2.1. General overview of vowels in PP

There are seven oral vowels in PP: /i, e, ɛ, a, ɔ, o, u/ (Balduino, 2022, 2025; Santiago, 2024; Santiago et al., 2022). These seven oral vowels vary according to stress position, as indicated in Table 1, where a synthesis of unstressed vowels in PP is presented. In this table, vowels only phonetically identified are highlighted in grey.

The vowel system of PP in pretonic and postonic non-final contexts is generally more reduced than the stress vowel system. This reduction is seen in the possible loss of contrast between the mid vowels [e] ~ [ɛ] and [o] ~ [ɔ] in derived words. For example, in belo [ˈbɛ.lu] ‘beautiful’ compared to beleza [be.ˈle.za] ‘beauty’, and in bola [ˈbɔ.la] ‘ball’ compared to bolada [bo.ˈla.da] ‘hit by a ball.’ This process leads to a simplified system of five vowels /i, e, a, o, u/, in which the low-mid vowels disappear, being raised to mid-high vowels (Balduino, 2022, 2025).

In PP, oral vowels in postonic syllables may undergo neutralization. In the data examined by Balduino (2025), vowel raising was observed, as in chávena [ˈʃa.vi.nɐ] ‘cup’, fósforo [ˈfɔs.fu.ɾʊ] ‘matchstick’, or lâmpada [ˈlɐ̃.pə.dɐ] ‘lamp’, characterized by the neutralization of contrasts between [e] ~ [i], [o] ~ [u], and [a] ~ [ə]. Additionally, cases of lowering were found, as in âncora [ˈɐ̃.kɔ.ɾɐ] ‘anchor’, marked by the alternation between [o] ~ [ɔ], though this was not observed for the front vowels. Therefore, similar to the pretonic subsystem, the postonic vowel system in PP consists of five oral vowels /i, e, a, o, u/, with the possibility of expansion to [i, e, ə, a, ɔ, o, u] or reduction to [i, a, u] (Balduino, 2025).²

In final unstressed positions, the vowel subsystem /i, a, u/ is similarly reduced, though there can be up to five phonetic realizations: [I, ə, ɐ, a, ʊ]. The vowels [i] and [u] are realized in reduced forms as [I] and [ʊ] due to their unstressed final position. This subset shows a pattern of vowel raising, such as the shift of [e] to [i] or [I], as seen in chefe /ʃɛfe/ [ˈʃɛ.fI] ‘boss,’ and of [o] to [u] or [ʊ], as in pato /pato/ [ˈpa.tʊ] ‘duck.’ We also observe the raising of [a] to [ɐ], as in pata/ [ˈpa.tɐ] ‘paw,’ and in some instances, the appearance of [ə] as an allophone of [a], as in pata /pata/ [ˈpa.tə].³ The centralized vowels [I, ʊ] are frequent realizations of [i] and [u], while [a] often alternates with [ɐ] and [ə].

In addition to raising and lowering phenomena, Balduino (2025) also observes vowel deletions in PP. Vowel deletion mainly affects high vowels [i] or [u] in unstressed syllables and often co-occurs with their devoicing, as indicated in (1).

According to Meneses (2012), vowel devoicing is a common phenomenon in many natural languages, typically affecting high vowels [i] and [u] when produced between voiceless segments, as also demonstrated in PP. This phenomenon has already been reported in studies such as Balduino (2020, 2022, 2025) and Santiago et al. (2022).

Regarding vowel duration, stressed vowels in PP are generally longer than unstressed ones, as evidenced in Table 2. Here, the duration of stressed and pretonic vowels is presented in the speech of men and women (see Balduino, 2025).

Based on Table 2, we observe that low vowels [ɛ, a, ɔ] in stressed positions are longer than mid vowels [e, o] and, notably, longer than high vowels [i, u], consistent with predictions from the literature and findings for BP and EP (cf. Escudero et al., 2009). Furthermore, [ə] is the shortest vowel in the stressed inventory, even shorter than the high vowels. Finally, when comparing stressed and pretonic vowels, it is clear that stressed vowels are consistently longer than their pretonic counterparts.

After presenting some general aspects of the phonological and phonetic framework of the vowels in PP, we will present the methods in the following section.

3.1 Data collection and fieldwork

The corpus of this study comes from 51 sentences (cf. appendix) containing potential contexts of degemination: V₁ and V₂ were identical and juxtaposed in final and initial word boundaries as indicated in (2).

Four male and four female informants repeated each sentence three times, all monolingual and from Principe.⁵ For analysis, we excluded the first occurrence of the sentence of each speaker, examining approximately 102 occurrences and 520 occurrences of sentences in total.⁶ The corpus was recorded in loco in 2016 and 2019. We did not have all speakers read sentences during the fieldwork, as not everyone felt comfortable with this method. For those who were comfortable reading, we provided printed sentences. The speakers read these alone and then repeated their reading three times. For those who preferred not to read, we spoke the sentence once and asked them to reproduce it three times as naturally as possible. We want to emphasize the importance of discarding the first round of repetitions, as the repetitions that followed were more automated.

The fieldwork conducted in 2016 and 2019 collected distinct sentences for the analysis of external vowel sandhi in PP. From the data collection carried out in 2019, we utilized 32 words, accounting for 64 occurrences of sentences for six speakers, totaling 384 occurrences. To expand the number of data evaluated in this study, we also considered the data collection conducted in 2016, which comprised 38 occurrences of sentences for two speakers, totaling 152 occurrences. This results in 536 occurrences, of which we have discarded some data due to a series of factors such as recording quality or interruption, as well as the implementation of phonological processes such as vowel lowering and/or raising that could alter the quality of V₂ if it were a mid-high vowel (see section 2.1).

We also discarded data that were unknown to the speakers. To this end, the sentences analyzed here underwent a validation phase before data collection began. In this context, the phrases were presented to a speaker of São Tomé and Príncipe Portuguese, who identified sentences that could lead to interpretation issues among speakers. The highlighted cases primarily concerned some lexical items that are common in Brazilian Portuguese but differ in São Tomé and Príncipe Portuguese. These items were excluded from the analysis. Additionally, if any word or sentence sounded strange to the speaker during the experiment, it was also discarded.

It is worth mentioning, finally, regarding the composition of the corpus, that we discarded data related to phonological contexts that we will not be addressing for the time being, such as parenthetical structures. This initially resulted in 426 occurrences.

Regarding the composition of the corpus, it is important to note that its collection is closely tied to the demands of fieldwork. Fieldwork has long been a fundamental part of linguistic research due to its documentary and descriptive nature, which is a key aspect of this study and also serves one of its main goals: providing a preliminary description of gemination and degemination in PP. However, fieldwork presents several challenges, and these are reflected in the composition of the corpus and the way this analysis was conducted.

First, it requires thorough preparation and respectful immersion into the speech community. This process requires multiple visits to the field, as was the case with our fieldwork conducted in 2016 and 2019 (in addition to other fieldwork trips conducted by researcher Agostinho in years before 2016). Thus, engaging with the community involved integrating into various social and bureaucratic environments, which facilitated interactions with speakers (Bowern, 2008; Brickell, 2018; Vaux & Cooper, 1999). In São Tomé and Príncipe, for instance, there is no ethics committee to formally approve research projects. Nevertheless, it was crucial to contact local authorities to obtain permission to conduct our research. Once this approval was secured, additional challenges arose, particularly in terms of the data collection environment.

There were no laboratories, specialized equipment, or even basic infrastructure like reliable electricity to enhance the quality of recordings. Despite these structural limitations, our integration into the Principean community helped us identify local needs and plan recording sessions accordingly, aiming to gather high-quality data. In the absence of acoustic laboratories, we conducted the recordings in private rooms within public spaces provided by local institutions, such as the Príncipe radio station in São Antonio do Príncipe. These settings were generally quiet, allowing the sessions to proceed with minimal external noise, as only the informant was present. To avoid disruptions due to electricity outages, all test materials were pre-printed and presented to the informants in hard copy. We brought our recorders from Brazil and used the Zoom H4n and Olympus DM-650 digital recorders.

Beyond the infrastructure challenges, it’s important to highlight the difficulties of analyzing a variety that remains largely understudied, especially in comparison to the extensive literature on dialects such as European and Brazilian Portuguese. While we refer to these different varieties as ‘Portuguese’, there are significant dialectal differences (as well as some similarities), widely reported in the literature (compare Balduino, 2025; Câmara Jr., 1970; Mateus & D’Andrade, 2000). Therefore, it is not feasible to approach a variety like Principe Portuguese with the theoretical assumption that it will behave phonologically in the same or even a similar way to related varieties like those spoken in Brazil and Portugal.

That said, we emphasize that, despite the methodological challenges described here, studying the phonological phenomena of PP through data collected on-site—via fieldwork and immersion in the speech community—is essential. Comparing these findings with the literature on BP and EP is also significant, as it provides a broader perspective on the language we politically refer to today as Portuguese, which spans diverse regions worldwide.

For now, our investigation of phenomena like gemination and degemination is necessarily limited to the dataset and phonological contexts collected during fieldwork in 2016 and 2019, with the latter serving as an extension of the 2016 corpus. Looking ahead, with sufficient financial support, expanding this corpus to encompass a wider range of phonological contexts will be crucial. Employing more varied methods and ideally implementing stronger methodological controls will allow for a deeper and more comprehensive analysis of PP.

Next, in section 3.2, we describe how the corpus was organized and analyzed.

3.2. Corpus, methods, and variables

Data occurrences were manually analyzed in Praat. Thus, we examined the contexts of the application of degemination to detect an output with only one segment, characterizing the phenomenon, or the existence of two vowels, demarcating the non-occurrence of the sandhi. In Figure 1, on the one hand, we have an example of degemination since the spectrogram only indicates the presence of one vowel. In Figure 2, on the other hand, we verify that when the process is not implemented, it is possible to discriminate between two vowels in the spectrogram. Although V₁ and V₂ are adjacent, [a] and [a] can be acoustically observed, in addition to the presence of a pause (||). Figures 1 to 3 were made using a script by Torreira (2005).

Figure 1: Waveform, spectrogram, labelling layers of segments (tier 3), and words of the sentence A menin[a]bre sempre a porta ‘the girl always opens the door’. Source: Torreira’s (2005) script.

Figure 2: Waveform, spectrogram, labeling layers of segments, and words of the sentence A madeir[a] [a]spera estragou ‘the rough wood has spoiled’. Source: Torreira’s (2005) script.

Figure 3: Waveform, spectrogram, labeling layers of segments and words of part of the sentence A barb[a][a]spera está [grande ]‘the rough beard is [long]’ in which the word “grande” was omitted from the spectrogram. Source: Torreira’s (2005) script.

For the acoustic description of the phenomenon, we measured the target segments. To do so, we compare: (i) the duration of the vowel sequence characterizing a hiatus, with pause (Figure 2) or not (Figure 3), and (ii) the vowel durations that resulted from degemination to assess the durational differences between both groups. The measurements were manually made in Praat, and, later, adopting the scale in milliseconds, we normalized the values obtained using the z-score in R-studio.

During the acoustic analysis, we also identified the possibility of lengthened vowels (Figure 4). This result surprised us because our initial hypothesis was focused solely on distinguishing cases of degemination and hiatuses based on duration — gemination was not an output expected. Based on the consulted literature on BP and EP regarding degemination, which guided the formulation of the analyzed contexts, vowel lengthening is not a predicted output (Bisol, 1996, 1999; Tenani, 2002; Veloso, 2003).

Figure 4: Waveform, spectrogram, labeling layers of segments and words of the sentence O desport[uː]nico é saudável ‘The unique sport is healthy’. Source: Torreira’s (2005) script.

As a result of this outcome, we will consider duration as a potential acoustic correlate to distinguish degemination and gemination from hiatus. A hiatus, corresponding to two vowels, would represent a sequence of segments, and therefore, it would exhibit a longer duration, equivalent to two fully produced vowels. Gemination, on the other hand, would have a shorter duration than a hiatus, since the vowel sequence is reduced, but longer than that of a simple oral vowel (degemination). In this sense, it behaves like a (semi-)long monophthong or a diphthong in PP.

To verify the reliability of this parameter, we will, as the first step of the analysis, in section 4.1, use a linear regression model to examine whether differences in duration are indeed significant for distinguishing between different phenomena, particularly gemination in relation to hiatus and degemination. This theoretical-methodological position arises from the novelty of the observed phenomenon, making it important to test whether duration, in the case of PP, is indeed an important parameter for distinguishing the different results produced in the context of vowel junctions. During the acoustic analysis, we did not examine other cues such as changes in intensity, nor did we observe glottalization or brief pauses between the two vowels in hiatus.

The construction of this corpus and the methods of analysis are based on studies such as Bisol (1996, 1999), Tenani (2002), Veloso (2003), Balduino and Araujo (2023). Therefore, sentences were created to test linguistic aspects such as the segments involved (V₁#V₂), the prosodic word (ω), and the phonological phrase (ϕ), potential contexts for implementing or not degemination in PP. Additionally, we also intend to examine whether the syllable resulting from the process may exert any influence on degemination. Next, we briefly discuss each of the phonological contexts, or predictor variables, considered.

The sentences were constructed to diversify the quality of V₁=V₂, such as a=a (256 occurrences), i=i (32 occurrences), e=e (32 occurrences), o=o (49 occurrences), and u=u (140 occurrences) to test possible effects of vowel quality. Regarding V₁ [i] and [u], the target vowel of the sandhi process may arise from the raising of a final unstressed vowel (see section 2.1) or correspond to the phonological vowels /i/ and /u/ when V₁ serves as the nucleus of a stressed vowel.

The accent and branching of ω and ϕ domains were diversified as well, in addition to being composed of different contexts of lexical stress. It is not one of our goals to analyze the boundary of the intonational phrase (IP).⁷

To delimitate the prosodic boundaries, we worked, at first, with the formation algorithms proposed by Frota (2000) and Vigário (2003), adapted from Nespor and Vogel (1986/2007). This choice is because these algorithms, which have also been reported for BP and EP, are well-suited for describing various phenomena in Portuguese, as indicated by several studies (Frota, 2000; Tenani, 2002; Vigário, 2003). This includes studies focused on São Tomé and Príncipe Portuguese and PP (Balduino & Araujo, 2023; Balduino et al., 2017; Braga, 2018). Additionally, these algorithms serve as a basis for comparing PP with other similar varieties of Portuguese.

The algorithms proposed by Frota (2000) and Vigário (2003) initially guided the creation of carrier sentences. Subsequently, the analysis of degemination will be conducted based on these algorithms, considering the boundaries and the accent of ω and ϕ. Thus, after identifying the segmental contexts relevant to degemination, we focused on analyzing prosodic boundaries.

In Brazilian Portuguese, vowel sandhi phenomena occur across various prosodic boundaries (cf. Tenani, 2002). The same is observed in PST, another variety spoken in São Tomé and Príncipe (Balduino & Araujo, 2023). Based, then, on previous studies, we evaluated the following contexts: (i) within the same ϕ (cf. c.; d.) and (ii) between distinct ϕs, as indicated in Table 3.

As mentioned earlier, the accents of ω and ϕ were also examined. Concerning lexical stress, we opposed sentences in which (i) the ω stress does not fall on the target vowels, forming a sequence of unstressed vowels (WW); (ii) one of the vowels of a V₁#V₂ sequence carries the ω accent, producing a stressed and unstressed (SW) or unstressed and stressed (WS) vowel chain, and (iii) V₁#V₂ corresponds to two stressed vowels (SS), as shown in Table 4. For the analysis of lexical stress on V1 and V2 in relation to degemination, we worked with the 2016 and 2019 corpora.⁸

Considering the accent of ϕ, we evaluated whether (i) the accent of ϕ differs from the lexical stress, as shown in Table 5. In this table, the ϕ accent is highlighted in bold, and we also consider whether the attribution of the phrasal accent is on the ϕ boundary or within the same ϕ. In general, the levels presented in Table 5 can be obtained by crossing the levels of the two predictors shown in Tables 3 and 4. Thus, there is an overlap of variables, which we will consider in our analysis. We will work with the variable whose regression model can account for a larger amount of data. For now, we will keep both predictor variables to determine which one is more relevant for explaining the process of degemination in PP.

Finally, considering that in the analyzed corpus the combination of target vowels could yield either CV or CVC syllables, we controlled for this context to test whether the resulting syllable might influence preferences for degemination, gemination, or hiatus. Typically, since CVC syllables in European Portuguese are longer than CV syllables (see Balduino, 2025), we aim to assess whether gemination is avoided in CVC contexts, as the rhyme already contains two timing units (the nucleus and coda), making it challenging to accommodate an additional timing unit represented by vowel lengthening. Conversely, in CVC contexts, degemination may be favored, given that there is only one timing unit in the rhyme (the nucleus). The contexts that reveal the resulting syllable are presented in Table 6.

Having established the method of analysis and the constitution of the corpus, highlighting the phonological contexts or predictor variables analyzed, we will follow the following analytical steps:

• (i) to test whether duration is a relevant variable for distinguishing gemination, degemination, and hiatus. For this, we will analyze duration through a linear regression model, considering not only the phenomena under focus but also other variables, such as the stress of the vowels involved, the boundary and phonological phrase accent, the resulting syllable, and the vowel quality of segments involved in the juncture process;

• (ii) to compare the duration of geminated vowels with their corresponding stressed and unstressed vowels;

• (iii) to analyze the processes of gemination, degemination, and hiatus, we will use logistic regression with three distinct models, each representing one of the processes as the response variable. The predictor variables include vowel quality, vowel lexical stress, boundary, phonological phrase accent, and the resulting syllable. We hypothesize that degemination is favored in contexts of low prominence, irrespective of the ϕ boundary (V₁#V₂; ϕ + ϕ and ϕ [ω + ω]). Although lexical stress and phonological phrase stress always coincide in these cases, we propose that the phonological phrase boundary does not act as a blocking or favoring factor for degemination;

• (iv) to discuss the results obtained.

These steps are presented in Section 4 below.

4.1. General distribution of data and duration

Like other vowel sandhi processes, degemination dissolves vowel encounters, driving, according to Bisol (1992, 1996), Tenani (2002), and Balduino and Araujo (2023), the disappearance of these vowels and the formation of a single syllable: (C).V#V(C) → CV(C).

This process was initially investigated in 51 controlled sentences in PP. We considered 102 occurrences per informant, operating with 508 tokens that presented a potential context of degemination, i.e., identical V₁#V₂ at word boundaries. The spectral and durational analysis of each of the occurrences revealed, however, a different result from BP (Bisol, 1992, 1996) and even from PSTP (Balduino, 2022) since four different resolutions were verified for identical V₁#V₂ at the word boundary:

• (i) Degemination (vowel deletion), resulting in just one short V due to the total deletion of one of the vowels;

• (ii) Gemination (vowel lengthening), resulting in v: (long vowel), which is longer in relation to the vowel resulting from deletion but smaller than a hiatus (V₁#V₂);

• (iii) Hiatus, with the maintenance of V₁#V₂, however, without gemination (or the production of a long vowel);

• (iv) Pause, with V₁ ||V₂ maintenance separated by a pause.

• (v) Data distribution is presented in Table 7, in which we observe the number of applications (Apl.), the percentage (%), and the average duration in milliseconds of each of the results shown between (i) – (iv).

In Table 7, the hiatus, or V₁#V₂ maintenance, was the least preferred possibility by speakers at 14.1% (60/426) as opposed to degemination, the most recurrent solution with 34.3% (146/426). The dissolution of the vowel cluster, by pause and gemination, was observed, respectively, in 22.5% (96/426) and 29.1% (124/426) of the data. Also, there were different results related to vowel durations in degemination (V), gemination (v:), and hiatus (V₁#V₂), as represented in the boxplot of Figure 5. From now on, we will focus on data related to degemination, gemination, and hiatus. Thus, excluding data related to pauses, we are working with 330 data tokens.

Figure 5: Boxplot showing the distributions of vowel durations resulting from deletion (V), lengthening (v:), and vowels corresponding to the hiatus (V₁#V₂).

The boxplot in Figure 5 shows the values equivalent to the first and third quartiles of the duration of the vowels resulting from deletion (V) and lengthening (v:) are presented, in addition to the duration of the hiatus (V₁#V₂). Each box also has a vertical line that extends from the minimum value at the bottom to the maximum value at the top, showing the dispersion of the data, including the outliers. The black points represent these outliers after the ends of the vertical lines or whiskers, representing values significantly higher or lower than the rest of the data. In the boxplot in Figure 5, it is possible to verify that while v presents minimum duration values, the hiatus has the longest duration in milliseconds, followed by the duration of geminated vowels.

Still in Figure 5, the medians (vertical lines inside the box) are different for each analyzed category, which is also demonstrated by the average duration shown in Table 3: degemination (V) 80 ms, gemination (v:) 173 ms, and hiatus (V₁#V₂) 280 ms. By comparing these measures from the different groups, we confirm the hiatus’s longer duration relative to the degemination and gemination durations since the mean and median of the hiatus are expressly higher than the same measures of deletion and lengthening. Concerning degemination and gemination, we also have a durational distinction between such groups: geminated vowels are durationally longer in relation to v, resulting from degemination. Finally, also in Figure 5, we visualize the distribution and outliers of the data, with the outliers being more common in the deletion data.

The data distribution, shown in Figure 5, indicates the percentage difference between the medians and means of the vowels resulting from degemination, gemination, and hiatus. This distinction is even more evident when we analyze data using a linear regression model with mixed effects. In this first model, we used vowel duration as the response variable. Result (degemination, gemination, and hiatus), Tonicity (WW, SW, WS, SS), Resulting Syllable (CVC: closed syllable by coda such as [kas], Cv: open syllables such as [ka]), Vowel Quality [i, e, a, o, u], and Gender (female and male) were evaluated as predictive variables. Participant and Sentence were considered as random effects. Tonicity refers to the combination of lexical stress in syllables 1 and 2 that are involved in the juncture process, where S represents a strong or stressed syllable and W represents a weak or unstressed syllable (see Section 3.2).

At first, we also included Phonological Phrase Prominence as a variable. However, we observed its collinearity with the Tonicity variable. We chose to remove the first one because the R² of the model is a little higher for Tonicity 0.856 than for Phonological Phrase Prominence 0.83.⁹ The results can be seen in Table 8, where the coefficients are given in the same unit as the response variable (in milliseconds (ms)), and the variables chosen as relevant by the model were Result, Tonicity, and Resulting Syllable (p > 0.001).

According to the model reported in Table 8, the vowel duration is greater according to (i) syllable stress (SS and WS being the contexts with the highest average duration); (ii) the type of syllable (being the vowels inserted in CVC syllables longer than the vowels of the CV syllables) and (iii) the type of resolution attributed to the juxtaposition of vowels on word boundaries. Thus, there is a significant durational difference between geminate and degeminate vowels and the maintenance of the hiatus, thus attesting to distinct processes triggered by the vowel encounter of identical vowels in a word boundary.

The statistical model includes duration as the dependent variable to determine whether the observed phenomena were indeed correlated with duration. As detailed in the methodology, we were surprised to identify four distinct outcomes when examining the data. Our initial intention was to use duration to differentiate between degemination and hiatus. While we anticipated finding degemination, hiatus, and pause based on existing literature from Brazilian and European Portuguese, the identification of vowel lengthening, indicative of gemination, was an unexpected discovery for PP. Despite the circularity inherent in using duration, as discussed in the methodology and this analysis, we chose to begin by assessing the impact of this variable on the type of process observed. This decision is especially significant because, to our knowledge, this is the first time such results have been reported for any variety of Portuguese. Therefore, the model aims to determine whether duration is a significant correlate for identifying and classifying processes. Our conclusion, based on the model’s findings, is that it is.

Comparing the duration of geminated vowels with their stressed and unstressed counterparts (data from Balduino, 2022), we observe that the geminated vowels maintain a longer duration compared to the other groups. This allows us to infer once again the presence of vowel lengthening as a possible result of vowel junction at word boundaries (see Table 9). Thus, although stress has been a relevant factor for greater vowel duration, we note that in cases of gemination, it is not possible to infer that the observed lengthening is a result only of the stress of the involved vowels. The duration of geminated vowels, considering vowel quality, was sometimes equivalent to twice the duration of a stressed vowel ([e, a, o]) and sometimes more than double ([i, u]).

Gemination is a process delimited by the “duplication” or lengthening of segments in different word positions or its boundary. In PP, gemination is a possibility of external vowel sandhi when V₁ and V₂ are identical vowels. In this case, V₁ and V₂, when juxtaposed, can generate a phonetic vowel lengthening, as in maracujá amarelo maracuj[aː]marelo ‘yellow passion fruit’. However, this is not the only option in the system since the degemination, or the prohibition of the coincidence of two identical vowels, in a word boundary generates the deletion of one of the moras responsible for the vowel lengthening, as in maracujá amarelo maracuj[a]marelo ‘yellow passion fruit’. There are, therefore, two distinct and opposite results: sometimes the V₁#V₂ sequence generates a gemination, and sometimes it results in degemination. Finally, the third option is established by maintaining the hiatus. In any of the three results, the durational difference of the vowel segments is relevant.

According to Bisol (1992, 1996), V₁ and V₂ resyllabification is motivated by the collision of two syllabic peaks in BP. Such a collision allows both vowels to be joined by the action of the Principle of Obligatory Contour (OCP) and shortened by the action of a shortening rule, resulting in degemination. Like Bisol (1992, 1996), we understand degemination as the result of a resyllabification that organizes new syllables, at a post-lexical level, through the application of universal principles such as (i) prosodic licensing; (ii) the sequential sonority, and (iii) the obligatory outline (Bisol, 1996). Unlike BP, however, the resolution of the V₁#V₂ hiatus can be established by creating a syllable CV or CVː. In (3), (4), and (5), based on Bisol’s representations (1996) for vowel sandhi, the different possibilities of syllabic restructuring or not are specified for PP.

In (5), an encounter between V₁ and V₂ promotes a nuclear shock and may or may not initiate the vowel sandhi process (Bisol, 1992, 1996). On the one hand, if the sandhi process does not occur, both vowels are maintained as in (3). In this case, either the hiatus is preserved, or a pause is inserted between the vowels. On the other hand, if sandhi occurs, a resyllabification process is initiated. In this phase, as a result of the clash of syllabic peaks, the segments [n], [a], and [a] are disassociated from their respective syllables and, through resyllabification, are once again associated by the action of prosodic licensing (Bisol, 1992, 1996) and by the Principle of Sequential Sonority, which licenses the loudest segment to the core.¹⁰ Achieving the operation of these principles, the nucleus now has two temporal units, or two moras, in (4). At this moment, the gemination of two identical segments can be verified to avoid two contiguous identical segments. Still in (4), the representation of the gemination process can be undone in (5) by degemination, which deletes one of the moras and results in a degeminated (and, therefore, light) syllable, characterized by the duration of only a syllabic temporal unit.

We have so far proven the implementation of gemination and degemination in PP and hiatus maintenance. However, we have not yet discussed which linguistic factors may be important in implementing such processes. Focusing, therefore, on degemination, gemination, hiatus maintenance, and inserting pauses as productive linguistic resolutions for identical vowel segments on a word boundary, we analyze, in the following sections, the phenomena according to linguistic variables such as vowel quality, syllable structure, lexical stress, and boundary and accent of the phonological phrase. Considering that degemination is a process of syllable shortening and weakening, while gemination is a strengthening phenomenon, since it creates and maintains the production of two syllabic moras in a single syllable, resulting in a heavy syllable, we intend to evaluate, in the following sections, whether positions of high prominence (stressed syllables, vowels that carry some lexical or phrasal prominence) are more conducive to gemination, or hiatus maintenance (in the latter case, preserving its moras and two distinct syllables).

4.2. Degemination, gemination, and hiatus: An overview

In the previous section, we discussed the differences among degemination, gemination, and hiatus, emphasizing duration as a key factor, as established in steps (i) and (ii) of our methodology. This section will analyze these processes using logistic regression, fulfilling steps (iii) and (iv). We will develop three distinct models — one for each process. Each phenomenon was examined as the response variable. Tables 10, 11, 12 present the results of the three logistic regressions, which incorporate the predetermined variables: boundary and accent of the phonological phrase (ϕ), lexical stress, vowel quality, and resulting syllable. Sentence and speaker were treated as random effects. Given the correlation between lexical stress and phonological phrase accent, we selected the model with the highest R², as previously noted. Notably, except for the case of ˈϕ + ˈϕ, which blocked degemination and was thus excluded from this analysis, no contexts blocked gemination (Table 11) or hiatus (Table 12). In all examined cases, boundary and ϕ accent emerged as significant variables (p < 0.001).

Comparing the three models, we observe distinct results. While all models identified the phonological phrase as the only relevant variable for the three phenomena, the influencing factors differ:

• 1) Degemination was favored within the phonological phrase (ϕ [ω + ω]) (PR: 0.92) and at the boundary of the phonological phrase (ϕ + ϕ) (PR: 0.91) when neither vowel carried phrasal or lexical stress.

• 2) Gemination, in contrast, was primarily favored within the phonological phrase (ˈϕ [ω +ˈω]) (PR: 0.70) when V₂ carried the stress of ϕ. It was also favored at the boundary of the phonological phrase (ϕ + ˈϕ) (PR: 0.695) if V₂ corresponds to an accented ˈϕ, and at the boundary (ˈϕ + ϕ) (PR: 0.59) in case V₁ also carried the accent of the first ϕ.

• 3) Hiatus was favored at the boundary of the phonological phrase (ˈϕ + ˈϕ) (PR: 0.84) when V₁ and V₂ carried the stresses of ˈϕ₁ and ˈϕ₂, respectively. It was also preferred within the phonological phrase (ˈϕ [ω +ˈω]) (PR: 0.73) when V₂ carried the stress of ϕ. In this latter case, we have a factor similar to the one that favors gemination.

• 4) The R² values of the models vary, indicating the extent to which each model explains the variation in the dependent variable. Among the three processes, gemination is the best explained by the predictor variables.

Overall, the results support our initial hypothesis: degemination, characterized by the deletion of one vowel due to a sandhi process, is linked to the absence of phrasal prominence. In contrast, gemination is favored in contexts where either V₁ or V₂ carries the stress of ϕ, regardless of the boundary. Hiatus, however, is preferred in contexts with accented ϕs or when V₂ carries the stress of ϕ, always at the boundary and within ϕ.

With the phonological phrase as a relevant domain for degemination, gemination, and hiatus, we will dedicate the following section to the description and analysis of these processes within phonological phrases. This discussion will draw on analyses proposed for both Brazilian and European Portuguese to compare the different urban varieties of the language.

4.3. Degemination, gemination, and hiatus: A general description considering the phonological phrase

Regarding the phonological phrase domain of the phenomena, degemination, gemination, and hiatus occur between two words, a locus that corresponds to the prosodic word boundaries (8a–b), within and between phonological phrases (6c–d).

In Tables 13, 14, 15, we resume the absolute application and percentage of each resolution, considering the following domains investigated: same phonological phrase (ϕ) and phonological phrase boundary (ϕ + ϕ). As indicated in the following tables, none of the investigated contexts blocked degemination (Table 13), gemination (Table 14), and hiatus (Table 15). Unlike in the logistic regression analysis, where we examined only the 330 instances of degemination, gemination, and hiatus without pauses, the percentage analysis presented here includes all 426 data points. This broader approach provides an overview of the phenomenon’s implementation, allowing us to observe the data distribution, including cases where sandhi was not applied due to the insertion of a pause.

Focusing on the domain of ϕ, we find that the ϕ boundary does not block degemination, gemination, or hiatus by inserting a pause or not. Besides, ϕ branching does not correspond to an inhibiting context, as indicated in (7). Focusing on vowel deletion and lengthening, we see its implementations on the boundary of ϕ regardless of having a branched structure, like (7a), or unbranched, like (9b). As for vowel deletion and vowel lengthening, their implementation is attested on the ϕ boundary regardless of having a branched structure, like (7a), or unbranched, like (7b). Moreover, these phenomena are also observed within a ϕ composed of two resyllabified lexical heads, like (7c). Finally, the hiatus and the pause also occur in the same contexts.

Still considering the domain of ϕ, another relevant question for our analysis concerns the hypothesis of the phenomenon occurring in vowel sequences in which one of the elements bears the prominence of the phonological phrase or ϕ accent. This issue, addressed by authors such as Bisol (1992, 1996 and later examined in works such as Abaurre (1996) and Tenani (2002, 2004, 2007), emerges from the fact that, in BP, degemination occurs even in the face of an accentuated V₂ as indicated in (8).

Examples such as (8) suggest that V₂ becomes sensitive to the process once the stress is lost. In (8), there are two PhPs: [como]ϕ and [uva madura]ϕ. According to the PhP algorithm for Portuguese proposed by Frota (2002) and adopted in this paper, the word uva ‘grape’ does not bear the PhP prominence. The degemination observed in (8) is therefore possible because the stressed vowel of uva ‘grape’ does not bear the PhP prominence. Degemination in a stressed V₂ is later revisited by Bisol (1996), who examines it through a rhythmic interpretation. The author argues that a rhythmic rule at the phrasal level prevents stress clashes by assigning main stress to the final word within the phrase. Consequently, tonicity does not inhibit degemination.

Although Bisol (1996) points out the effect of the phonological phrase accent for degemination, the author does not formalize this conclusion. As a result, Bisol (1992, 1996) focuses his analysis on lexical prominence and does not explain why the process occurs when the primary stress of the word is “weakened for prosodic and rhythmic reasons” (Bisol, 1992, p. 87), nor does it shed light on the nature and implications of such a weakening. Considering these issues and based on the description proposed by Bisol (1992, 1996), who, despite not going deeper into the relevance of the prominence of the phonological phrase for sandhi consolidation, proves the importance of a multi-level perspective to examine the phenomenon, Abaurre (1996) analyzes not only the nature of the prominence that blocks some occurrences of vowel sandhi but also proposes an explanation that resumes the interface between phonology and syntax for this phenomenon.

According to Abaurre (1996), degemination blocking and other sandhi processes, such as elision, are verified if the lexical stress of the item, assigned in the lexical component, is post-lexically interpreted as the accent of the phonological phrase. In (8), degemination is possible because V₂, even carrying the primary stress of uva ‘grape’, is not interpreted as the accent that marks the relative prominence of the phonological phrase, which is attributed to the primary stress of madura ‘mature’. Therefore, if the lexical stress of uva ‘grape’ coincided with the accent of the phonological phrase, a situation elucidated by the sequence como uvas ‘I eat grapes’, sandhi could not be applied. For the author, as degemination has the phonological sentence as its domain, it is natural that the accent of this same prosodic domain is responsible for blocking the process (Abaurre, 1996, p. 50). Thus, according to Abaurre (1996, p. 50), assuming the phonological phrase accent as a degemination blocker and not the lexical stress of the word as proposed by previous works, is the most adequate resolution to explain the impossibility of a syllable optimization promoted by resyllabification in BP.

Considering this discussion, degemination, and gemination in PP were examined according to the ϕ accent and, later, by contrasting the lexical stress of the target vowels with the prominence of ϕ. The same was done concerning the hiatus and even the pause insertion. Knowing that degemination is permissive in PP when one of the vowels carries the lexical stress or even in tonic sequences (a less productive scenario), we have shown that the prominence of ϕ is more relevant for the phenomenon implementation. In Table 16, the number of productions and non-productions of vowel deletion was delimited according to the context of the prominence of ϕ.

On the one hand, V₁ does not prevent degemination, even bearing the prominence of ϕ, a productive context (40.4%). V₂, by carrying the accent of ϕ, disfavors the process if V₂ is at the ϕ boundary (7.8 %) or in the same ˈϕ [ω + ˈω] (1.3%). On the other hand, in sequences of the type ϕ + ϕ and ϕ [ω + ω]/ [ˈω + ω] in which neither V₁ nor V₂ has the prominence of ϕ, degemination is largely performed, reaching the respective percentages of 78.9% and 60%.

Similarly to BP, PP disfavors deletion in contexts where V₂ receives the phonological phrase accent. However, it also seems more permissive since blocking is not as categorical in all domains as in the Brazilian variety (Tenani, 2002). Observing data in (9), in which V₂ carries the lexical stress and the phonological phrase accent, we note that the process was allowed mainly on the phonological sentence boundary among the few data in which degemination was found (9d-e). (9a) is the only sentence produced by only one informant in which deletion was attested.¹¹

In Lung’Ie, a language with which PP was in contact during its emergence, the resolution of vowel clusters is also more permissive in comparison with BP and EP, and it is observed in relation to its tonal system (cf. Agostinho & Hyman, 2021, 2023). Moreover, when analyzing sandhi processes in Lung’Ie, Agostinho (2015, 2016, 2019) indicates that both the elision of the first vowel and the degemination are verified before the accentuation of the first or second vowel within the domain of ɸ. For the author, there is a tendency to preserve segments at the beginning of the word in Lung’Ie processes, with the phenomena being applied from left to right and systematically blocked at the IP boundary. Considering the vast context of processes implementation, such as degemination in PP and Lung’Ie, a deeper analysis of the linguistic contact between these languages can corroborate the description and analysis of structural aspects of PP.

Having said that and evaluating sequences in which V₂, even carrying the word stress, does not carry the ϕ accent as a result of the restructuring of such a constituent, we contrasted the occurrences of sentences such as (9) (d-e) in which ω and ϕ prominences fall on V₂, as well as data like (10), whose ω and ϕ accents diverge because of restructuring.

Evaluating the data in which the ϕ accent is moved to the rightmost items sempre ‘always’ and caro ‘expensive’, the phenomenon was verified in 23 occurrences (among 30 occurrences with the target context), which allows us to reaffirm the relevance of prominence and domain of ϕ for blocking degemination in PP. In the Principense variety, degemination data analysis suggests that the maintenance of the phrasal accent, which carries the syntactic information and is responsible for pointing out some syntactic relations in the language and its boundary, can categorically block the phenomenon. Therefore, in this context, the preservation of this prosodic and syntactic information prevails in relation to the syllabic restructuring caused by deletion, with ϕ being the main domain of the process in PP (cf. Balduino, 2022) and in BP (Abaurre, 1996, p. 48).

In Table 17, focusing on gemination, we notice a distinct distribution of such process. Long vowels were mainly attested within a phonological phrase when v2 carried the phonological phrase prominence ˈϕ [ω + ˈω] (42.1%). In phonological phrase boundaries, ˈϕ + ϕ (35.1%) and ϕ + ˈϕ (32.8%) were also productive to phonetic lengthening. Still, the same did not happen in ϕ + ϕ, in which neither V₁ nor V₂ carried the prominence of the sentence (8.8%), or in ˈϕ + ˈϕ, in which both vowels contained the prominence (15%). This result reinforces the initial hypothesis that gemination is an intermediate resolution for the dissolution of the vowel sequence, being preferred in contexts in which V₁ or V₂ corresponds to strong positions, but not both vowels, because, in the case of two prominent positions, these tend to be preserved by maintaining the hiatus or inserting a pause. In a weak position, degemination is more common.

Adding to this argument are the hiatus results (Table 18), which are more recurrent in phonological phrase boundaries and phrasal accent sequences.

The results showed the productivity of two processes of external vowel sandhi for identical V₁ and V₂ in PP: degemination and gemination, which co-occur with hiatus. Degemination and gemination, in addition to being durationally distinct, are favored by divergent prosodic factors since degemination is the most common output in unstressed lexical and phrasal contexts. In contrast, gemination was preferred if V₁ or V₂ carried some prominence. Such results support the presence of two moras related to the nucleus of the syllable that was resyllabified as a consequence of the sandhi process. We show that positional prominence favors strengthening processes in PP, such as gemination, or even the maintenance of stressed syllabic nuclei by preserving the hiatus, while weak positions, such as unstressed vowels in word boundaries, favor the phenomenon of shortening, that is, degemination. This result characterizes PP in relation to varieties such as BP and EP, demonstrating that the Principense variety is more permissive with regard to the external vowel sandhi.

Long vowels as a sandhi resolution had already been attested in languages such as Luganda, a Bantu language. In such a language, a sequence of vowels in successive syllables (V.V) is not allowed. According to Myers (2020), if the first vowel is non-high, it is deleted with compensatory lengthening of the second vowel in the sequence (e.g., e + a = aː). Otherwise, if the first vowel is high, the two vowels are joined together in a diphthong (e.g., i + a = ja) (Myers, 2020). Although diphthongization is also possible in PP, compensatory lengthening has only been attested in a sequence of identical vowels in successive syllables (V𝝰.V𝝰), two identical vowels. By comparing Luganda to PP, we observe a behavior grammatically distinct, even though both present vowel lengthening as a sandhi possibility. Luganda is a language in which compensatory lengthening as a sandhi resolution follows the language’s grammar since vowel length is contrastive in Luganda (Vogel & Spinu, 2013). In PP, however, vowel lengthening is not contrastive and, until now, has only been attested as a resolution for two identical vowels, generating a Cv: syllable to avoid a CV.V syllable in a high prominence context (Becker et al., 2018; Beckman, 1998, 1999). In a low-prominence context, the preferred outcome is a CV syllable. This outcome agrees with the findings of Balduino (2022) for elision and diphthongization as sandhi resolutions for distinct vowel sequences (V.V): elision is favored in unstressed lexical (89%) and phrasal (76%) sequences, resulting in a CV syllable; diphthongization is more common when one of the vowels bears lexical (64.5%) or phrasal prominence (72.5%), resulting in a CGV syllable. These results may configure an effect of positional prominence in PP.

It is reported in the literature that languages, typologically, present less vowel contrast in unstressed syllables than in stressed ones (Crosswhite, 2004; Flemming, 2004; Kingston, 2007). Regarding the PSTP, the relevance of positional prominence was attested in processes such as devoicing and deletion of unstressed vowels, raising and lowering of pretonic vowels, and hetero and tautosyllabic vowel nasalization (Balduino, 2022, 2023; Balduino & Araujo, 2023; Balduino et al., 2020; Balduino et al., 2023). Given this fact, it is possible to hypothesize that in PP, the vowel segments that do not carry the lexical stress can suffer lenition and deletions to maximize the contrasts of prominence within the word, maximizing, consequently, the contrast of prominence between the segments in stressed and unstressed syllables (cf. Crosswhite, 2004; Nevins, 2012) — and even between the constituent segments of vowel sequences, a fact that would preserve V.V or generate results like v: and GV. This hypothesis is raised from our results and should be evaluated in future studies.