified by  in any way. It is suggested by the fact that the typical realization of its reflex is a mid vowel in most Tupian branches (Mundurukú, Tuparian, Karo, Puruborá, and Mondé). Since Rodrigues' (2005) *o is reinterpreted as an unrounded vowel *ə in our account, it is now unproblematic to reconstruct *o in PT stems such as *amẽko 'jaguar', *jacjo 'armadillo', *jaḳo 'lizard', *jeko 'monkey', *jõk 'flea', *jopi(-ʔa) 'egg', *ḳo 'to ingest', *ndo 'hill, rock', *ndok 'to eat (intr.)', *õp 'to give', *õt 'I' (and the first person prefix *o-), *toḳo 'to bite', *top 'to see', *waco 'alligator', among many others (we do not list these well-established etymologies in Appendix 1 for reasons of space).
PT *ɯ VS. *ɨ In this section, we will argue that it is necessary to reconstruct two distinct vowel phonemes in place of  Proto-Tupian *ɨ. According to , Proto-Tupian *ɨ would have undergone a number of splits in the daughter languages depending on the immediate consonantal environment, yielding ɨ/i in Yudjá, e/i in Karitiana, i/ʉ in Tuparí, i/ɨ in Karo, and ɨ/i in Puruborá. Unfortunately,  does not specify the consonantal environments which would have triggered the putative fronting of *ɨ in the daughter languages, nor is he explicit on whether these environments were identical for each constituent branch (Juruna, Arikém, Tuparian, Karo, and Puruborá). In what follows, we show that Rodrigues' (2005) reconstruction collapsed two correspondence sets into one and that two distinct vowels must therefore be reconstructed for Proto-Tupian. We symbolize them as PT *ɯ and *ɨ 15 , as in the minimal pair PT *jɯ 'liquid' vs. PT *jɨ 'urine', still retained in Karitiana as se 'liquid' and si 'urine'. Their reflexes are identical in some branches (PMG *ɨ, PMu *i, PJu *ɨ, Mo i); for this reason, in what follows we are concerned only with the remaining branches (that is, Tuparian, Arikém, Karo, and Puruborá). At the end of the section, however, we will see that the distinction between the PT vowels in question is indirectly preserved in the Mundurukú branch as well.
PT *ɯ is reconstructed for the correspondence set which involves the following reflexes: PTpr *ɨ (> Ma/Sk/ Ak ɨ, Wy/Tu ʉ), Kt/Ari e, Pu ɨ. In Karo, one usually finds i in the word-initial position (i-cɨ 'water', itɨ 'deer'), ə̃ if the syllable is nasal (nəp 'louse', wakə̃ja 'agouti'), and ɨ elsewhere (i-cɨ 'water', ma-ʔɨp 'tree', tɨt 'to cook', itɨ 'deer', jaɨ 'howler monkey'); we are as of now unable to account for the apparently aberrant reflex pək 'to burn'. In Karo ju 'blood', the vowel u continues the PT sequence *əɯ and is thus not necessarily irregular. Below we list the PT etyma where evidence from multiple branches converges to the reconstruction of PT *ɯ as opposed to PT *ɨ (reflexes in branches which do not distinguish between them are omitted).
The following example can be considered regular if it turns out that Lemos Barbosa (1951) transcription ‹mixon› B of the Arikém cognate (see Rondon & Faria, 1948, p. 199) stands for mʉ̃sɒ̃ (for the development PT *w > Kt/Ari m in nasal environments, compare Kt mĩɲõ 'Brazil nut').
PT *ɨ is reconstructed for the correspondence set which involves the following reflexes: PTpr *i (preserved as i in all daughter languages), Kt/Ari i, Pu i. Below we list the PT etyma where evidence from multiple branches converges to the reconstruction of PT *ɨ as opposed to PT *ɯ (reflexes in branches which do not distinguish between them are omitted).

IMPLICATIONS FOR THE SUBGROUPING OF TUPIAN
The following groups share more than one innovation related to the evolution of PT *ə and *ɨ: Arikém and Tuparí (3 innovations), Mundurukú, Juruna, Sateré-Mawé, Awetí, and Tupí-Guaraní (2 innovations), and Mundurukú and Mondé (2 innovations). Of these, the former two sets are strong candidates for valid clades: both include non-trivial, 20 PT *mbɨʔa/*pɨʔa 'liver' > PMu *pia̰ is not necessarily an exception, as in this case one might suspect that the PT absolute form with *mb-was generalized in PMu. positionally conditioned innovations (merger of *ə and *o following labials in Tuparí and Arikém, preceding vowels in Mundurukú, Juruna, Sateré-Mawé, Awetí, and Tupí-Guaraní). In contrast, the set comprising Mondé and Mundurukú is in all likelihood spurious (or paraphyletic): in addition to being incompatible with the proposal which links Mundurukú to Juruna and Mawé-Guaraní, there is indirect evidence which suggests that the fronting of PT *ɯ in Mundurukú counterfed the loss of *p before front vowels, an innovation specific to that branch. Therefore, the triple merger of PT *i, *ɨ, and *ɯ as *i has probably occurred independently in the phonological history of Mondé and Mundurukú. That way, evidence from the development of the PT vowels supports the identification of two mid-level clades within Tupian. The node comprising Tuparí and Arikém is defined by the sound change *ə > *e (default) / *o (after labials); we suggest the label Tuparikém for this subgrouping hypothesis. 21 The vowel inventory of Proto-Tuparikém (*/i ɨ e a o/) is preserved without changes in Proto-Tuparian, whereas in Proto-Arikém these vowels yielded /i e ae ɒ ʉ/ (> Karitiana /i e a o ɨ/) by means of a vowel shift identified by Storto and Baldi (1994). The second clade includes Mundurukú, Juruna, and Mawé-Guaraní (that way, our findings partially corroborate  hypothesis regarding the validity of his Eastern branch) and has the merger of PT *ə and *o before vowels as its defining innovation. It may have proceeded in two stages: first, PT *ə and *o may have changed into *o and *u in a chain shift (this is precisely the state reconstructed by  for Proto-Tupian); in turn, the vowel *o (from PT *ə) may have been raised to *u in prevocalic contexts (thus merging with *u from PT *o). After that, Proto-Eastern Tupian *o, *u yielded PMG *o, *u; PMu *ɨ, *o; PJu *a (*u next to labials in stem-final syllables), *u. As the Eastern Tupian languages reach their greatest diversity between the Lower Madeira and the Lower Iriri, the Proto-Eastern Tupian Urheimat has to be sought in that region.

CONCLUSION
This paper has presented a reconstruction of the Proto-Tupi (PT) inventory of oral vowels alternative to that advanced by , this being clearly the accepted view on the PT vocalism since its adoption in reference works on the family such as Rodrigues (1999) and Rodrigues and Cabral (2012). Our proposal is summarized in Table 6. Table 6. PT vowels and their reflexes (proposal). A = before PT *ɨ or *ɯ in the next syllable; B = before a vowel; C = next to a labial in a stem-final syllable; D = after a labial; E = after a coronal. It is interesting to note that the defining innovation of this branch did not affect Kepkiriwat, an extinct language of Rondônia sometimes classified as Tuparian (cf. Hanke et al., 1958, p. 188;Rodrigues, 1999, p. 109;Galucio, 2001, pp. 5-6;Aragon, , pp. 6, 10-11, 2014Rodrigues & Cabral, 2012, p. 497, inter alia). The default reflex of PT *ə in Kepkiriwat appears to be o rather than e, as in ‹uóque› R , ‹uóc› B 'house, village', ‹óp› B 'leaf', ‹gó› B 'cultivated field' (Rondon & Faria, 1948, pp. 181, 187, 191) < PT *ək 'house' or maybe the first person form *o-jək 'my house', *jəp 'leaf', *ŋgə 'cultivated field'. This suggests that Kepkiriwat is not a Tuparian language but rather forms a branch of its own. The issue awaits further investigation.
We have argued that this new proposal is superior to the  reconstruction in that it avoids the postulation of unexplained bifurcations of reflexes and the proposal of exception-ridden splits that, moreover, lack phonetic plausibility. Rodrigues' (2007) proposal of a series of labialized consonants to PT is rejected too, as the segments in question lack reflexes different from those of their plain counterparts and because the positional developments of contextual vowels were shown to be spurious.

ACKNOWLEDGMENTS
We are grateful to two anonymous reviewers for their comments on the presentation and the substance of the paper. These have certainly improved the quality of our submission, and the authors are fully responsible for any remaining errors or shortcomings. We are especially grateful to the editors and technical staff of the Boletim for their swift and high-quality work in preparing the proofs and dealing with our observations on necessary adjustments and revisions. This appendix includes all etymologies that contain the relevant vowels in at least two branches of the family, as well as several other etymologies that were mentioned in the body of the text. In what follows, cognates which cannot be regularly derived from the reconstructed etyma are marked with (!), and the irregular reflexes of specific segments are underlined (except in cases of irregular deletion of segments).