Rhotic Variation in Brazilian Portuguese

Abstract

We present acoustic and articulatory data from an experiment designed to test the phonetic variability of rhotics in Brazilian Portuguese, focusing on the São Paulo variety. Ultrasound tongue imaging was used to examine the realisation of rhotics in a range of phonological environments. Our analysis reveals that word-initial and intervocalic fricatives are acoustically and articulatorily distinct for most speakers. We attribute a tendency for utterance-initial fricatives to display longer duration, less voicing, and greater tongue-dorsum displacement than word-medial intervocalic counterparts to phonetic enhancement at the site of a major prosodic boundary. Similarly, rhotic taps in utterance-final position show a tendency for devoicing and frication (aspiration or assibilation) speaker-dependently. By comparison, word-medial pre-consonantal and intervocalic taps are characterised by shorter durations and greater voicing: hence, a pattern of phonetic reduction in prosodically weaker environments. We relate our findings to theoretical debates around the phonological status of rhotics in Portuguese. Whilst not providing conclusive proof in favour of any one particular approach, our results highlight the need to recognise the reality of prosodically driven strengthening in developing a full account of rhotic variation in the variety.

1. Introduction

This article presents an analysis of rhotic variation in Brazilian Portuguese (BP), and specifically the variety spoken in São Paulo (i.e., the paulista variety). Our study draws on acoustic and articulatory data from an experiment using ultrasound tongue imaging (UTI). The contribution of this element of the study is descriptive in nature. We also make a contribution to theoretical debate around the theme of context-dependent articulatory strengthening and weakening as well as considering the broader impact of our results for phonological analysis of rhotic allophony in the language.

In this section, we begin by outlining some key facts about rhotic variation in Portuguese, both from phonetic and phonological perspectives. We then discuss previous work of relevance on the phonetic characteristics of rhotic consonants cross-linguistically.

1.1. Portuguese Rhotic Variation

Portuguese as a language displays extensive cross-dialectal variation in the realisation of rhotic consonants. Brazilian dialects, more specifically, show patterns of variation that may depend on regional, social, and stylistic factors as well as phonological contextual factors. The data in

Table 1. Portuguese rhotic variation by phonological environment.

Environment	Variant(s)	Examples
a. Intervocalic	[ɾ]	caro ‘expensive’
b. C2 in tautosyllabic cluster	[ɾ]	prato ‘plate’
c. Word-final pre-vocalic	[ɾ]	mar alto ‘high sea’
d. Intervocalic	[h], [x], [ʁ], [r]	carro ‘car’
e. C2 in heterosyllabic cluster	[h], [x], [ʁ], [r]	Israel ‘Israel’
f. Word-initial pre-vocalic	[h], [x], [ʁ], [r]	rato ‘rat’
g. Word-medial pre-consonantal	[h], [x], [ɹ], [ɻ], [ɾ]	carta ‘letter’
h. Word-final pre-pausal	[h], [x], [ɹ], [ɻ], [ɾ]	bar ‘bar’

below illustrate the major patterns (see Câmara 1972; Cristófaro Silva 2003; Mateus and d’Andrade 2000; Rennicke 2015).

In environments (a–b), a tap occurs in almost all varieties of Portuguese, and taps occur most commonly in (c). Thus, dialectal variation principally affects environments (d–f) and (g–h). Intervocalically, orthographical <rr>, as in carro, is typically a fricative. Glottal realisations, which can include both [h] and a voiced variant, [ɦ], predominate in the São Paulo variety under investigation here. Velar [x], which may also occur with voicing (hence, [ɣ]), is characteristic of Rio de Janeiro, i.e., the Carioca variety. Uvular [ʁ] is the most wide-spread variant in urban dialects of European Portuguese (EP), whereas the alveolar trill occurs in a more restricted, typically rural, dialect space in the present day.

The same patterns of phonetic variation that affect intervocalic <rr> are also observed when rhotics occur as the second member of a heterosyllabic cluster. Words exhibiting such sequences are not numerous in the language, such that (e) in Table 1 can be considered a marginal pattern. Descriptions of Portuguese generally state that the same dialect-specific variants that occur intervocalically in environment (d) also occur word-initially in environment (f), at least with regard to place and manner of articulation. Voicing is more variable, with greater incidence of, e.g., [ɦ] and [ɣ], in intervocalic contexts (further details below).

Environments (g–h) in Table 1 admit variants that may or may not also occur in other environments. In EP, rhotics in canonical coda position are realised as taps. However, BP dialects display variation between fricative realisations like [h] or [x] that are, again, characteristic of Rio de Janeiro. Particular to São Paulo is the variable use of rhotic approximants (i.e., [ɹ] or [ɻ]) in pre-consonantal and pre-pausal coda contexts.

1.2. Descriptive and Theoretical Research on Portuguese Rhotics

Regarding the phonetics of this variation, the most extensive study on rhotics in BP to date comes from Rennicke (2015). Rennicke studied rhotic productions in 14 speakers of BP from Lavras in the state of Minas Gerais using a combination of semi-structured interviews and a sentence-completion task. Due to its relative proximity to the state of São Paulo, dialects spoken in the south-west of Minas Gerais display certain similarities to the paulista variety: importantly, this includes the occurrence of rhotic approximants in coda position. Rennicke (2015, pp. 88–89) observes that speakers residing north of an isogloss line that bisects Minas Gerais favour fricative realisations of coda rhotics instead of approximants.

Following the taxonomy laid out in Cristófaro Silva (2003: further comments below), Rennicke employs a categorisation system that includes strong-R and weak-R—i.e., fricative realisations vs tap-like realisations, respectively. This relies on an impressionistic identification method to identify rhotic variants, which includes observation of voicing bars in the spectrogram to distinguish voiced from voiceless realisations (cf. Sebregts 2014 for Dutch). Fricative productions with observable high-energy noise were categorised as velar or uvular, whereas low-energy fricatives were categorised as glottal.

In accordance with these criteria, participants in Rennicke’s study seemed to overwhelmingly favour [h] and [ɦ] in strong-R contexts, which comprise the word-initial and intervocalic environments in (d) and (f) in Table 1. Marginal occurrence of other variants, including [ʁ], [χ] and deletion, was also noted. Realisations of weak-R were more highly variable. In VRV-sequences, nearly all realisations are tap-like; however, Rennicke notes that a high proportion of taps are partly weakened. Both full [ɾ] and weakened [ɾ̞] also occur with high frequency in cluster contexts. Devoicing of taps is relatively common, particularly in post-tonic pre-pausal environments and following voiceless consonants. Other variants, such as approximants and rhotic deletion, occurred sporadically.

Rennicke treats coda rhotics separately from strong- and weak-R. The analysis reveals extensive variation in both word-medial and word-final coda contexts. In word-medial (pre-consonantal) codas, rhotic approximants are the dominant realisation. These include [ɹ], retroflected [ɻ], and an R-coloured vowel transcribed as [ɚ]. Acoustically, Rennicke distinguishes [ɹ] and [ɻ] through observation of F3: rhotic realisations displaying a shallow rise in the F3 trajectory are transcribed as [ɹ], whereas [ɻ]-productions are identified through a falling F3 contour over the VR sequence. Unlike [ɻ], [ɚ]-realisations are characterised by a flatter formant structure throughout the duration of the vowel, where the frequency of F3 is low and close to F2. Deletion is also fairly common in VRC contexts.

Word-finally, the situation is somewhat similar. Retroflex approximants are the most common variant across all contexts examined, particularly in pre-consonantal and pre-pausal contexts. Pre-vocalically, word-final rhotic realisations include [ɻ] and various tap-like productions (i.e., [ɾ], [ɾ̥], and [ɾ̞]). Word-final rhotic deletion is reasonably common, whereas other realisations occur with much lower frequency (e.g., glottal and aspirated variants).

Summing up the findings, Rennicke proposes that BP rhotic realisations span a gradient lenition trajectory comprising [ɾ] > [ɾ̞] > [ɹ] > [ɻ] > [ɚ]. A multitude of factors are argued to be responsible for the relative reduction of a given rhotic including phonological positional factors, as discussed above, as well as interactions with vowel quality, lexically specific variation, and stylistic factors, such as emphatic or controlled speech patterns. Rennicke also makes the point that a fuller understanding of some of the patterns of variation would only be possible with further work of an articulatory nature: it is to this need that the current paper aims to make an initial response.

In addition to Rennicke’s work, rhotic variation has been the subject of variationist sociolinguistic research. Oushiro and Mendes (2013) studied rhotic productions in a 1.5-million-word spoken corpus of paulistano Portuguese (see also Ricardo and Schwindt 2023 on rhotics in Porto Alegre BP). Focusing on coda rhotics, the authors first draw attention to geographical patterns, with coda [h, ɦ] having associations with Belo Horizonte (state capital of Minas Gerais)1 and [x, ɣ] strongly indexing Carioca speech, as already mentioned. For São Paulo, a distinction is made between coda [ɾ], which is characteristic of normative, educated city speech, and use of coda [ɻ], which is expected to occur in more rural areas and in the speech of less highly educated individuals. Relatedly, Rennicke (2015) and Oushiro (2021), inter alia, draw attention to the relative stigmatised status of [ɻ].

As well as studies with an empirical focus, Portuguese rhotics have been the subject of theoretical analysis. Mateus and d’Andrade (2000) approach rhotic variation as a case of positional allophony. They make a fundamental claim in this regard, namely, that all surface rhotics derive synchronically from a single phoneme, /ɾ/. Accordingly, intervocalic dorsal fricatives (as in carro) are analysed as deriving from an underlying geminate (i.e., /ɾɾ/). In this case, a two-stage derivation first syllabifies instances of /ɾɾ/ heterosyllabically and then assigns [dorsal] place to the onset: hence, /ɾɾ/→/ɾ.ʁ/ (referring to EP). Secondly, the coda /ɾ/ deletes, thereby generating [ʁ]. On the surface, this [ʁ] is therefore representationally indistinguishable from word-initial [ʁ], which is derived by a simpler [dorsal]-default rule that applies syllable-initially, i.e., [_σɾ→[ʁ].2

Cristófaro Silva (2003) takes a different approach that aims to account for rhotic patterns in a range of dialects on the basis of a set of shared representations. Unlike Mateus and d’Andrade’s monophonemic approach,3 Cristófaro Silva assumes a three-way lexical contrast, namely, between a phonemic tap and two further rhotic phonemes notated as /$\overline{\mathrm{R}}$/ and /R/, respectively (i.e., the weak, strong, and coda Rs referred to by Rennicke 2015). This approach is perhaps best suited to dialects exhibiting a three-way surface pattern, such as the São Paulo dialect. Assuming the occurrence of intervocalic alveolar trills in this variety (rather than glottal fricatives), Cristófaro Silva posits the rules listed in

Table 2. Rhotic allophony; adapted from Cristófaro Silva (2003, pp. 142–43).

a.	/$\overline{\mathrm{R}}$/	‘Strong-R’	→	[r]	Word-initial, intervocalic, and post-consonantal contexts
b.	/R/	‘Coda-R’	→	[ɹ]	Coda contexts
c.	/ɾ/	‘Weak-R’	→	[ɾ]	Onset clusters and intervocalically

.

On the one hand, a possible objection to this characterisation of the facts is its lack of predictive power. That is to say that deriving the surface forms from a three-way underlying contrast, the distribution of surface variants relies exclusively on the asymmetric patterning of /ɾ, $\overline{\mathrm{R}}$, R/ at an underlying level. On the other hand, a potential advantage of this analysis is that it does not rely on simplification of putative lexical /ɾɾ/-sequences to generate intervocalic fricatives.

1.3. Articulatory Studies on Rhotic Consonants

Phonetic studies—and particularly articulatory studies—on Portuguese have generally focused on phenomena other than rhotic allophony: e.g., the laterals, obstruents and oral/ nasal vowel contrasts (Barlaz et al. 2018; Charles and Lulich 2018; Howson et al. 2022; Martins et al. 2008, inter alia). However, one previous UTI study on BP rhotics by Howson and Kochetov (2018) is particularly relevant in regard of the aims of the current research. Howson and Kochetov (2018) studied rhotic fricatives and taps in four vowel contexts in the speech of six speakers of BP from São Paulo. They found evidence that both types of rhotic ([ʁ] and [ɾ] in their transcriptions) are sensitive to coarticulation with surrounding vowels, with particularly strong effects observed for /i/. Phonological context also exerted an effect on the articulation of the rhotic fricatives, with greater variability in tongue position in intervocalic <rr> as compared to word-initial fricatives in corresponding vowel environments. The authors therefore suggest that, in comparison to word-initial [ʁ], the intervocalic fricative lacks an oral place of articulation.

Furthermore, research on other languages provides a relevant point of comparison for investigation of rhotic variation in BP. For example, Recasens and Pallarès (1999) studied coarticulatory patterns involving /r/ and /ɾ/ in Catalan using a combination of electropalatography (EPG) and acoustic methods. Across both vowel contexts studied (/i_i/ and /a_a/), they found a significant difference in F2, which reached a higher frequency with the tap than the trill. This is attributed to increased dorsal activity in the latter. Durational differences were also noted, with rhotic articulations in the /i_i/ environment tending to be longer than in /a_a/. Regarding articulation, both trills and taps were articulated at a more anterior location on the palate in /i_i/ than in /a_a/. The authors further argue that achieving a specific dorsal target for the production of the trill makes it more resistant to coarticulation with surrounding vowels than the tap, particularly in /i/ contexts.

In a follow-up EPG study on Majorcan and Valencian Catalan, Recasens and Espinosa (2007) examined trill and tap realisations in a broader range of phonological contexts and vowel environments. Evidence for somewhat stronger trill productions is reported for pre-vocalic as opposed to intervocalic contexts (i.e., larger and longer context periods in word-initial /r/ than in word-medial intervocalic /r/: see Baltazani and Nicolaidis 2013 for a similar result from an EPG study on Greek). Varying qualities of the pre- and post-rhotic vowels produce complex effects on the duration and contact profiles of /r/ and /ɾ/. Dialectal differences in the degree of dorsum retraction also occur in the data, for example, the tendency for coda tap-reinforcement in Eastern Catalan. Valencian speakers, on the other hand, tended to have weaker articulations, particularly in syllable-final environments. Context-dependently, rhotic devoicing and the occurrence of epenthetic vowels following tap-realisations are also reported, although these phenomena are strongly speaker-specific.

In addition to EPG, UTI has been applied to studying rhotic articulations. Proctor (2011) presents a comparative analysis of rhotics in Russian and Spanish. For Spanish, articulatory tracking reveals fine-grained articulatory differentiation of [ɾ] and [r]. Proctor (2011) highlights that both rhotics in Spanish have a dorsal target (see also Boyce et al. 2016). For the tap, this is described as being similar to a schwa-like mid-central vowel. The trill, by contrast, is characterised by a retracted dorsum and an [o]-like tongue-body configuration. Comparisons of Russian plain and palatalised trills similarly confirm that [r] has a dorsal target that resembles a mid-central vowel (hence, a similarity to Spanish [ɾ]). [r^j] has a more fronted articulation than the plain trill due to the influence of the secondary palatalisation gesture.

English rhotics have been studied extensively using articulatory methods. For example, Campbell et al. (2010) studied the production of Canadian English /ɹ/ using lip-movement tracking and ultrasound. They found that the timing of lip, tongue-body, and tongue-root gestures varied according to the syllabic position in which the /ɹ/ occurred. In word-initial /ɹ/ contexts, labial and tongue-body movements were produced earlier and with greater overall magnitude than the radical gesture. Although with less consistency, this pattern was reversed when /ɹ/ occurred in word-final contexts (pre-vocalic and pre-consonantal).

Zhou et al. (2008) examined North American English /ɹ/ using a combination of MRI and acoustic methods. They note that /ɹ/ is produced with two main tongue configurations, either with bunching of the tongue body or retroflexion of the tongue tip (i.e., tip-down vs tip-up realisations). Acoustically, similar to what Rennicke (2015) and Sebregts (2014) describe for Portuguese and Dutch, respectively, both types of /ɹ/ are characterised by a lowering of F3, such that the frequency of F3 comes very close to F2. Computer modelling of the vocal tract based on MRI images from bunched and retroflex realisations reveals a more obvious acoustic differentiation at higher frequencies, specifically F4 and F5.

Acquisitional studies of North American /ɹ/ have provided further insights into the bunched/retroflex distinction. Klein et al. (2013) show that speech therapeutic interventions designed to correct misarticulated /ɹ/ can result in accurate productions that may be either bunched or retroflex (cf. Tiede et al. 2010). They suggest that a preference for one tongue shape or the other may partially reflect the morphology of the vocal tract in children. They also point out that consonantal /ɹ/ and vocalic /ɹ/ (i.e., realisations that are more /ɚ/-like) are acquired differently, and that adult populations with typically developed speech show an array of tongue shapes in /ɹ/ that sometimes do not neatly fit into the bunched or retroflex categories (Delattre and Freeman 1968). Similarly, on the basis of a UTI-study involving four children acquiring Canadian English, Magloughlin (2016) argues that children go through an exploratory period with bunched and retroflex /ɹ/ before settling on a dominant adult pattern.

Further investigation into the articulatory properties of American English rhotics has added to this picture. In a UTI study involving 27 speakers of American English, Mielke et al. (2016) report that some speakers use only a single rhotic variant, whereas others show what is described as “idiosyncratic allophony”, i.e., a pattern of alternation between bunched /ɹ/ and retroflex /ɻ/ that is dependent on contextual factors. The variable speakers produced retroflex rhotics most commonly in word-initial pre-vocalic position and in the context of /a/ and /o/. With /i/ and in word-final postvocalic position, /ɹ/ tended to be bunched. The fact that some patterns of alternation between the two rhotic types are clearly speaker-specific rather than general is attributed to perceptual factors: the difficulty in distinguishing perceptually between /ɹ/ and /ɻ/ leads to an inconsistency in the use of either variant in specific environments for some speakers.

Dediu and Moisik (2019) further discuss the hypothesis that vocal-tract anatomy may contribute to variation in the use of bunched vs retroflex rhotics, as mentioned in Klein et al. (2013). Using MRI, the authors studied rhotic productions among 80 participants, the majority of whom were L2 speakers of American English. The results support the generalisation that whereas the bunched and retroflex variants have highly similar, if not perceptually indistinguishable, acoustic profiles, the two configurations may show different coarticulatory behaviour. With respect to their main research question, Dediu and Moisik (2019) also argue that anatomical differences between participants principally affect tongue bracing. This is relevant for retroflex articulations, which require a greater degree of radical-pharyngeal bracing than bunched /ɹ/. Thus, speakers with narrower vocal tracts may favour bunched realisations over retroflex ones.

Rhotic variation in English varieties and languages other than North American English has also been studied. Heyne et al. (2020) collected UTI data from 62 speakers with the aim of ascertaining whether similar variation between bunched and retroflex articulations occurs in New Zealand English. They showed that some speakers do show a pattern of categorical alternation between /ɹ/ and /ɻ/. Similar to the results of Mielke et al. (2016), these speakers favoured use of the retroflex /ɻ/ in word-initial position and in back-vowel contexts. Low central vowels also patterned with tip-up articulations in initial and intervocalic positions. Word-finally, and in the context of alveolar and velar consonants, tip-down articulations were more commonly observed. There were also non-alternating speakers, 25 of whom consistently used bunched articulations and 12 of whom consistently used a retroflected tongue configuration. In general, the authors concluded that the same patterns of variation reported for North American English occur in New Zealand, but that bunched tip-down articulations are generally less common.

Additionally, rhotic productions produced by English speakers from a variety of locations in England were studied using UTI by King and Ferragne (2020). Like New Zealand English, the authors note that tip-up retroflex articulations are much more common in English than American speech. Of the 29 speakers recorded, the majority favoured retroflex rhotics, whereas 7 speakers produced a higher proportion of bunched articulations, and 3 used a mixture of the two types of /ɹ/. Uniquely among studies looking at English /ɹ/, King and Ferragne (2020) also examined patterns of lip configuration in addition to tongue shape. They confirm that bunched /ɹ/ is produced with greater lip protrusion than retroflex /ɻ/; and interestingly, lip protrusion in /ɹ/ does not correlate directly with rounding on flanking vowels. The authors attribute this to a compensation strategy that arises due to bunched /ɹ/ having a smaller front cavity than retroflex /ɻ/.

Regarding Scottish English, Lawson et al. (2011) discuss variation in approximant realisations that include tip-up, front-up, and bunched articulations. These articulations form a socially stratified continuum that is not dissimilar to the variation described for the paulista rhotics. In Scottish English, middle-class speakers favour realisations in which the post-dorsum has visible inward bunching: a similar feature is observed for Canadian French rhoticised vowels in Mielke (2015). The working-class participants in Lawson et al. (2011) more regularly produced tip-/front-up articulations lacking observable bunching. Lawson et al. (2014) further note that blending of approximant-R with a preceding vowel, resulting in a rhoticised vowel, is a regular feature of middle-class speech. Acoustically, these realisations are characterised by a progressively falling F3, as noted by Rennicke (2015) for some BP realisations. Lawson et al. (2013) also find evidence for some working-class speakers producing rhotics with delayed tongue-tip raising, which results in the gestural maximum occurring within a period of voicelessness. These realisations also lack the obvious F3 drop that characterises approximant and [ɚ]-realisations.

In a study on Dutch, Strycharczuk and Sebregts (2018) examined allophonic alternation of rhotic consonants in three phonological environments, namely, word-initially, word-finally, and in a “fake-geminate” /-r#r-/ external-sandhi context. Similar to BP, Standard Dutch exhibits a uvular rhotic in onset position and a coda approximant, the latter of which can be produced with tongue bunching or retroflexion. These patterns were confirmed by ultrasound data in Strycharczuk and Sebregts (2018), in which evidence of coda-R reduction was also found. In the sandhi context, some speakers favoured [ʁ]-like articulations that strongly resemble the onset allophone. Other speakers, and in particular those who showed a tendency to reduce coda-R, produced a fake-geminate R that has an intermediate configuration between [ʁ] and the coda approximant realisation.

Articulatory research on non-European languages is less common. However, the results of two studies are particularly relevant. Dravidian as a language family is well-known for displaying a high number of phonological contrasts in place of articulation, in sonorants as well as in obstruents. Scobbie et al. (2013) report UTI data from a single speaker of Malayalam, which includes realisations of [ɾ], [r], [l], and two retroflex sonorants: [ɭ] and [ʐ]. [ɭ] is characterised by extensive raising and retroflexion of the tongue tip (cf. comparisons of geminate [tː] vs retroflex [ʈː] reported in Kochetov et al. 2014 for the related Dravidian language, Kannada). The same is visible to a certain degree in tongue traces from [ʐ]; however, the degree of retroflexion is observably smaller than in [ɭ], both in terms of bunching in the front body and arching of the tongue tip. Clear differences in the tongue configuration for [ɾ] and [r] are also noted, with the trill displaying retraction at the root and a flatter front-body position.

Tabain and Beare (2018) studied the alveolar∼retroflex contrast in Arrernte (Pama Nyungan, Australia). Regarding the rhotics, Arrernte has both an alveolar tap and a retroflex approximant. UTI data show that [ɾ] and [ɻ] are not consistently differentiated in spatial terms for all speakers examined. Three of the six speakers studied articulated [ɾ] with the tongue root in a more retracted position than [ɻ]. All speakers produced the retroflex with the front body of the tongue in a higher configuration (as measured at the temporal midpoint of each rhotic); however, the difference was more extreme for some speakers than others. Tabain and Beare (2018) draw attention to a small hollowing or bunching in the front-body area of the tongue that resembles the Malayalam [ʐ], and perhaps also its alveolar trill (cf. Scobbie et al. 2013, p. 110). This occurs in both [ɾ] and [ɻ], which may suggest that some Arrernte speakers retroflect the tap as well as the approximant to some extent.

Huang et al. (2024) examined tongue and lip configurations in Southern Mandarin rhoticised schwas. Comparisons focus on phonemic /ɚ/ (i.e., /ɚ/ lacking any kind of morphological status) and suffixal /ɚ/ (i.e., /ɚ/ that fulfilled a diminutive derivational function historically). Despite descriptions that describe it as a retroflex vowel, Huang et al. (2024) find no evidence for tip-up realisations in either type of /ɚ/. Instead, they describe two subtypes of tip-down bunched realisations based on EMA and UTI recordings, namely, dorsum-up and dorsum-down. Dorsum-up articulations show some degree of tongue retraction, whereas dorsum-down realisations are characterised by a convex tongue shape. Small articulatory differences between morphemic and non-morphemic /ɚ/ are observed, and the authors further note on the basis of acoustic measures that non-morphemic /ɚ/ has a tendency to show a higher F1 than its morphemic counterpart. They further suggest that suffixal /ɚ/ is produced with some degree of diphthongisation. Interestingly, there are also differences in lip configuration, where realisations of the suffix /ɚ/ do not show increased lip protrusion that differentiates them from plain (i.e., non-rhoticised) vowels. By contrast, non-morphemic /ɚ/ displays similar degrees of lip protrusion to the monophthong /o/.

UTI has also been applied to studying rhotic productions among Standard Mandarin speakers in an L2 context. Chen et al. (2024) examined the articulation of English /ɹ/ in a group of Mandarin–English bilinguals with varying levels of proficiency in English. In addition to patterns of acoustic variation, Chen et al. (2024) show that the most proficient English speakers in the sample varied categorically: they used a bunched tongue configuration in articulating Mandarin /ɹ/ and switched to a retroflex articulation when speaking English. The Mandarin bunched /ɹ/ interestingly displays clear similarities to the bunched configurations observed for middle-class Scottish English in the studies cited above. By contrast, the English /ɹ/ that highly proficient Mandarin–English bilinguals used in Chen et al. (2024) is generally characterised by a higher front-body position than bunched /ɹ/. Less proficient speakers showed some tendency to use a retroflected articulation when speaking English, although this occurred only variably. The least proficient English speakers used a bunched Mandarin-like articulation consistently, i.e., when producing speech in both Mandarin and English.

1.4. The Current Study

In light of the complexities that BP rhotics present and the current state of knowledge about the phonetics of rhotics cross-linguistically, as summarised above, this study has two main aims. Firstly, we ask to what extent detailed phonetic study of Portuguese rhotics confirms the occurrence of contextual patterns previously reported on the basis of impressionistic observation. Secondly, we aim to identify and describe fine-grained patterns of articulatory variation, particularly in view of previous findings about contextual reduction in BP (and other languages like Catalan and Dutch). Additionally, we aim to relate the experimental findings to theoretical claims about the operation of allophony, as described in Section 1.2.

The rest of the paper is structured as follows. Section 2 below presents our methodology. The results are then presented in Section 3. Section 3.1 considers the acoustic properties of rhotics in Brazilian Portuguese. We move on to discuss the articulatory properties of these rhotics in Section 3.2. Section 4 discusses the impact of our findings for our specific research questions. Section 5 concludes the paper.

2. Methods

2.1. Participants

Seventeen L1 speakers of BP from São Paulo participated in the study. Ten self-reported their gender as female, and seven as male. Ethical approval for the study was obtained from the Ethics Committee at the University of Edinburgh prior to speaker recruitment. Participants responded to an advertisement posted at the University of São Paulo to volunteer for the study. The minimum level of education for the participants was therefore undergraduate degree level. Full informed consent of participants was sought prior to data collection.

Recordings took place in a soundproof laboratory at the University of São Paulo. None of the participants reported any diagnosed hearing or speech pathologies prior to taking part. Due to poor ultrasound image quality, data from three speakers were excluded from the analysis. Our analysis is therefore based on data from 14 speakers. This represents a convenience sample of participants who volunteered for the study with a reasonable binary gender split (8 females, whom we refer to as BPF1–BPF8, and 6 males, referred to as BPM1–BPM6).

2.2. Recording Setup

Our data were collected using an Echoblaster 128 ultrasound machine and the Articulate Assistant Advanced (AAA) software package (Wrench 2003–2023). The ultrasound data were recorded with an average rate of 76 frames per second and with simultaneous audio recording sampled at 22.1 kHz using a Sony ECM-55B microphone. The microphone was attached to an ultrasound headset during recording. Participants wore the headset for the duration of the experiment. The headset holds the ultrasound probe in position under the chin during recording and allows for adjustment of the probe angle and placement to suit the physiology of individual participants.4 An ultrasound probe with a depth of 70 mm and a 123.8° field of view was used for the experiment.

2.3. Materials

During the experiment, participants were asked to read a set of sentences from a computer screen. The stimuli were programmed into AAA prior to data collection and were presented to participants in a randomised order.

Table 3. Experimental test words by context.

	Test Contexts		Target Words	Gloss
a.	Word-initial fricatives	##R & V#RV	rabo	‘tail’
			ramo	‘branch’
			rapa	‘scrape.imper’
b.	Word-medial fricatives	VRRV	barra	‘bar’
			farra	‘fun, a laugh’
			marra	‘unwillingly’
c.	Word-medial onset taps	VRV	apara	‘chip’
			para	‘stop.imper’
			vara	‘stick’
d.	Word-medial coda taps/approximants	VRC	arma	‘weapon’
			barba	‘beard’
			harpa	‘harp’
e.	Word-final taps/approximants	R## & VR#V	bar	‘bar’
			mar	‘sea’
			par	‘even’

below lists the target words that form the basis of our analysis. These words were embedded into carrier phrases for the purpose of examining contextual variation in rhotic production. The carrier phrases are presented in

Table 4. Carrier sentences used in the experiment.

	Test Environments	Carrier Phrase	Gloss
a.	[Utt [WrdR–	_____ foi a palavra que disse.	‘The word that I said was _____.’
b.	[Utt … [Wrd–R–] … ]	Digo _____ outra vez.	‘I say _____ once again.’
c.	–RWrd] Utt]	Outra vez, digo _____.	‘Once again, I say _____.’

.

All the words listed in Table 3 are real lexical words. They were chosen specifically due to the occurrence of /a/ immediately adjacent to the target rhotic in each environment. Furthermore, non-rhotic consonants occurring within the target words were limited to labials and labio-dentals (e.g., ramo, farra, barba, etc.).5 This set of consonants was chosen so as to minimise lingual coarticulation with the rhotic productions that are the focus of the study. All words except for the monosyllabic bar, mar, and par, and trisyllabic apara [ɐ.ˈpa.ɾɐ] are disyllables with trochaic stress.

The items in (b–d) in Table 3 were embedded into carrier sentence (b) exclusively. The words in Table 3 (a) were placed into carrier sentence (a) for the purpose of testing utterance-initial realisations (##R), and into sentence (b) for testing utterance-medial word-initial realisations (V#RV). Similarly, the items in Table 3 (e) were placed into frame sentence (b) to test word-final pre-vocalic rhotic realisations (VR#V), and into sentence (c) for testing word-final pre-pausal realisations (R##). Participants read each word-phrase combination twice (=six tokens per speaker per test context). They also read other sentences in the same format that did not contain target words with rhotic consonants. The number of additional phrases exceeded the target ones at a greater than 2:1 ratio. Thus, the non-target phrases functioned as distractors and prevented participants from fixating on rhotic productions in the experiment. Allowing some time for calibration of the equipment for each speaker, recording took approximately 30 minutes per participant.

2.4. Measurements and Analysis

Prior to analysing the UTI data, audio recordings were extracted from AAA and exported to Praat (Boersma and Weenink 1992–2023). A small number of tokens containing disfluencies or reading errors were discarded (see Table 5). The remaining tokens were hand-segmented and subjected to acoustic analysis: some examples are shown in Section 3.1.3. Fricative realisations were segmented according to the occurrence of frication in the waveform and spectrogram. The segment boundaries for rhotic taps were placed where we observed a visible reduction in amplitude that typically coincided with obvious discontinuities in the established waveform patterns, corresponding to flanking vowels. Rhotic approximants occurred in pre-pausal and pre-consonantal environments, speaker-dependently. The offset boundary for approximants was placed where there was a clear loss of acoustic energy (coinciding either with an utterance break or consonantal closure). The onset boundary was sometimes more difficult to place due to gradience in transitions from preceding vowels. In these cases, we relied on observation of changes, typically loss of amplitude, of the upper formants and established waveform patterns. Where necessary, we also examined tongue movements in the UTI recordings so as to make a best estimate of the onset segment boundary.

We calculated the duration of each rhotic on the basis of the acoustically segmented material by script. The script also extracted pitch and formant-frequency readings. We measured f₀ at 11 equidistant temporal points in the longer rhotic realisations: the fricatives, approximants, and fricated taps. This method was not practical for the simple taps due to their short duration. Accordingly, we extracted pitch measurements at 3 points in [ɾ]-tokens (corresponding to the acoustic onset, the midpoint, and the acoustic offset). We determined voicing to be present in the signal at any measurement point where pitch values were extractable.

Values in Hertz were recoded using binary values: 1 indicated the presence of voicing at a given measurement point and 0 indicated the absence of voicing. These values were used to calculate a voice-ratio measurement for each token (see Ramsammy and Strycharczuk (2016) on similar use of voice ratio in the analysis of fricatives). Voice ratio is calculated by summing the binary-coded values and dividing by 11 in the case of the fricative and approximant rhotic realisations, and dividing by 3 in the case of [ɾ]-tokens. This yields values in a continuous range between 0 and 1, in which 0 represents complete absence of voicing and 1 indicates full voicing throughout the duration of consonant. The duration and voice-ratio data were modelled with mixed-effects linear regression using the lmer and lmerTest (Kuznetsova et al. 2020) functions in R (R Core Team 2018). Post hoc comparisons were conducted using estimated marginal means with conservative Bonferroni corrections (emmeans package: Lenth 2023).

Formant-frequency measurements were extracted for a subset of the data that included only rhotic approximants. We aimed to track movement in formant contours over time in order to visualise the occurrence of expected acoustic correlates of rhotic retroflexion. We took F1, F2, and F3 readings at 11 equidistant points over the full acoustic duration of each approximant token, i.e., the same points at which pitch values were also calculated. As we were particularly interested in tracing F3 over the vowel–rhotic transitional phases, we included two additional measurements before the acoustic onset of the rhotic: these were equidistant to the 11 measurements taken during the rhotic realisation. The full range of formant-frequency measurements therefore span the range −0.2 to 1 in normalised time (where values extracted at −0.2 and −0.1 correspond to the final periods of the [a] preceding the rhotic, 0 corresponds to the rhotic onset, and 1 corresponds to the rhotic offset).

The script was programmed to calculate 5 formants below 5 kHz with a 25 ms window length: we found this to be the ideal range for accurately measuring rhotic approximants produced by both male and female speakers. We normalised the raw data on a speaker-by-speaker basis using z-transform. These data were then submitted analysis using GAMs (mgcv package: Wood 2023), in which phonological context and normalised time were fixed predictors (see Section 3.1.2). Plots were generated using ggplot2 (Wickham et al. 2023).

Following acoustic analysis, the annotation files were re-imported into AAA for the purpose of analysing the UTI data. Initial traces of the tongue surface were generated for all ultrasound frames in the recordings using the batch process function in AAA. Automatically generated traces occurring within the boundaries of pre-annotated rhotic realisations were hand-corrected where necessary. We then selected individual ultrasound frames that coincided most closely with the temporal midpoint of each annotation and extracted xy-coordinate values from a maximum of 42 points along the tongue-surface contour in those frames. We determined the acoustic midpoint to be the most suitable measurement location as this allows for maximal comparability between short [ɾ] and longer [h]- and [ɹ]-productions.

To partially correct for variation in the positioning of the ultrasound probe during recording for each participant, x- and y-values were scaled and centred at zero. This increased the visual comparability between speakers in plots. We calculated fitted splines on a speaker-by-speaker basis using GAMs. As per usual practice in ultrasound studies using the SS-ANOVA method (e.g., Davidson 2006; Howson and Kochetov 2018; Kochetov et al. 2014; Strycharczuk and Sebregts 2018, etc.), the fitted splines presented in Section 3.2 are plotted with 95% confidence intervals with the tongue tip to the right.6 Given that the rhotic productions that we included in the ultrasound analysis were recorded in a single vowel context (i.e., with flanking /a/), we did not encounter issues in fitting GAMs using Cartesian coordinates (cf. Mielke 2015).

3. Results

3.1. Acoustic Analysis

3.1.1. Overview

In this section, we present the results of the acoustic analysis of rhotic realisations. First and foremost, the high degree of inter- and intra-speaker variation in rhotic productions in syllable-final test contexts must be highlighted. Table 5 lists realisations per context based on broad categorical labels (approximant, fricative, tap, and trill). Word-initially, all speakers produced fricative realisations both in (a) the utterance-initial environment and in (b) the word-initial intervocalic environment. All speakers also consistently produced fricatives in (c) the word-medial <rr> environment. Some variation in voicing was observed across contexts (a–c): we discuss this further in Section 3.1.2. As expected, we observed only taps in (d) the word-medial VRV environment.

Table 5. Variation in rhotic productions across test contexts.

Test Context	Realisation	N	% Total
a. ##R	fricative	82	100
b. V#RV	fricative	84	100
c. VRRV	fricative	84	100
d. VRV	tap	84	100
e. R##	approximant	28	33
	fricative	6	7
	tap	45	54
	trill	5	6
f. VR#V	approximant	9	11
	fricative	0	0
	tap	67	81
	trill	7	8
g. VRC	approximant	21	26
	fricative	3	4
	tap	51	64
	trill	5	6

Table 5. Variation in rhotic productions across test contexts.

Test Context	Realisation	N	% Total
a. ##R	fricative	82	100
b. V#RV	fricative	84	100
c. VRRV	fricative	84	100
d. VRV	tap	84	100
e. R##	approximant	28	33
	fricative	6	7
	tap	45	54
	trill	5	6
f. VR#V	approximant	9	11
	fricative	0	0
	tap	67	81
	trill	7	8
g. VRC	approximant	21	26
	fricative	3	4
	tap	51	64
	trill	5	6

In the coda contexts, tapped realisations also predominated. One speaker, BPM1, produced only [h] utterance-finally, but used a tap in the word-final intervocalic environment. BPM1 also produced three pre-consonantal tokens with [h] but otherwise used an approximant in this context. Another speaker, BPF1, favoured the use of alveolar trills in the majority of coda-R realisations, but also produced taps in the utterance-final context. BPF4 articulated a single alveolar trill in the utterance-final context, but otherwise used only taps in this environment. Broadly speaking, all other variation in coda rhotics constitutes a distinction between approximant and tap realisations.

With regard to individual speakers, speakers BPF2, BPF4, BPF6, BPF7, BPF8, BPM1, BPM3, and BPM6 are binary speakers, according to our initial categorisation. This means that they produced exclusively fricatives in environments (a–c) in Table 5 and exclusively taps in environments (d–g).7 The remaining participants, i.e., BPF3, BPF5, BPM2, BPM4, and BPM5, are ternary speakers who produce exclusively fricatives across contexts (a–c), exclusively taps in (d), and a mixture of taps and approximants in (e–g). None of the participants in this study produced rhotic approximants without also producing taps in at least one of the coda test contexts. Speaker BPM5 is the only ternary speaker who showed no within-context variation. He produced only approximants in utterance-final and pre-consonantal coda contexts and only [ɾ] in the word-final pre-vocalic environment.

3.1.2. Fricative and Approximant Realisations

We now turn to consider the non-tapped rhotic realisations in more detail (the tap data are presented in Section 3.1.3). Figure 1 below illustrates the durational variation in fricative and approximant rhotic realisations by test context. Recall here that fricative realisations account for 100% of the data in ##R, V#RV, and VRRV environments, as per Table 5. By contrast, the approximants that occur in R##, VR#V and VRC account for only a subset of realisations in these environments.

In the fricatives, we observe a progressive decline in duration across the three phonological contexts in which they occur. The longest fricatives occur utterance-initially (mean duration = 119 ms). In the word-initial utterance-medial environment, this drops to 105 ms. Word-medial intervocalic fricatives are the shortest, with a mean duration of 94.3 ms. Similarly, although there is considerable variation,8 utterance-final realisations of [ɹ] have an average duration of 109 ms. The mean duration of utterance-medial word-final [ɹ] is 102 ms with the word-medial pre-consonantal approximants having a much shorter mean duration of 60.2 ms. Overall, these values indicate that both types of rhotic are produced with the greatest duration at the periphery of the utterance. Rhotics that occur immediately before or after a grammatical word boundary in an intervocalic context are somewhat shorter. Rhotics that occur word-medially are observably shorter, with the pre-consonantal codas displaying the shortest duration of all.

Statistical modelling confirms the significance of these findings. Given the phonetic differences between the fricative and approximant realisations, we modelled the two data subsets separately. In both cases, duration values in ms were the dependent variable. Phonological context and participant gender were included as fixed predictors along with a random intercept for participant. The contextual differences between the fricative realisations achieve significance after p-value adjustment: ##R∼V#RV β = 13.9, t = 4.42, p < 0.001; ##R∼VRRV β = 24.5, t = 7.81, p < 0.001; V#RV∼VRRV β = 10.7, t = 3.41, p < 0.01. With the approximants, word-final realisations are longer than pre-consonantal realisations: R##∼VRC β = 53.6, t = 9.64, p < 0.001; VR#V∼VRC β = 52.8, t = 6.36, p < 0.001. However, the durational difference between the utterance-final and utterance-medial approximants does not achieve significance: R##∼VR#V β = 52.8, t = 0.11, p > 0.1. No significant gender differences were observed in either model.

Regarding voice ratio, we observe an opposite tendency to duration. As shown in Figure 2, the highest levels of voicing are observed in word-medial contexts: that is, for fricative realisations in VRRV, in which the mean voice-ratio value is 0.747, and for approximants in VRC, in which it is 0.965. Again, although there is indication of a high degree of variation, utterance-initial R shows the lowest tendency for voicing (mean voice ratio = 0.2). Utterance-medial word-initial R is somewhat more prone to voicing, reaching a mean voice-ratio value of 0.559. With the approximants, utterance-final R## often displays devoicing over the final phase of the articulation. This accounts for the relative low voice-ratio mean of 0.5. Interestingly, this also occurs utterance-medially, where approximants in the VR#V are the least voiced of all test contexts (average voice ratio is 0.475 in this context).

The voice-ratio data were modelled in the same way as the duration data reported above. In the fricative model, the cline in voice ratio across phonological contexts achieves significance (p < 0.001 in all comparisons): ##R∼V#RV β = −0.365, t = −8.38; ##R∼VRRV β = −0.544, t = −12.8; V#RV∼VRRV β = −0.188, t = −4.45. Mirroring the analysis of duration, voice ratio is significantly higher in VRC than in R## (β = −0.422, t = −8.18, p < 0.001) and VR#V (β = −0.537, t = −6.98, p < 0.001). However, the difference between approximants in R## and VR#V is non-significant: β = 0.115, t = 1.583, p > 0.1. Again, no significant gender differences were observed in either model.

In addition to duration and voice-ratio measurements, we also examined the formant-trace data extracted from approximant R-realisations. Based on existing descriptions, we were principally interested in visualising changes in F3 over vowel-to-rhotic transitions. Recall that Rennicke characterises [ɻ] as displaying a gradual narrowing of the formant space between F2 and F3 in pre-vocalic rhotics, whereas [ɹ] is expected to show a rise in F3 in the transitional periods between the vowel and the following rhotic.

In initial modelling, we found that fitting GAMs to the full set of data, covering 13 time points, led to inconsistencies: for example, where partial devoicing in some approximants caused interruption of established formant trajectories. The occurrence of glottal creak, particularly at word-final boundary points, sometimes also contributed to unreliable readings (see Figure 4 in Section 3 for an example of this in an assibilated [ɾ^ʂ]-realisation). Consequently, we chose to work with a smaller temporal window for the analysis. This covers the two measurements points in the pre-rhotic vowel and all measurement values up to and including the acoustic midpoint of each approximant R (i.e., −0.2–0.5 in normalised time). Plots showing fitted F1, F2, and F3 contours over this timescale for the five speakers in the dataset who produced approximated rhotics are shown in Figure 3.

Approximants produced by speakers BPF3 and BPF5 display a progressive narrowing of the formant space between F2 and F3. As noted, this suggests a retroflected realisation of R, particularly in the utterance-final context. A similar, but visibly less extreme pattern occurs in utterance-final R with BPM5. The flatter F3 contours that we see for BPM2 and BPM4 may be indicative of a less retroflected approximant. The slight F3 rise for BPM4 suggests a more [ɹ]-like quality according to Rennicke’s diagnostics; it is also possible that the flatter F3 trajectory, in general, indicates a more central alveolar approximant (with relatively little contextual variation for these speakers). In all cases, the patterns shown in Figure 3 are impressionistically somewhat less dynamic than the examples presented in Rennicke (2015). This may indicate that this set of speakers favour a more central type of approximant, context-dependently, or otherwise that the degree of retroflexion in their rhotic approximants is rather small on the whole. We discuss these patterns further with reference to the UTI data in Section 3.2.2 below.

3.1.3. Tapped Realisations

The rhotic realisations that we broadly categorised as tap-like present several complications for analysis. In addition to [ɾ] (the prototypical tap), we encountered assibilated and aspirated taps, i.e., [ɾ^ʂ] and [ɾ^h], respectively. Taps followed by a period of glottal frication occurred only in the utterance-final context: they were produced by BPF2, BPM3, and BPM6 exclusively. Assibilated taps, which closely resemble realisations reported for certain dialects of Spanish,9 were observed in R## and VR#V. These were produced exclusively by female speakers (BPF6, BPF7, and BPF8). A typical assibilated realisation is shown in Figure 4. Observe here the occurrence of high-frequency noise following the production of [ɾ]. This coincides with a loss of periodic voicing following several periods of glottal creak that precede the articulation of [ɾ]. This contrasts with the utterance-final aspirated tap shown in Figure 5, where the frication is more diffuse, and with the devoiced trill, also shown in Figure 5, in which multiple short apico-alveolar contacts are observable in the spectrogram.

Additionally, rhotic taps in the VRC context were frequently produced with a short svarabhakti vowel excrescence between the tap and the following consonant.10 This also occurred—although much more rarely—in utterance-medial VR#V. Examples of this are shown in Figure 6. Note here that the post-rhotic vocalic material is considerably shorter than the duration expected to be observed in a full vowel.

Regarding duration, we observe small differences in simple-tap realisations across test contexts. [ɾ]-realisations are longest utterance-finally, where the mean duration reaches 31.8 ms. [ɾ] displays the shortest duration in the word-medial intervocalic environment (mean of 19.7 ms), and mean values fall between these extremes in VRC (20.7 ms) and VR#V (23.6 ms). These values are within the expected range for taps: for instance, Recasens and Espinosa (2007) refer to 15–30 ms (exclusive of positional variation) as being the recorded durational range for [ɾ] cross-linguistically.

In the utterance-final context, where we observe the greatest differentiation of variants, trilled [r] and assibilated [ɾ^ʂ] have similar durations: i.e., average values of 147 ms and 148 ms, respectively (recall that the trill data represent realisations produced by a single speaker, BPF1, so the values reported here should not be taken as anything more than indicative). [ɾ^h]-realisations are considerably shorter, reaching a mean of 88 ms here. A further point to highlight is that the duration of [ɾ^ʂ] appears to be somewhat more stable in VR#V-realisations, as compared to R##. Utterance-medially, the mean duration of [ɾ^ʂ] is 98.2 ms, which is shorter that the duration of [r]-realisations produced by BFP1 in this context (i.e., 124 ms). BPF1’s trills in the VRC context have a comparable mean duration of 119 ms.

Due to the extensive idiosyncratic variation, it was not possible to fit meaningful models to the duration data other than for the simple taps. As with the fricative and approximant-realisations, [ɾ]-durations were modelled with duration in ms as the dependent variable, context and gender as fixed effects, and a random intercept for speaker. Variation by speaker gender was insignificant (β = 2.038, t = 1.219, p > 0.1). Some comparisons against R## (in which [ɾ] displays longer duration) did achieve significance: R##∼VRV, β = 11.61, t = 3.286, p > 0.01; R##∼VRC, β = 10.55, t = 2.944, p > 0.05. The R##∼VR#V comparison is non-significant (β = 7.68, t = 2.166, p > 0.1). VRV realisations are shorter than VR#V (β = 3.93, t = 3.238, p > 0.05) but non-significantly shorter than word-medial taps in the VRC environment (β = 1.06, t = 0.762, p > 0.1).

Voice ratio is also observably affected by realisation and phonological context: we see that voicing is most consistent in word-medial VRV taps and least consistent in the utterance-final context. [ɾ]-realisations register a mean voice-ratio value of 0.25 utterance-finally, which rises progressively across the other test contexts: 0.758 in VR#V, 0.793 in VRC and a near-ceiling value of 0.914 in VRV. The lowest voice-ratio values are observed for [ɾ^h], which displays a mean value of 0.033. Utterance-final [ɾ^ʂ]-realisations are mainly unvoiced in all contexts: mean voice-ratio values are 0.15 in R##, 0.145 in VR#V, and 0.09 in VRC. The trills produced by BPF1 display more consistent voicing in VRC and VR#V, in which voice-ratio averages are 0.8 and 0.688, respectively. This drops to 0.145 in the utterance-final context.

As with duration, heavy skew in the voice-ratio data prevents full inferential analysis. Nevertheless, a model built on the data subset containing only the simple-tap realisations with the same structure as the duration model confirms the significance of some of the observed patterns. These bear a strong similarity to the results for duration. Comparisons with R##, in which voicing is least consistently observed, all achieve significance at least at the 0.05 level: R##∼VR#V, β = −0.453, t = −3.03, p > 0.05; R##∼VRC, β = −0.517, t = −3.42, p > 0.01; R##∼VRV, β = −0.609, t = −4.032, p > 0.001. Realisations of [ɾ] are only marginally more voiced in VR#V than in VRC (β = −0.148, t = −2.888, p > 0.05), whereas no significant difference between VRV and VRC emerges (β = −0.085, t = −1.435, p < 0.1).

In general, therefore, the analysis reveals a parallel patterning of duration and voicing: i.e., tap-realisations that are significantly differentiated in duration by phonological context also tend to display significant differentiation in voicing. By contrast, where context exerts only a minimal effect on [ɾ]-durations (e.g., the VRV∼VRC comparison), voicing is also not significantly affected.

3.2. Articulatory Analysis

In this section, we present the results of the articulatory analysis. It was necessary to restrict this part of the analysis to five of the test contexts discussed above, namely, utterance-initial ##R, utterance-final R## and word-medial VRV, VRRV, and VRC. In VR#V-tokens, rhotic realisations were strongly coarticulated with the back vowel that occurs word-initially in carrier phrase (b) in Table 4. Likewise, coarticulation with the word-final back vowel in carrier phrase (c), used to test V#RV, meant that tokens produced in this context were not directly comparable to the word-medial environments in which we used an /a_a/ frame, as mentioned. Despite this, focusing the analysis on the three word-medial environments and the two utterance-peripheral environments allows for our main research questions to be addressed.

As an initial point of reference for the detailed discussion of contextual variation that follows, Figure 7 shows relative tongue positions for the five test contexts as produced by a representative binary-allophony speaker.

There are two major distinctions to draw attention to here. Firstly, the tap realisations in VRV, VRC, and R## are mainly distinguished from the fricative realisations in VRRV and ##R by the position of the front body of the tongue and the tongue tip. As expected, [ɾ] is produced with a tip-up configuration, whereas the fricative realisations show a tip-down orientation. Secondly, although the tap splines show a high degree of overlap, the fricatives are not spatially identical: the dorsum and front body of the tongue show a significantly greater displacement in utterance-initial ##R than in VRRV. We comment further on these patterns with reference to the other participants in Section 3.2.1, Section 3.2.2 and Section 3.2.3 below.

3.2.1. Fricative Realisations

In the environments in which all speakers consistently produce fricative realisations of R, we observe two main patterns in the articulatory data. Figure 8 shows the rarer pattern: speakers BPF1 and BPM2 produce utterance-initial fricatives that do not differ in terms of tongue position from word-medial <rr>-fricatives. In both cases, the realisation is a period of glottal frication which, in articulatory terms, resembles the lingual configuration for the flanking /a/ vowels. There is a suggestion of an extremely minor amount of fronting in ##R for BPM2, but this is only differentiated from VRRV in a small section of the post-dorsum.

Unlike BPF1 and BPM2, the other speakers in the sample all produce fricatives in the ##R environment that display some degree of tongue advancement relative to more retracted realisations word-medially in VRRV: see Figure 9. For a number of speakers, utterance-initial R is produced with the tongue dorsum raised towards the uvula. This is most clearly the case for BPF2, BPF3, BPF4, and BPM4, and to a less extreme extent BPF5 and BPM6. BPF2, BPF4, BPF5, and BPM4 achieve the more uvular utterance-initial target through a small amount of fronting of the tongue root as well as dorsum raising. The tongue tip, however, remains in a similar position to the more glottal realisations observed in the VRRV context (in which the tongue root is more retracted and the dorsum is much lower).

BPF8 and BPM3 also display root fronting in ##R. In comparison to the other speakers, both the dorsum and front body of the tongue have a more raised configuration in ##R. Realisations produced by BPM1, BPM5, and BPM6 in these test contexts trend in the same way; nevertheless, differentiation of VRRV and ##R is observably less extreme for these speakers, such that fricatives in both environments appear to have a more glottal than uvular place-of-articulation target. BPF6 and BPF7 show a lesser differentiation of VRRV and ##R fricatives. BPF6 does display the same tendency for utterance-initial root fronting as the other speakers; however, there is no evidence of dorsum or front-body raising. BPF7 shows almost no root fronting and only a very minor degree of front-body raising in ##R relative to a slightly lower configuration in VRRV.

Despite this, the generalisation applicable to the majority of speakers is that utterance-initial fricatives are produced with a larger displacement of the tongue, particularly at the root and dorsum, than word-medial ones. The degree of differentiation varies across speakers, but in general, the articulatory data match up with the acoustic data, in which longer durations and more devoicing were shown to occur utterance-initially than in word-medial VRRV.

3.2.2. Approximant Realisations

Figure 10 below shows comparative plots for the speakers who produced rhotic approximants in the R## and VRC test contexts. Significant differences between the splines calculated for these environments are observed except with BPF3. This speaker’s realisations differ observably from the other speakers in Figure 10 in that BPF3 produces a tip-down approximant in which bunching is observed in the dorsal region in both R## and VRC. These realisations bear some resemblance to the “dorsum-down” pattern reported by Huang et al. (2024) for Southwestern Mandarin. Here, the tongue tip shows a smaller displacement in pre-consonantal R than in utterance-final R. This is the area of the tongue in which positional variation is greatest (notice the large confidence intervals here). However, given that the tongue displays a near-identical configuration across the root, dorsum, and front-body sections, approximant realisations of R are not robustly differentiated by context for this speaker.

Clear contextual differences are noted for the other speakers in Figure 10, all of whom produce tip-up approximants. Like BPF3, rhotic approximants produced by BPM4 in the two coda environments shown are not distinct in terms of the configuration of the tongue root. However, the tongue body is much higher in utterance-final realisations than in VRC. In the latter context, the tongue has a flat configuration across the dorsum and front body, possibly with some small pre-dorsal hollowing. This is also visible in Figure 11 as a small dip in the splines around the 0 mark on the x-axis for BPM2 (and note that the same occurs in BPM4’s VRC-spline around 0.2). This kind of bunching can be a characteristic of rhotic retroflexion, as seen in Arrernte (Tabain and Beare 2018) and Malayalam (Scobbie et al. 2013). It contrasts with the more arched position that occurs, for example, in R##-realisations produced by BPM4.

Regarding BPF5 and BPM5, these speakers also differentiate utterance-final coda R from word-medial pre-consonantal R. In the former environment, the root is advanced and the front body of the tongue has a larger displacement than in VRC. A minor dip occurs in the VRC spline for BPM5 around 0.2 on the x-axis in Figure 10; however, nothing resembling this is visible for BPF5. Interestingly, these speakers’ splines for the two contexts bear similarities to ultrasound data reported in Tabain and Beare (2018) for Arrernte, where a higher tongue position and more advanced root is characteristic of [ɻ] (as compared to [r], which is produced with a more retracted root and lower front body), at least for some speakers. This may therefore suggest that both BPF5 and BPM5 retroflect the tongue to a greater degree in utterance-final coda R than in the VRC environment.11

The UTI data from approximant realisations raise questions in light of the acoustic analysis presented above. In particular, the fact BPF3 appears to make a minimal articulatory distinction between R## and VRC is perhaps unexpected since this was the speaker for whom the formant traces extracted for these two environments were most distinct of all the approximant users (cf. Figure 3). Despite that, the higher front-tongue position observable in both BPF3 and BPF5’s utterance-final approximants conceivably contributes to the more extensive narrowing of the F2/F3 space: i.e., tongue tip raising causing an observable increase in F2 across the V-to-R transitions in Figure 3.12 The same pattern was noted for BPM5, who also articulates utterance-final R with the front body in a higher position. It is notable, likewise, that BPM4, for whom the V-to-R formant transitions did not display the characteristic drop in F3 expected of [ɻ], also produces utterance-final approximants with more of a central/palatal arch in the tongue. F3 also generally patterns lower in R## than VR#V and VRC for this speaker. The slight bunching that we observe in the male speakers’ VRC splines in Figure 10 does not seem to have a strong acoustic correlate, other than perhaps a somewhat lower F2 frequency in the dynamic V-to-R measurements.

3.2.3. Tapped Realisations

We now turn to consider the articulatory data from the rhotic realisations that we broadly classified as taps. As noted above, this is the category that displays the most inter-speaker variation. It must also be highlighted that imaging tap movements with an ultrasound system that achieves ca. 75 frames per second means that measurements are not consistently extractable at or near the gestural maximum of [ɾ]. Whereas the UTI frame rate offers a significant improvement on, for example, EPG data, it must be acknowledged that our observations about the spatial variation in [ɾ] are to some extent limited by the capabilities of the equipment.

Figure 11. Tongue splines for [ɾ]-realisations in R##, VRC, and VRV test contexts (binary speakers).

Figure 11. Tongue splines for [ɾ]-realisations in R##, VRC, and VRV test contexts (binary speakers).

With this caveat in mind, we begin by considering the binary speakers: i.e., those who exclusively produce fricatives word-initially and in VRRV, and tap-like realisations elsewhere. Figure 11 shows comparative plots for tap-like realisations in the R##, VRC, and VRV test contexts. We observe here that tongue splines for taps produced in the VRC and VRV environments are not significantly differentiated for the majority of speakers. Indeed, speakers BPF4, BPF6, BPF7, and BPM3 show little to no differentiation in tongue position across the three test contexts (but note that the tongue reaches a slightly higher displacement in R## for BPF7 and BPM3). Splines for BPF2 show greater differences across the three environments, with the front body of the tongue exhibiting a larger displacement in utterance-final tap realisations. In some cases (as with BPF2, BPF4, BPF6), the tongue has a marginally lower configuration in the word-medial intervocalic context; however, this only appears to be significant for BPF2. BPM1’s [ɾ]-spline (i.e., in VRV) resembles taps produced by the other speakers. Recall that this participant produced [h]-realisations of coda R: thus, the tongue position in R## and VRC resembles the /a/-like configuration that we observe in Figure 9.

Speaker BPF8 presents a particularly interesting pattern. As previously noted, this participant is one of the female speakers who regularly produced assibilated rhotic realisations utterance-finally. In R## here, we see significant raising of the tongue body across the dorsal and front-body areas, extending into the tip. This possibly suggests active compression of the airsteam by channel narrowing for the generation of frication across the front area of the tongue. Nevertheless, it is remarkable that there is no evidence for this kind of gesture in the other consistent assibilating speakers (i.e., BPF6 and BPF7 here). In fact, tap productions produced by BPF6 and BPF7 look to be among the most homogeneous across test contexts, as noted. It seems to be the case, therefore, that assibilated taps produced by BPF6 and BPF7 are precisely that: [ɾ]-like articulations in which the tongue reaches its maximal displacement and remains in position for a short period. During that period, airflow is maintained but voicing tends to drop out, thus resulting in the generation of relatively weak apico-alveolar voiceless frication. BPF8 is clearly different in this regard: the tongue contour shown in Figure 11 looks to reflect a more obviously controlled sibilant production, possibly resembling something closer to [ɾ^ɕ] than [ɾ^ʂ].

Finally, we come to the word-medial rhotic realisations produced by the ternary speakers. Figure 12 shows splines for VRC and VRV in which these speakers tended to favour the use of taps (despite consistent use of approximants utterance-finally). BPF1 is included here for comparison: recall that this speaker articulated alveolar trills in VRC. In this case, the fitted splines show that the tongue orientation for [r] is not markedly different from that for [ɾ], at least when measured at the durational midpoint of both types of realisation. The dorsum reaches a slightly more retracted position in [r], as expected, but this difference is marginal. The other speakers in Figure 12 all produce approximants in VRC versus simple taps in VRV. For BPM2 and BPM5, the tongue configuration in the VRC-approximants is not robustly different from intervocalic [ɾ]. The root is slightly more advanced for the tap than for the approximant in both speakers, and the dorsum has a larger displacement in BPM5’s approximant realisations. This suggests an articulation that is more [ɹ]-like than [ɻ]-like: as noted above, BPM5 produces approximants that look to be more retroflected in the R## environment.

Otherwise, BPF2 and BPM4 show approximant realisations in VRC in which the tongue has a higher configuration at the front body than in intervocalic [ɾ]. The characteristic bunching that is observable in BPM4’s pre-consonantal approximant is also absent from the tap spline. There is also a small degree of bunching visible in BPF2’s VRC spline, although in this case its locus aligns around –0.15 on the x-axis, i.e., in the dorsal region. Interestingly, BPF5’s approximant spline shows a significantly lowered tongue tip relative to its position in [ɾ]. As with some of the utterance-final realisations discussed above, this may indicate a more palatal (or palatalised realisation) than [ɹ] or [ɻ].

3.3. Summary of Results

Our analysis confirms the existence of both binary and ternary allophony patterns among BP speakers in our participant sample. Binary speakers produce rhotic fricatives in ##R, V#RV, and VRRV contexts and taps elsewhere. Ternary speakers have the same patterning of rhotic fricatives, but taps are restricted to VRV; they instead articulate rhotic approximants in R## and VRC contexts. One speaker (BPF3) uses a bunched tip-down configuration in these environments, whereas the other speakers all produce tip-up approximants. We have also highlighted that tap-realisations are not uniform across test environments, specifically with regard to the occurrence of tap frication (of both glottal and assibilated qualities) and excrescent vowels context-dependently.

In addition to the categorical allophonic alternations, we have shown that there is variation in the acoustic and articulatory profiles of rhotic variants that is dependent upon prosodic context. Rhotic fricatives are longer, less voiced, and articulated with a larger displacement of the tongue in ##R than VRRV: i.e., [ʁ] shows evidence of strengthening when it occurs at the utterance edge. Similarly, with the exception of the bunched [ɹ]-realisations produced by BPF3, approximants are articulated with longer durations and larger tongue displacements in utterance-final environments than in word-medial VRC. Whereas similar spatial variation does not occur with the taps, [ɾ] is longer and less voiced in utterance-final position (for those speakers who use this variant in R##) than utterance-medially.

4. Discussion

The results presented above confirm a number of points about rhotic variation in BP. In agreement with previous work that has surveyed the phonological patterning of rhotic variants, our data reveal complex inter- and intra-speaker variation. And as traditional phonological accounts have outlined, we have encountered categorical allophonic patterns in the data. However, further than a broad distinction between, e.g., strong-, weak-, and coda-R, a finer categorisation of phonetic variability emerges.

One striking finding concerns the patterns of variability that seem to align with specific phonological boundaries. We have shown that rhotic fricatives in utterance-initial position are longer, less voiced, and often produced with larger displacement of the tongue dorsum than fricatives in the VRRV context. These phonetic characteristics are precisely what we would expect to observe under a scenario of edge strengthening: a controlled enhancement in the context of a high-level prosodic boundary.

This phenomenon has been documented for a number of languages on the basis of articulatory experimentation. As noted in Section 1.3, word-initial rhotics show a tendency for gestural enhancement in Catalan (Recasens and Espinosa 2007) and Greek (Baltazani and Nicolaidis 2013) relative to more reduced articulations in word-medial environments. Electromagnetic midsagittal articulometry studies on French (Tabain 2003) and English (Byrd et al. 2005) have also shown that articulatory gestures are bigger and are produced with faster velocities (and longer overall durations) when consonants occur adjacent to higher prosodic boundaries, like the utterance or IP, than when they occur domain-medially or adjacent to lower prosodic boundaries (e.g., phonological word or syllable). Similarly, research using EPG (Fougeron and Keating 1997; Keating et al. 2003, 1999) and UTI (Ramsammy and King 2023) confirms the existence of the same kind of prosodically-driven dynamic variation for other languages.

What we have observed for fricatives in ##R (and acoustically for V#RV) versus VRRV partially replicates one of the results from Howson and Kochetov (2018), who claim that BP fricatives are gesturally specified for a uvular/pharyngeal place-of-articulation target. Arguably, it is the occurrence of /ʁ/ at the left edge of a P-word boundary (utterance-medially in their data) that maximises the potential for realisation of this target. Word-medially in VRRV, by contrast, this target has a greater tendency to be undershot, resulting in what Howson and Kochetov (2018) refer to as a placeless contextual variant of /ʁ/. In addition to illustrating that word-medial fricatives are produced with smaller tongue displacements than in ##R, we interpret our acoustic findings as further evidence that VRRV admits gradient weakening—or lenition, in Howson and Kochetov’s terms—due to its occurrence in a weak prosodic position. Conversely, the utterance boundary in ##R promotes a stronger, enhanced (or fortified) fricative realisation that affects both its articulatory and acoustic qualities.13

Like the fricative realisations, we have observed that taps are shorter and more voiced in word-medial contexts (VRV and VRC) than in utterance-final position. To some extent, loss of voicing and increased duration at the periphery of the phrase is suggestive of the same sort of strengthening as in utterance-initial fricatives. However, there is a complication in that it is also utterance-finally that we most commonly observe aspiration and assibilation of [ɾ]. It is reasonable to assume that these realisations reflect different articulatory timing strategies. In the former case, the tongue tip produces an [ɾ]-gesture and then rapidly lowers as vocal-fold vibration ceases, thereby generating a period of glottal frication. In [ɾ^ʂ], voicing also falls out, but the tongue tip must remain in a raised position as airflow is sustained. This points to both types of articulation being planned and controlled: the question that arises, then, is whether these can feasibly be conceived as edge-strengthening phenomena. On the one hand, it is possible that both are active edge-marking strategies that signal the end of an utterance, in which case tap frication could be considered an enhancement. On the other, the failure of voicing could be described in terms of partial undershoot or failed execution of a laryngeal target.

Indeed, Rennicke (2015) interprets the occurrence of devoiced and aspirated taps as weakening in line with the lenition trajectory that she proposes: [ɾ] > [ɾ̞] > [ɹ] > [ɻ] > [ɚ]. However, an alternative way of looking at this is that speakers actively manage the relative timing of laryngeal and lingual gestures in order to fortify rhotics at both utterance peripheries. Thus, what happens in a categorical way utterance-initially, namely, devoicing and frication, may have a more continuous and variable parallel utterance-finally. If this is the case, then any fortition–lenition trajectory that is proposed for a language like BP must have a sensitivity to prosodic environment built into it. Whereas a scale like [ɾ] > [ɾ̞] > [ɹ] > [ɻ] > [ɚ] may capture a general trend, it does not account for what our results indicate, namely, that strengthening and reduction are at least partially predicable on the basis of prosodic factors.

Regarding the approximants, it is worth considering to what extent edge strengthening may be responsible for the variability we have observed between VRC and R##. The tongue displacements shown in Figure 10 for R## certainly appear larger than the VRC counterparts; and furthermore, it is clear that at least some of the speakers produce approximants with different degrees of retroflexion in these contexts. In general, UTI is somewhat limited in recording very fine-grained differences in retraction and arching of the tongue tip. Other imaging techniques such as fMRI may be necessary to gain a fuller understanding of the contextual variation in retroflex rhotic articulations in BP (see recent work by Kochetov et al. 2023 on retroflex consonants in Kannada). Nevertheless, the current articulatory data coupled with the acoustic findings suggest that speakers may produce [ɹ]-like and [ɻ]-like approximants along a continuum, with the degree of retroflexion—along with variation in gestural duration and voicing—correlating with prosodic boundaries, at least in part.14

Recognising the reality of edge strengthening is also important in respect of the theoretic claims about the phonological status of rhotics in Portuguese. For example, recall that two of the speakers do not differentiate fricatives in spatial terms in the utterance-initial versus word-medial intervocalic environments, and that some other speakers show only relatively small differences here. These facts, perhaps, offer some support for the claim that rhotic fricatives in the two environments are representationally identical at the surface phonological level (as in Mateus and d’Andrade 2000, for example). Theoretical models of this type would nevertheless have to rely on the assumption of some sort of controlled, prosodically driven variability in phonetic production to account for the variation in the realisation of surface [ʁ] (as a derived variant of underlying /ɾ/) for the majority of speakers who show this pattern.

In fact, the same issue arises if one takes the opposite theoretical position and assumes that synchronic BP has a contrastive rhotic fricative, as in Cristófaro Silva (2003). Such a model still assumes a representational equivalence between fricatives in ##R and VRRV, such that controlled edge strengthening in the former context and domain-medial reduction in the latter is the most coherent way of explaining the variability we have described. However, the articulatory and acoustic differences in fricatives in these contexts could also be predicted by the third type of theoretical model, i.e., one in which only intervocalic fricatives are phonemic (cf. Câmara 1972). In this case, it could be concluded that the tendency for a more reduced fricative—or perhaps even a placeless one, as Howson and Kochetov (2018) propose—reflects the phonetic realisation of a separate phonological category. The realisations that are spatially and temporally different in word-initial contexts would thus reflect the realisation of a computationally derived fricative, e.g., by a /##ɾ/→[ʁ] rule, rather than a lexical one. In other words, observable non-isomorphisms between utterance-initial and word-medial fricatives might be taken as evidence that these are representationally distinct in the phonology in some way.

Regarding the putative derivation of intervocalic fricatives, there may be good diachronic reasons for being suspicious of analyses that rely on /ɾɾ/→[ʁ] mappings. Portuguese regularly reduced all other inherited geminate sonorants to singletons: as in Lat. annus > ano ‘year’, olla ‘pot’ vs olaria ‘pottery’. It is possible that /ɾɾ/ may have been exceptional in this regard; but aside from orthographic conventions, it is equally possible that inherited rhotic geminates followed the same pattern of degemination as other sonorants. Crucially, as none of the theoretical approaches described above make very specific predictions about what fully derived versus partially derived allophony versus full lexical contrast might look like phonetically, these issues will be resolved here. That is to say that no phonetic data are capable of speaking decisively on the issue of which of the rhotic variants we have described may be contrastive and which contextually generated.

More concretely, there are sociophonetic implications to our findings. Given the reported stigmatised status of approximants in general in BP, we wonder whether any of the speakers we have categorised as using binary rhotic allophony actively suppressed use of approximants. A sentence-reading task involving UTI may well have caused some speakers to default to a more formal register; and it should also be reiterated that all participants in our sample were educated to university level. If some participants chose to favour the prestige variant—i.e., coda taps—then this patterns in the expected direction with regard to gender: only two out of eight female speakers produced approximants under experimental conditions, whereas three out of the six male participants did so. For those speakers who did produce them, social stigma could have been a relevant factor in the degree to which participants retroflected approximants. In this sense, there may be finely tuned socio-indexical associations that span the range of more “acceptable” [ɹ]-like productions to more highly stigmatised [ɻ]. This is something that merits further investigation, along the lines of what Lawson et al. (2014) have achieved for Scottish English.

Moreover, we believe that further research on the sociophonetic status of rhotic assibilation in BP is necessary. In our data, [ɾ^ʂ] was only produced by female participants. It is clear in the Latin American dialects of Spanish in which assibilated rhotics have emerged, that the change is being led by women (Mazzaro and de Anda 2020). Whereas assibilation is a stigmatised feature in Spanish, we suspect that the situation is rather different in BP. In fact, we wonder whether [ɾ^ʂ] represents an incipient prestige variant, and one that is perhaps a younger innovation in Portuguese than in Spanish. If this is the case, then a specific gender patterning, namely, greater frequency of occurrence in the speech of high-status women, would not be unexpected.

Before concluding, it is worth remarking on some of the limitations of this study. As already alluded to, the use of controlled speech stimuli recorded under laboratory conditions with UTI instrumentation may have prompted speakers to monitor their speech quite closely. It is therefore possible that our data are under-representative in certain aspects, particularly as regards the retroflex rhotic realisations that could be said to characterise paulista BP: i.e., in being well above the level of sociolingusitic awareness. Relatedly, we were not able to control for age or socio-economic status in the participant sample. The findings from previous sociolinguistic studies, such as Oushiro (2021) and Oushiro and Mendes (2013), lead us to believe that some of the variation we have described occurs due to complex interactions of phonological, phonetic, and sociolinguistic factors. Further research will certainly be necessary to disentangle these aspects of variation from one another.

5. Conclusions

This study has documented patterns of contextual variation in rhotics in the paulista variety of Brazilian Portuguese. In response to previous descriptive work, we applied acoustic and articulatory analysis to the complex rhotic allophony pattern that exists in the language. Beyond basic categorical alternations between taps, fricatives, and approximants that existing accounts describe, we have shown that there are much more fine-grained dimensions to the contextual variability of rhotics. Variation in duration and voicing as well as differentiation of tongue shape and position often aligns with prosodic structure. Thus, we have attributed some of the variation we have observed to gradient phonetic enhancement in the context of major prosodic boundaries (i.e., edge strengthening) and concomitant reduction or undershoot of articulatory targets in weaker prosodic environments. Regarding phonological analysis, we have argued that no single pre-existing theoretical approach to the alternations under discussion necessarily captures the data we have presented better than any other. It may well be the case that new theoretical thinking will need to be applied in the future to the issues that rhotic variation in BP presents. We have further highlighted sociophonetic patterns of interest in the data and outlined specific areas that would benefit from additional empirical research.