2.1 Phonological factors

/r/-sandhi seems conditioned by both prosodic and segmental phonological factors. Regarding prosody, /r/-sandhi is believed to be more frequent within than across tone units with no intervening pause (Joseph Reference Joseph, Bohumil, Osamu and Joseph1999). Moreover, linking /r/ seems to be disfavoured when the second hiatus vowel is stressed (e.g. Foulkes Reference Foulkes1997, Hannisdal Reference Hannisdal2006, Hay & Maclagan Reference Hay, Margaret, Preston and Nancy2010, Mompean & Gómez Reference Mompean, Alberto, Wai-Sum and Eric2011, Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014, Yuen, Cox & Demuth Reference Yuen, Cox and Demuth2015, Pavlík Reference Pavlík2016), which coincides with a left-dominant foot boundary in English (Giegerich Reference Giegerich1999).

Concerning segmental factors, one potential predictor is the quality of the vowels flanking the /r/-sandhi site. Some studies describe no effect of the preceding vowel in both linking /r/ and intrusive /r/ (e.g. Mompean & Mompean-Guillamón Reference Mompean and Mompean-Guillamón2009, Pavlík Reference Pavlík2016). However, other studies suggest that at least some back vowels favour /r/-sandhi. Hay & Sudbury (Reference Hay and Andrea2005) showed that intrusive /r/ was more common in 19^th-century New Zealand English (NZE) after /ɑː/ than after schwa and least of all after /ɔː/. In contrast, Foulkes (Reference Foulkes1997) for Derby English, Hay & Warren (Reference Hay and Paul2002) for contemporary NZE, and Hannisdal (Reference Hannisdal2006) for Received Pronunciation (RP) – or Standard Southern British English (SSBE) – found intrusive /r/ to be more common after /ɔː/ and /ɑː/ than after schwa. Finally, the influence of the following vowel has also been considered. Hay & Sudbury (Reference Hay and Andrea2005) found that back vowels favoured linking /r/. Pavlík (Reference Pavlík2016) reported that linking /r/ and intrusive /r/ were more common in RP before mid central or close front vowels.

Two further segmental factors potentially affecting variation in /r/-sandhi usage are the presence of another /r/ in the vicinity (before, after or at both sides) of the potential /r/-sandhi site and the length of syllables flanking that site. In this respect, Hannisdal (Reference Hannisdal2006) found that the presence of another /r/ in the vicinity clearly inhibited /r/-sandhi in RP, although no such effect was described in other studies on RP (e.g. Mompean & Mompean-Guillamón Reference Mompean and Mompean-Guillamón2009, Pavlík Reference Pavlík2016) or NZE (Hay & Sudbury Reference Hay and Andrea2005). Finally, Hay & Sudbury (Reference Hay and Andrea2005) found no influence of syllable length in NZE, but Navarro (Reference Navarro2013) noted that monosyllabic words triggered higher rates of linking /r/ than polysyllabic words for the syllable before the /r/-sandhi site in a corpus of Lancashire English.

2.2 Lexical and grammatical factors

A few studies have looked at lexical factors that might influence variability in /r/-sandhi. Lewis (Reference Lewis1975) and Hannisdal (Reference Hannisdal2006), for example, found linking /r/ to be less frequent after honorifics (e.g. Dr., Mr., Sir, etc.). Most studies, however, focus on grammatical features. According to McCarthy (Reference McCarthy1993), /r/-sandhi production should not be particularly sensitive to syntactic constituency. However, /r/-sandhi might be less frequent at clause boundaries because such boundaries mark the division of two separate syntactic planning units (Foulkes Reference Foulkes1997). In this respect, Hannisdal (Reference Hannisdal2006) noted that, although the presence of a clause boundary did not prevent the insertion of /r/ significantly, the frequency of /r/ was somewhat lower at clause boundaries than in other environments. She found, in addition, that speakers with lower linking /r/ rates tended to avoid /r/ also before syntactic boundaries, but found no such effect for speakers with higher linking /r/ rates.

Another grammatical variable potentially influencing /r/-sandhi production is the word class (lexical vs. function) of the items flanking the /r/-sandhi site. Hay & Sudbury (Reference Hay and Andrea2005) found no evidence of this effect. However, Bauer (Reference Bauer1984), Foulkes (Reference Foulkes1997) and Viollain (Reference Viollain2014) observed that linking /r/ was more common after monosyllabic function words, and a recent study by Pavlík (Reference Pavlík2016) also found a small effect. Finally, Hannisdal (Reference Hannisdal2006) determined that a combination of two lexical words disfavoured the use of linking /r/, and found higher frequencies between function words. One possible explanation for this is that function words are normally unstressed and carry little informational value, facilitating linking /r/ (Hannisdal Reference Hannisdal2006).

/r/-sandhi has also been claimed to be sensitive to morphological structure, such as the type of morpheme boundary (word-external vs. word-internal). As a sandhi phenomenon, /r/-sandhi typically occurs across word boundaries but it can also be found at word-internal morpheme boundaries in affixed words (e.g. feve[ɹ]ish, draw[ɹ]ing) and compounds (e.g. four-yea[ɹ]-old, salsa[ɹ]-evening). Linking /r/ before suffixes is considered to be categorical (i.e. obligatorily present) and often excluded from data analyses in studies of /r/-sandhi. Intrusive /r/ in affixed words and both linking /r/ and intrusive /r/ in compounds show some degree of variability. However, a few studies suggest that /r/-sandhi becomes lexicalised in word-internal positions. Mompean & Mompean-Guillamón (Reference Mompean and Mompean-Guillamón2009) found that intrusive /r/ was more common word-internally in the noun withdrawal than across word boundaries. Similarly, Hannisdal (Reference Hannisdal2006) found linking /r/ to be more common within compounds than across word boundaries. Finally, Hay & Warren (Reference Hay and Paul2002) and Hay & Maclagan (Reference Hay, Margaret, Preston and Nancy2010) found that the rate of intrusive /r/ is a partial function of suffix identity, with intrusive /r/ being increasingly more common after nominal suffixes (e.g. -ism), adjectival suffixes (e.g. -ish) and verbal suffixes (e.g. -ing).

2.3 Usage-based factors

The structural factors mentioned above may correlate with – and sometimes be explained by – usage-based factors such as lexical, morphological and vowel type frequency as well as collocability and collocability frequency. The general finding is that higher frequencies favour higher rates of /r/-sandhi (see Hay & Maclagan Reference Hay and Margaret2012 for an exemplar-based explanation). Hay & Sudbury (Reference Hay and Andrea2005), for example, found that a high lexical frequency of the word after the /r/-sandhi site disfavoured the use of linking /r/, and Hay & Maclagan (Reference Hay and Margaret2012) found that a high lexical frequency of the previous word favoured it. Pavlík (Reference Pavlík2016), however, found word frequency to have no influence. In contrast, he found that the frequency of /r/-sandhi was predicted by the frequency of the vowel type. For example, /r/ was favoured in sites followed by mid central and close front vowels, the most frequent /rV/ sequences in connected speech. Finally, Hay & Warren (Reference Hay and Paul2002) found that affixes with high type frequencies (e.g. -ing) were more likely to attract intrusive /r/ than less frequent ones (e.g. -ism).

Regarding collocability, linking /r/ seems more common in frequent collocations than in free combinations (Hay & Sudbury Reference Hay and Andrea2005, Hannisdal Reference Hannisdal2006, Mompean & Mompean-Guillamón Reference Mompean and Mompean-Guillamón2009, Navarro Reference Navarro2013, Sóskuthy Reference Sóskuthy2013, Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014), which suggests that common phrases containing highly predictable sequences of words are stored and accessed as a whole. Finally, Pavlík (Reference Pavlík2016) suggests that, although collocability has an effect on /r/-sandhi frequency, it is the frequency of the collocation that has the largest effect.

One interesting finding in some of the studies above is the interaction between frequency of use and degree of constriction used for /r/ as measured by F3. The third formant is the most salient characteristic of /r/ (Lisker Reference Lisker1957), often leading to a strong percept of rhoticity (Lawson, Stuart-Smith & Scobbie Reference Lawson, Stuart-Smith and Scobbie2018). Yet English /r/ displays gradient features in its realisation (see Section 4.2.2 below for more details). F3, for example, can be higher or lower, with lower values resulting in strong robust /r/. As far as /r/-sandhi is concerned, there is evidence that speakers who use intrusive /r/ (Hay & Maclagan Reference Hay, Margaret, Preston and Nancy2010) or linking /r/ (Hay & Maclagan Reference Hay and Margaret2012, Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014) more often have stronger, more robust /r/s. The effect also seems to emerge in potential linking /r/ cases with words that tend to occur more often before vowels. These are more likely to elicit an r-sound in linking /r/ contexts than words which tend to occur before consonants (Hay & Maclagan Reference Hay and Margaret2012). Finally, lower F3 values are also more common in frequent collocations (Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014).

4.1 Materials

A corpus containing the annual Christmas broadcasts produced by QEII was compiled for this study (henceforth QXB corpus). The Christmas message is a traditional broadcast made by the British sovereign to the people of the Commonwealth of Nations each Christmas since 1932, when King George V delivered the first of such messages (to the then British Empire). Each message serves the same communicative purpose, that is, an annual season’s greetings to the public in which, unlike other official speeches, the monarch does not deal with government policy but offers instead her views on personal milestones and the year’s major events. Despite its personal character, the Christmas message has, in fact, a great political importance in the relationship between the monarch – or more generally the monarchy – and the country since it is the only occasion in which she can convey a message of her own without needing the Prime Minister’s consent (Catsiapis Reference Catsiapis and Peter2005). The messages have been televised since 1957 and recorded in advance since 1960. The location is usually Buckingham Palace.

At the time of the analysis, the corpus spanned a period of 67 years (December 1952–December 2018) and contained 66 broadcasts, with an annual broadcast except for 1969, when the Christmas message was replaced with a documentary on the Royal Family. Table 1 shows the years (n = 63) for which the full or part of the broadcast could be accessed and analysed. This corpus expands substantially the dataset analysed in some studies by Harrington (Reference Harrington2006, Reference Harrington, Jennifer and Hualde2007) and Hinton (Reference Hinton2015) on potential diachronic change in QEII’s speech, which included only 15 and 13 broadcasts, respectively. The corpus analysed amounts to over six hours of speech and 38,800 words. The audio material was obtained from YouTube and AP Archive (http://www.aparchive.com/). The transcripts were obtained from the official website of the British monarchy (https://www.royal.uk/history-christmas-broadcast). The quality was generally considered acceptable for subsequent perceptual and acoustic analyses. Broadcasts of lesser acoustic quality underwent a noise reduction process using Audacity (version 2.1.0; Audacity Team 2012), a noise profile filtering which was set at a 12 dB reduction, 6 for sensitivity, and 3 bands of frequency smoothing.

Table 1

QEII’s corpus of Christmas broadcasts (QXB). Decades 1950s to 2010s and years. Availability for the study: NA (non-available); white cell (no broadcast); ^(p) partial broadcast.

The corpus is stylistically consistent, as mentioned above, given QEII’s unchanging role and communicative purpose as well as the broadcasts’ format and style (i.e. formal). Thus, there are no shifts in role (she always addresses the audience as the Queen) and the speech sample is genuine (often not the case of actors/actresses). Moreover, the corpus only contains adult non-pathological data. The Queen was sufficiently mature (26 years) at the time of the first message to rule out changes in her voice potentially affecting /r/-sandhi; and although aging has an impact on speakers’ general voice quality, we do not expect it to have an impact on structural patterns of /r/-sandhi production. Her accent can be described as an upper-crust or marked variety of RP as opposed to a more mainstream, neutral or unmarked variety (see Wells Reference Wells1982, Upton Reference Upton, Clive and Bernd2008, Cruttenden Reference Cruttenden2014 for discussions on varieties of RP). There are indications, however, that the Queen’s accent seems to have shifted from a more aristocratic form of RP towards a more mainstream variety characteristic of the traditional professional middle classes in England (Harrington et al. Reference Harrington, Palethorpe and Watson2000, Robson Reference Robson2016).

In order to address RQ2, QEII’s production was compared with a reference corpus to find out whether her use of /r/-sandhi was in line with group-level patterns across time. As mentioned above, the group-level data are from a study of 72 RP speakers (36M; 36F), aged 25–55, and a corpus of around 58,000 words that had been compiled from some of their public speeches (Mompean Reference Mompean, Sasha, Paola, Marija and Paul2019). The speakers were public figures or newsreaders, their date of birth ranged from 1879 to 1990, and the data were produced from 1927 to 2017. They were grouped around the concept of demographic generation/cohort. This can be defined as a group of people born within a given period of time and who experience the same significant social events. The cohorts identified were WW1 (1879–1900), WW2 (1901–1925), Silent (1926–1945), Baby boomers (1946–1960), X (1961–1980), and Millennials (1981–2000), with each cohort comprising six males and six females. The procedure for the identification and analysis of potential /r/-sandhi sites in this corpus was the one described in Section 4.2 below. Regarding linking /r/, the number of potential cases per speaker was 10, so the corpus amounted to 720 potential cases. The number of potential intrusive /r/ cases in the whole corpus was 57.

4.2 Procedure

4.2.2 Analysis and coding of /r/-sandhi sites

As with most phonetic phenomena, /r/-liaison realisations display gradient features and variability. To start with, there are different realisations of /r/-sandhi depending on the accent and, to some extent, on the historical period under study. The labiodental approximant [ʋ], often stigmatised by prescriptivists, seems to be gaining ground as a feature of some British accents, particularly southeastern England (Foulkes & Docherty Reference Foulkes and Docherty2002). In RP, [ʋ] has been described as ‘an upper-class affectation’ (Wells Reference Wells1982: 282). A voiced alveolar tap [ɾ] (or tapped /r/) seems to have been common up until the 1950s in RP (Fabricius Reference Fabricius and Raymond2017). Since then, and currently, the most typical realisation in RP is a postalveolar approximant [ɹ] and this is the only realisation found in QEII’s speech. Moreover, there are differences between /r/-sandhi and canonical /r/ in terms of tongue tip height and duration (see Gick Reference Gick1999, Tuinman, Mitterer & Cutler Reference Tuinman, Mitterer and Cutler2011, cf. Mullooly Reference Mullooly2004). Finally, there seems to be variation in F3 values across /r/ productions, with lower F3 leading to stronger percepts of rhoticity (Lawson et al. Reference Lawson, Stuart-Smith and Scobbie2018).

Despite this variation, most studies treat /r/-sandhi as a binary variable that opposes the presence of constricted /r/ to no (or vocalised) /r/. Following this approach, a categorical decision was adopted as to the presence or absence of /r/-sandhi. For this, every instance of an /r/-liaison site was inspected acoustically with PRAAT by the author, a long-term observer of the variety under investigation (e.g. Mompean & Hernández-Campoy Reference Mompean, Hernández-Campoy and Isabel de La Cruz2001). Following Pavlík (Reference Pavlík2016), the criteria used were: lowered F2 and F3, a regular structure of voicing pulses, and relatively stable pitch and intensity contours.

Regarding glottalisation, it should be pointed out that the phenomenon occurs on a continuum from laryngealisation to full glottal closure, including intermediate stages such as creaky-voiced glottal approximants (i.e. [ʔ̞] or [ʔ̰]), and it may be difficult even for phoneticians to discriminate glottal stops from creaky voice. In fact, the occurrence of true glottal stops in English may be unusual as so-called glottal stops in English are often realised as creaky voice (Ashby & Przedlacka Reference Ashby and Przedlacka2014). To avoid these problems, and following previous studies (e.g. Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014), varying degrees of glottalisation were not differentiated in this study. In any case, typical acoustic cues for glottal stops (e.g. stop gap, absence of f0, sudden energy change), creaky voice (lowered f0, decreased acoustic intensity, irregularly spaced glottal pulses with periods of varying length), or a combination of these (e.g. creaky voice + glottal stop) were considered to determine the presence of glottalisation. Finally, cases involving no glottalisation were treated as hiatus, manifesting itself as two adjacent vowels without a clearly audible transition (V#V>VV) or gemination (e.g. (V#V>Vː), often the case of rapid realisations of sequences of two schwas. Figure 1 shows examples of linking /r/, hiatus and glottalisation from the QXB corpus.

Figure 1

Examples of QEII’s productions: linking /r/ in nor any (above left); hiatus in war is still (below left); glottalisation in near and far (above right) and far outback (below right). Blue solid lines represent pitch while yellow solid lines represent intensity. Glottalisation is represented with [ʔ̰] irrespective of its degree.

Coding reliability was checked by an independent phonetician on two corpus subsets. The inter-rater agreement was high, as measured by Cohen’s kappa (linking /r/: 0.9; intrusive /r/: 0.95). Although the coding was based on visual inspection of spectrograms and f0 contours – and their related values – most potential cases categorised as /r/ were perceived as an r-sound, as /r/-sandhi tends to be a perceptually noticeable phenomenon.

5.1 RQ1: How does QEII resolve hiatus at potential /r/-sandhi sites?

RQ1 addressed how QEII resolved hiatus at word boundaries where a potential /r/-sandhi site is found. In the corpus under study, intrusive /r/ was not used at all in any of the 53 potential cases. The limited dataset and the zero occurrences prevent us from carrying out an in-depth analysis of the phenomenon for the variables studied in RQ3. Yet one interesting finding regarding these cases is QEII’s avoidance of intrusive /r/ altogether. She did not use it in any of the four word-internal positions described in Section 4.2.1, within affixed words (draw[ɹ]ings) or within compounds (awe[ɹ]-inspiring, law[ɹ]-abiding, Teresa[ɹ] of Calcutta).

Table 2 summarises the results for linking /r/, which show that out of 760 cases it was used 76.1% of the time. Glottalisation was the most common strategy when /r/-sandhi was not used (21.3%) and pure hiatus was rare (2.6%). Thus, these data allow us to confirm the two hypotheses entertained, that is, that linking /r/ would be more frequent than intrusive /r/ and that glottalisation would be typical when /r/-sandhi does not apply. One interesting result from the current study is the use of both linking /r/ and glottalisation in a few cases (3.4%). This finding shows that, although linking /r/ typically occurs without glottalisation on the one hand, and glottalisation without linking /r/ on the other, both phenomena can co-occur at potential /r/-sandhi sites.

Table 2

Rates of presence/absence of linking /r/ in potential cases (n = 760) with and without concomitant glottalisation (glott.) or hiatus.

The results reveal that potential word-external cases of linking /r/ are non-categorical. However, at word-internal sites (see Section 4.2.1.), linking /r/ is categorical. An inspection of 100 random cases confirmed this when a potential linking /r/ site was followed by a suffix, as well as in the three cases in which it was preceded by a prefix. As for compounds, it was found in 29 out of the 34 cases (84.3%). Since three tokens were instances of Siste[ɹ] Ethel, and these contain an honorific (Sister), a semantic category claimed to prevent /r/-sandhi (Lewis Reference Lewis1975), linking /r/ can be considered nearly categorical in compounds excluding those three cases, which leaves the rate of word-internal linking /r/ at 93.6%.

5.2 RQ2: Is QEII’s use of /r/-sandhi in line with speech community trends?

This question was first analysed from a longitudinal perspective. As pointed out in Mompean (Reference Mompean, Sasha, Paola, Marija and Paul2019), an adjusted (multiple) linear regression (ALR) showed that the variable year of production (YoP) was not a good predictor of linking /r/ (R ² = .21) in the group-level corpus, although some decrease over time could be observed (Figure 2).

Figure 2

Graph of a linear regression line: linking /r/ in a longitudinal corpus of RP speakers (rate per speaker). The line (trend line) represents the fitted values and the shaded bar represents the 95% CI. The small dots are actual data points. Each shape represents an individual in a given cohort: dots (WW1), triangles (WW2), squares (Silent), crosses (Baby boomers), barred squares (X), and asterisks (Millennials).

The same decrease can be observed applying GAM modelling to the data. In this type of model, the rate of linking /r/ production depends linearly on at least two smooth functions of some predictor variables, with each variable being added individually. In this study, the predictor variables were year of production (YoP) mean and year of birth (YoB) mean. These two variables were explored because they are often taken into account in real-time (i.e. YoP) and apparent-time (i.e. YoB) studies. Four GAM models were fitted using a smoothing spline for each of the covariates. The deviance was used to measure the difference between models.

The results are shown in Table 3 as well as in Figure 3a and Figure 3b, which contain three plots each: best fit, spline, and residual autocorrelation function (ACF). Figure 3a relates to YoP while Figure 3b relates to YoB. The figures show no autocorrelation in the residuals for a lag of twenty years, as indicated by the fact that the lines at the different lags are within the area circumscribed by the dashed lines. In the first model, the variable YoP was added first, which produced a significant model fit (p = .045) and an undulating, non-linear decrease in the use of linking /r/ across time. Notice, however, that the wiggle around the late 1990s/early 2000s might be due to the lack of data in that part of the curve, so the actual decrease could be even clearer if that part of the curve were not taken into account. Moreover, adding YoB did not result in a significant improvement in model fit (p = .154). Similarly, after applying a second GAM model and adding YoB first, a significant model fit was obtained (p = .039), again with some non-linear decrease in the use of linking /r/ across time; yet adding YoP the model did not result in a significant improvement in model fit (p = .41). The graphs reveal a decreasing trend in the use of linking /r/ over the years. Yet, since the data have high variability, the confidence intervals for the curves are wide, which suggests that further analyses should be conducted with a larger sample size.

Table 3

Analysis of two GAM models applied (linking /r/ production) in a longitudinal corpus of RP speakers.

^*p < .05. ^** p < .01. ^*** p < .001

Figure 3a

GAM modelling using YoP. Left: GAM plot displaying the best fit (line), 95% CI (shaded bar), and actual data points (dots, as in Figure 2) for YoP. Right (above): spline curve of the relationship between linking /r/ production and YoP as well as the 95% CI (dashed lines). Right (below): residual ACF plot for a lag of twenty years.

Figure 3b

GAM modelling using YoB. Left: GAM plot displaying the best fit (line), 95% CI (shaded bar), and actual data points (dots, as in Figure 2) for YoB. Right (above): spline curve of the relationship between linking /r/ production and YoB as well as the 95% CI (dashed lines). Right (below): residual ACF plot for a lag of twenty years.

As for QEII’s longitudinal pattern of linking /r/ usage, a decrease over the years was also observed. However, a simple linear regression revealed that YoP was not a good predictor of linking /r/ (R ² = .05), given that R ² values close to zero suggest that the model explains none of the variability. Since the range of cases per year in the QXB corpus is 2–33, and this may obscure any effect of YoP, a second linear regression was carried out excluding the years with fewer than eight (first quartile) potential cases (Figure 4). The results reveal that YoP is still not a predictor of linking /r/ production (R ² = .01).

Figure 4

Graphs of the rate of linking /r/ per year of production: all years with potential linking /r/ cases (left) and years only with eight or more cases (right).

A general comparison between the group-level and QEII’s data suggests that both appear to drift downward – i.e. use linking /r/ less over the years (see Figure 5). A closer inspection of QEII’s data with those of her cohort and gender group shows that QEII also conforms to their patterns. Kruskal–Wallis tests reveal no statistically significant differences in the mean rate of production across cohorts, including all cohort members or speakers analysed by gender, although females seem to produce less linking /r/ than males.

Figure 5

Mean rate of linking /r/ in cohorts including all cohort members (top) or by gender (bottom).

Given the data above, QEII can be said to belong to the Silent generation, for which demographers and researchers normally use mid-to-late 1920s and early-to-mid 1940s as starting and ending birth years, respectively. As Figure 5 reveals, speakers born in the Silent generation produced linking /r/ at a mean rate of around 75%, while the females of that generation produced it with a mean of above 70%. With a mean rate of 76%, QEII would then be considered to also fall within the normal range of linking /r/ for her generational and gender group.

Regarding intrusive /r/, the data are more difficult to interpret given the limited data from both the general corpus, its cohorts and the QXB corpus, with only 57 and 53 potential cases, respectively. They are also difficult to interpret because the potential cases in the general corpus are mostly limited to one or two cases per speaker and the data in both corpora are also irregularly spread along the time span. Yet the data from the general corpus are clear in showing that intrusive /r/ – produced at a rate of 21% (12 out of 57 cases) – was found in all cohorts except WWI (see Table 4). It is present, for example, in QEII’s generational cohort, with four actual cases out of 15 potential ones. With 53 potential cases but no single occurrence of intrusive /r/ in her Christmas messages, QEII clearly differs from what appears to be a general trend in the group-level corpus from the WW2 generation on, that is, a low but constant rate of intrusive /r/ usage.

Table 4

Potential and actual occurrences of intrusive /r/ in a longitudinal corpus of RP speakers by generational cohort: WW1, WW2, Silent, Baby Boomers (BB), X, and Millennials (Mill.). Potential/actual cases.

Another way of assessing whether QEII’s use of /r/-sandhi is in line with speech community trends is to compare her data with those of a population at a given point in time. Hannisdal (Reference Hannisdal2006) analyzed the speech of 15 male and 15 female middle-aged RP newscasters in samples produced between 2002 and 2004 (Hannisdal Reference Hannisdal2006: 246–247 for scores). Potential cases per speaker ranged from 162 to 275 for linking /r/ (M = 201.5) and from 8 to 43 for intrusive /r/ (M = 18.6). This dataset was used alongside QEII’s scores (as subject 31) to compute the 95% confidence interval for a proportion in one sample with continuity correction. The data from this analysis are shown in Table 5. The data show, for example, that the point estimate – or single value which serves as a best guess or best estimate of actual intrusive /r/ – was .326 (or 33%), and with 95% confidence the true estimate lies between .288 and .367 (or 28.8% and 36.7%). QEII, who produced 76.1% and 0% of linking /r/ and intrusive /r/, respectively, is not within the 95% confidence interval, a range of values that you can be 95% certain contains the mean of the population. Given this, QEII’s means cannot be considered to lie within the means of the population.

Table 5

Analysis of CI in a population of 30 RP speakers from Hannisdal (Reference Hannisdal2006) and QEII.

5.3 RQ3: What factors condition /r/-sandhi usage in QEII’s speech?

RQ3 addressed the nature of the factors potentially conditioning /r/-sandhi usage in the QXB corpus. Table 6 reports the data from the final (best-fitted) logistic regression model as well as the odds ratio.

Table 6

Best-fitted logistic regression analysis and odds ratio for predictors in the QXB corpus.

^*p < .05. ^** p < .01. ^*** p < .001

The OR data allow for the statement that, for example, the likelihood that linking /r/ will be produced when Stressed linking σ = 1 increases by a factor of 3.6 as opposed to when Stressed linking σ = 0. In terms of statistical significance, the results of the logistic regression show that the following variables influence the probability of success or likelihood of occurrence of linking /r/ (i.e. they are predictors):

Phonetic variables regarding:

1. (a) Tone units: QEII is less likely to produce linking /r/ when there is a TU boundary.

2. (b) Stress: QEII is more likely to produce linking /r/ when the linking syllable is stressed and she is less likely to produce linking /r/ when the linked syllable is stressed.

3. (c) Vowel quality: QEII is more likely to produce linking /r/ when the vowel quality of the linking syllable is open back (/ɑː/).

4. (d) Neighbouring /r/: QEII is less likely to produce linking /r/ when the linking syllable or the previous one contains /r/.

Usage-based variables regarding:

1. (e) Lexical token frequency: QEII is less likely to produce linking /r/ when the lexical frequency of the linked word is high (QXB).

2. (f) Lexical collocation frequency: QEII is more likely to produce linking /r/ when lexical collocation frequency (linking + linked words) is high (BNC).

3. (g) Vowel quality collocation frequency: QEII is more likely to produce linking /r/ when vowel quality collocation frequency (linking + linked vowels) is high (QXB).

The results obtained indicate that phonological and usage-based variables condition, to various degrees, linking /r/ production. They reveal, however, that the grammatical variables studied have no influence.

6.1 QEII’s /r/-sandhi resolution and individual vs. group-level trends

The fact that QEII uses a high proportion of linking /r/ is unsurprising given that linking /r/ is supposed to be the rule rather than the exception at potential /r/-sandhi sites. In contrast, although some use of intrusive /r/ was expected to occur in the corpus, even if at very low rates, the Queen’s lack of intrusive /r/ is interesting, although perhaps not too surprising as the latter has traditionally been stigmatised in RP.

The current study, as well as the author’s previous exploration, suggest some decrease in the use of linking /r/ over the years in the speech of both QEII and a reference group. Statistically, the effect is not very strong, so the question of whether /r/-sandhi is in a state of ongoing sound change or a case of stable variation needs further investigation. As a matter of fact, ‘there is lack of consensus on what it means for a variable to be “stable” over time … on how to measure and predict stability vs. change’ (Gardiner & Nagy Reference Gardiner and Naomi2017: 78–79), and it may be difficult to ascertain whether change is extremely slow or does not occur. Moreover, change may not occur in linear, progressive increments (Dannenberg Reference Dannenberg2000), as often assumed in apparent-time studies. Figure 3 above, for example, shows a non-linear decrease in the use of linking /r/ over the years and not a steady, linear decrease. Rates seem to rise, for example, between 1940 and 1960 before they start to decrease again. Finally, as pointed out in Section 5.2, the data have high variability and, in the case of intrusive /r/, are also rather limited. Thus, further real-time analyses should be conducted that include more data points at multiple points across time. In the case of the reference corpus, this could include more potential cases per speaker, more years, more speakers or a combination of these. As for QEII, further material from other formal speeches could be gathered (e.g. State Opening of Parliament, formal reception speeches, etc.).

Irrespective of the issue of stable variation vs. sound change in progress, the differences in rates of production across studies reveal the existence of a wide range of inter-speaker and intra-speaker variation in the use of /r/-sandhi. This is likely to happen in the case of linking /r/ and that of intrusive /r/. Table 7 shows some rates in three recent corpus-based studies five years apart: Hannisdal (Reference Hannisdal2006) on TV newsreaders, Mompean & Gómez (Reference Mompean, Alberto, Wai-Sum and Eric2011) on radio correspondents, and Pavlík (Reference Pavlík2016) on radio newscasters. These studies provide rates of linking /r/ lower than QEII’s in the current study (i.e. 76.3%), which seems to point to the fact that QEII may be more of an /r/-linker than the average RP speaker, at least in the case of linking /r/.

Table 7

Rates of linking /r/ in potential cases of linking /r/ and intrusive /r/ in three recent studies.

Hannisdal’s (Reference Hannisdal2006) study is particularly useful here. In that study, 25 speakers (i.e. 83.3%) had rates below 70%, with 11 speakers in the 50%–59% range (the group’s mean was 59.8%). Only three speakers had rates between 70% and 75% and only two between 76% and 80%. As reported in Section 5.2, QEII’s rate is not within the 95% confidence interval.

The findings from the three corpus-based studies show that intrusive /r/ is relatively common in contemporary RP, ranging from 20% to 40% depending on the study. Yet QEII does not use it, despite its apparently greater acceptance in contemporary RP (Wells Reference Wells1999). As in the case of linking /r/, there is also high inter-speaker variation in the use of intrusive /r/. The data from Hannisdal (Reference Hannisdal2006), for example, show rates of 0% (3 speakers), 5%–10% (3 speakers), 10%–30% (11 speakers), 30%–50% (8 speakers), 51%–70% (3 speakers) and 71%–85% (2 speakers). As reported in Section 5.2, QEII’s rate of intrusive /r/ is not within the 95% confidence interval.

Similar inter-speaker variation in the use of intrusive /r/ has also been found in NZE (Hay & Warren Reference Hay and Paul2002, Viollain Reference Viollain2014). QEII could then be part of a minority of RP speakers that does not use intrusive /r/ (Wells Reference Wells1982), at least in formal public speech. Related to this explanation, it has been pointed out that some speakers also often avoid intrusive /r/ altogether in experimental conditions. In their study of intrusive /r/ in New Zealand English, for example, Hay & Warren (Reference Hay and Paul2002) pointed out that their 16 participants, who read 48 sentences targeting 48 examples of intrusive /r/, ranged from 0% to 77%, with a median of 21% and a mean of 29%. Interestingly, one subject produced no tokens of intrusive /r/. According to Hay & Warren, intrusive /r/ was ‘likely to be much lower than would occur in natural speech’ (2002: 49).

As to why QEII avoids intrusive /r/ altogether, the formal register of the broadcasts cannot explain by itself all the facts, as the studies referred to above (i.e. Hannisdal Reference Hannisdal2006, Mompean & Gómez 2011, Pavlík Reference Pavlík2016) have shown that speakers produce intrusive /r/ in formal contexts (e.g. newscasting). In contrast, gender cannot be overlooked as a possible explanation for some of the inter-speaker variation observed in the studies. Most studies on /r/-sandhi in RP (Hannisdal Reference Hannisdal2006, Mompean & Mompean-Guillamón Reference Mompean and Mompean-Guillamón2009, Pavlík Reference Pavlík2016) have found no general gender effect. However, the three speakers in Hannisdal’s study who avoided intrusive /r/ altogether were females. Related to this, it has been claimed that, in many speech communities, female speakers use a higher proportion of prestige forms than male speakers and tend to use fewer stigmatised forms than men, particularly in formal speech (Coates Reference Coates2004). It should be borne in mind that intrusive /r/ has traditionally been regarded as a vulgarism (Crystal Reference Crystal and David1984) and it may still carry some degree of stigmatisation in RP. This stigmatisation may be apparent when attempting to elicit intrusive /r/ in experimental studies, as pointed out above, or when speaking to an audience. In this latter case, the Queen might, therefore, use a high proportion of prestige forms (linking /r/) and no stigmatised forms (intrusive /r/) in her (surely very) speech-conscious delivery of her broadcasts. It could also be mentioned, as possible evidence of the prestige of linking /r/, that the latter has been claimed to be a feature of colloquial style and not so much of scripted speech in non-rhotic English (Hannisdal Reference Hannisdal2006, Navarro Reference Navarro2013, Viollain Reference Viollain2014). However, the prestige of linking /r/ for the Queen may be such that she uses it even in formal speeches. Alternatively, as pointed out by one reviewer, what might be prestigious could be to know and apply the difference between linking /r/ – possibly neutral – and intrusive /r/ – somewhat stigmatised. This difference might be difficult to maintain for some speakers who, trying to avoid intrusive /r/ in formal contexts, may also tend to suppress linking /r/ as some sort of hypercorrection effect (Wells Reference Wells1982, Brown Reference Brown1988).

Another possible explanation of QEII’s avoidance of intrusive /r/ could be related to her personal role. She has a unique status as both Queen of the United Kingdom and other Commonwealth states (political status) and a symbol/stereotype of her accentual variety (sociolinguistic status), which is often referred to in connection with herself as head of state (i.e. the Queen’s English). During the Christmas broadcasts, then, she projects a public persona (the Queen) who is – as another reviewer points out – professionally careful in what she says, so it would not be surprising if she were always careful in how she says it, particularly when addressing such a very wide audience (the UK and the Commonwealth) in a very formal context. Given the emphasis on stylistic practice in third-wave approaches to sociolinguistic variation, which consider speakers as stylistic agents who place themselves in the social landscape through their stylistic practice (Eckert Reference Eckert2012), QEII’s avoidance of intrusive /r/ could be a stylistic resource adding a layer of social meaning to her speech and indexing professional carefulness and attention to detail as well as general standardness.

A comparison of QEII’s /r/-sandhi usage patterns with the studies mentioned above as well as with the data from the longitudinal group-level corpus and Hannisdal’s (Reference Hannisdal2006) study points to a two-way pattern. On the one hand, QEII seems to be a more-than-average /r/-linker (linking /r/) who shares the apparently decreasing use of linking /r/ on a longitudinal dimension with the community’s trend. On the other hand, QEII seems to differ from the relatively low yet constant use of intrusive /r/ on a longitudinal dimension and also at a given point in time, as revealed by various corpus-based studies. This finding suggests that the same individual can be more or less in line and at the same time differ from group-level trends or norms depending on the linguistic feature analysed.

There are many reasons why individuals resemble or differ from such trends or norms in their language choices. In the case of intrusive /r/, as pointed out above, stylistic agency and/or the underlying prescriptivism of a language ideology may explain her avoidance of the phenomenon. Language ideologies – or sets of beliefs about languages as used in social contexts – often play a role in shaping and influencing linguistic structures and speech forms (Silverstein Reference Silverstein, Clyne, Hanks and Hofbauer1979). Prescriptivism can lead to a static situation in which a change in progress does not go to completion, is prevented from developing entirely, or is excluded from the standard forms of the language (Hickey Reference Hickey2012). If language users think, for example, that sounds not represented in spelling should not be pronounced (e.g. an r-sound in I saw[ɹ] it), their pronunciation is likely to reflect it eventually. Intrusive /r/ in RP and many other non-rhotic varieties seems to be in that situation. It is found across language cohorts and over decades, but never quite reaches the rate of linking /r/. At the level of individuals such as QEII, a prescriptivist language ideology strictly adhered to, apparently, suppresses the phenomenon altogether, therefore setting her aside from the general trend. It should be pointed out, at this stage, that language ideology regarding intrusive /r/ may not be the same across all non-rhotic groups of speakers. Some studies have shown that rates of intrusive /r/ are typically lower among speakers with a higher socioeconomic status/education in NZE (Hay & Warren Reference Hay and Paul2002, Hay & Sudbury Reference Hay and Andrea2005, Hay & Maclagan Reference Hay, Margaret, Preston and Nancy2010) or Singaporean English (Tan Reference Tan2012). The RP speakers in the general corpus and QEII are middle- to upper-class speakers, and their use of intrusive /r/ seems to fit this pattern. Interestingly, however, Foulkes (Reference Foulkes1997) found that intrusive /r/ was more frequent in Newcastle among working-class than among middle-class speakers in spontaneous conversation but not in reading passage style. Since intrusive /r/ is not a salient feature of the local Newcastle dialect, Foulkes speculated that intrusive /r/ could be seen as a non-regional feature often heard in the media and perceived as prestigious by middle-class speakers.

A noticeable finding in this study is how glottalisation is used to resolve hiatus at potential /r/-sandhi sites. Glottalisation can be described as a hiatus-breaking strategy when it occurs in isolation, which is attested at a non-negligible rate (over 20%). This finding is in line with a similar tendency in a corpus of modern BBC newscasts (Mompean & Gómez Reference Mompean, Alberto, Wai-Sum and Eric2011), where glottalisation was used in 31.3% of all potential /r/-sandhi sites. The role of glottalisation as a hiatus-breaker at word boundaries is also confirmed by similar findings from a recent corpus-based study on /r/-sandhi in Australian English (AusE) by Cox et al. (Reference Cox, Palethorpe, Buckley and Bentink2014) and by another study on a related external sandhi phenomenon known as glide insertion (Davidson & Erker Reference Davidson and Erker2014). As for Cox et al.’s (Reference Cox, Palethorpe, Buckley and Bentink2014) study, it is interesting to note that the rates found in that study and in the current one are remarkably similar (see Table 5), although QEII’s variety and Australian English are different – yet apparently similar in the case of /r/-sandhi. As Table 8 shows, glottalisation was used as the single hiatus-breaking strategy in over 20% of the potential cases.

Table 8

Rates of linking /r/ (with and without concomitant glottalisation) as well as rates of lack of linking /r/ (+glottalisation, hiatus) in potential /r/-sandhi sites: Current study (QXB) and Cox et al. (Reference Cox, Palethorpe, Buckley and Bentink2014).

Evidence on glottalisation as a more general hiatus-breaking strategy comes from a study on a sandhi phenomenon often referred to as glide insertion in V₁ [+high] + V₂ contexts, with the glide being [^j] if the high vowel is front (e.g. may I (e.g. [meɪ^jaɪ]) and [^w] if the high vowel is back (e.g. go inside [ɡəʊ^wɪnsaɪd] ). In their lab-based study of glide insertion in American English, Davidson & Erker (Reference Davidson and Erker2014) found that potential cases of glide insertion across word boundaries were actually resolved with glottalisation. This happened in nearly half of the cases, with a higher rate when the linked vowel was stressed. Although glide insertion was not the object of study in the current research, a preliminary inspection was carried out of 50 potential glide insertion contexts in the QXB corpus. This confirmed that QEII also uses glottalisation in such contexts (around 50% of the time), as shown in Figure 6, both when the second vowel is stressed (e.g. to all) or unstressed (e.g. obviously in).

Figure 6

Examples of glottalisation at potential glide insertion contexts when the second vowel is stressed (left) and unstressed (right).

As to why glottalisation may override /r/-sandhi, it could be argued that glottalisation is less marked than coronal articulations (Lombardi Reference Lombardi2002) and that it tends, therefore, to replace /r/-sandhi in areas with significant language and sociocultural contact (Britain & Fox Reference Britain, Susan, Markku, Juhani and Heli2009, Fox Reference Fox2015) and in first language acquisition, where glottalisation seems to be the default form in children’s English until forms are acquired at a relatively later stage (Newton & Wells Reference Newton and Bill2002). Thus, although some have tried to provide theoretical explanations of /r/-sandhi as a natural process and /r/ as a good sandhi segment (e.g. Broadbent Reference Broadbent1991, Orgun Reference Orgun2001, Lombardi Reference Lombardi2002, Uffmann Reference Uffmann2007), /r/-sandhi may, after all, be a historical artefact (Britain & Fox Reference Britain, Susan, Markku, Juhani and Heli2009), a synchronically arbitrary process with rare equivalents in the world’s languages (e.g. Blevins Reference Blevins and Jeff2008, Hall Reference Hall2013). Glottalisation might, therefore, represent a more ‘natural’ strategy to prevent hiatus. Increasing use of glottalisation could be, then, one reason for the decrease in the use of linking /r/ noted in various studies (Allerton Reference Allerton2000).

Irrespective of the motivation for the use of glottalisation instead of /r/-sandhi, studies such as Cox et al. (Reference Cox, Palethorpe, Buckley and Bentink2014) or the current study have implications in our understanding hiatus resolution in /r/-sandhi in non-rhotic accents. More specifically, the finding points to a more complex system of hiatus resolution at both speech community and individual levels than that typically modelled by most phonological accounts (see Section 6.2 below). In these models, /r/-sandhi is assumed to be categorical. Faced with the evidence that /r/-sandhi is not a categorical phenomenon, hiatus might seem to be the default value. However, hiatus is rare and glottalisation is the norm when /r/-sandhi is not used. To complicate matters even more, /r/-sandhi and glottalisation can also co-occur, although this co-occurrence is further discussed below.

The study has also shown that /r/-sandhi and glottalisation co-occur at external /r/-sandhi site, although at a relatively marginal rate. Figure 7 shows how linking /r/ and glottalisation co-occur at linking /r/ sites followed by stressed (e.g. to their own) and unstressed (e.g. desire of) syllables. The examples involve creaky voice but not full glottal stops, which are rare in QEII’s speech in these contexts.

Figure 7

Simultaneous glottalisation and linking /r/ at sites followed by stressed (left) and unstressed (right) syllables.

To the best of our knowledge, this combination of /r/-sandhi and glottalisation only seems to have been found previously by Cox et al. (Reference Cox, Palethorpe, Buckley and Bentink2014) in their study on contemporary AusE, which also found the combination to occur at a very low rate (3.4%). An interpretation of such a finding, however, is necessary. The use of linking /r/ alongside glottalisation could be interpreted as a combination of two strategies to resolve hiatus at potential /r/-sandhi sites. Yet the glottalisation observed alongside /r/-sandhi might be a word-boundary phenomenon that has nothing to do with an attempt to break hiatus. To start with, the co-occurrence of /r/-sandhi and glottalisation seems to be rare, as revealed by the very low rates in both the current study and in Cox et al.’s study. If glottalisation reinforced /r/-sandhi in an attempt to avoid hiatus its frequency would have been expected to be much higher. However, the low frequency of the combination may be one reason why it has been unnoticed until now. Some recent work has specifically addressed the possible use of /r/-sandhi, glottalisation, or both in L1 English learners of L2 German. Bissiri & Scobbie (Reference Bissiri and Scobbie2016, Reference Bissiri and Scobbie2017) carried out both acoustic and ultrasound recordings of four such learners producing English and German sentences with potential /r/-sandhi contexts (e.g. cider apples; jeder Abend ‘every evening’). The researchers were interested in the possible role of the presence/absence of a phrase boundary at /r/-sites and the stress/lack of stress in the linked vowel. Moreover, they were also interested in whether there was any transfer of glottalisation and /r/-sandhi from learners’ L1 (English) to their L2 (German). Bissiri and Scobbie’s preliminary findings revealed that /r/-sandhi was only articulated in L1 English without an intervening phrase boundary and when the linked vowel was deaccented. /r/-sandhi was also absent in L2 German. Interestingly, no overlapping of /r/-sandhi and glottalisation was found in L1 English. Glottalisation was found, however, at phrase boundaries and before accented linked vowels, and more frequently in G2 German than in L1 English.

One manner of checking whether the glottalisation occurring alongside /r/-sandhi does not specifically serve a hiatus-breaking purpose would be to inspect whether glottalisation also occurs in positions that attract creaky voice at phrase-internal word boundaries. Previous studies (e.g. Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014, Yuen et al. Reference Yuen, Cox and Demuth2015, Pavlík Reference Pavlík2016) and the current one have found that /r/-sandhi is disfavoured when the second hiatus vowel is stressed. In many such cases, however, glottalisation is used. Yet glottalisation seems a more general phenomenon at phrase-internal word boundaries. A case in point is word-initial glottalisation (Garellek Reference Garellek2012) or the phenomenon whereby word-initial stressed and/or accented vowels are preceded by a glottal stop and/or have creaky voice (e.g. Nakatani & Dukes Reference Nakatani and Dukes1977, Umeda Reference Umeda1978), irrespective of whether the segment before the word boundary is a consonant or a vowel or whether there is a pause. Research has found that word-initial glottalisation occurs with higher frequency at the beginning of tone unit boundaries and to a greater degree if the word is pitch-accented (Dilley, Shattuck-Hufnagel & Ostendorf Reference Dilley, Shattuck-Hufnagel and Ostendorf1996). Yet, the segmental context also plays a role, with glottalisation occurring significantly more frequently after vowels – particularly back vowels – and liquids at phrase-internal word boundaries (Umeda Reference Umeda1978, Dilley et al. Reference Dilley, Shattuck-Hufnagel and Ostendorf1996). This finding is consistent with an Electopalatography (EPG) study by Scobbie & Pouplier (Reference Scobbie and Marianne2010) on word-final /l/ produced by speakers of SSBE (or RP) and Scottish-accented Standard English. The researchers found that speakers from both groups often used glottalisation alongside /l/ in word-final contexts followed by a vowel-initial, lexically stressed syllable in an accented word. In the case of QEII, limitations of space prevent a thorough analysis of her use of glottalisation in the corpus under analysis. However, an inspection of 50 tokens of V₁/l/ + V₂ with an intervening phrase-internal word boundary revealed that glottalisation was common alongside /l/ followed by stressed (e.g. will always) or unstressed (e.g. all a) vowels. Figure 8 shows some examples with /l/ followed by a stressed or an unstressed vowel.

Since word-final /l/ and /r/-sandhi would already prevent hiatus, the concomitant presence of glottalisation could serve some prosodic function given that glottalisation is used for various prosodic and communicative functions (see e.g. Redi & Shattuck-Hufnagel Reference Redi and Stefanie2001). Some of these functions could be, depending on the case, cuing open juncture, marking the boundary between two rhythmic feet or between two intonation units, or a combination of these. In other words, glottalisation alongside /r/-sandhi may be acting as an external boundary signal rather than simply as a hiatus-breaking strategy. Glottalisation without /r/-sandhi, however, can be considered as a hiatus-breaking strategy, although it may also be performing other prosodic functions.

Figure 8

Glottalisation at V₁/l/ + V₂ contexts in QEII: stressed V₂ (left) and unstressed V₂ (right).

6.2 QEII and factors conditioning /r/-sandhi usage

The results confirm the hypotheses that phonetic and usage-based variables condition /r/-sandhi usage. The hypothesis that grammatical variables such as syntactic boundaries and word class condition /r/-sandhi usage is, however, not confirmed.

As regards phonetic factors, the study has shown that the absence of a TU boundary predicts the use of linking /r/. This is in line with suggestions that /r/-sandhi is more frequent within phonological phrases than phrase-finally (Joseph Reference Joseph, Bohumil, Osamu and Joseph1999) or within clauses than across clause boundaries (Foulkes Reference Foulkes1997), which may effectively mark prosodic boundaries. Besides, the results suggest that the presence of a stressed linking syllable and the absence of stress in the linked syllable facilitate linking /r/. These results indicate that a prototypical example of linking /r/ would be in an SW prosodic site and least likely of all in a WS site. This result is in line with Cox et al.’s (Reference Cox, Palethorpe, Buckley and Bentink2014) finding that the likelihood of occurrence of linking /r/ is higher in SW prosodic sites (86% in her study) as opposed to WS sites (49%). These findings suggest that that prosodic and metrical effects, found to be critical in the use of /r/-sandhi at group level (Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014), also occur at the idiolectal level.

While the influence of prosodic factors is clear in this study, the contribution of segmental ones is more modest. The logistic regression analysis suggests that the presence of /r/ one syllable before the linking syllable prevents the use of linking /r/, pointing to a dissimilation effect. This contrasts with previous studies such as Pavlík’s (Reference Pavlík2016), as well as Mompean & Mompean-Guillamón’s (Reference Mompean and Mompean-Guillamón2009), which found little or no effect of a neighbouring /r/ in the production of linking /r/ at all in RP. However, our findings coincide with those reported by Hannisdal (Reference Hannisdal2006), who found that the presence of /r/ in the vicinity inhibited /r/-sandhi. Finally, our analysis suggests that the presence of an open back vowel /ɑː/ in the previous syllable facilitates the use of linking /r/. However, no other linking or linked vowel quality category had a significant effect. Thus, the findings obtained in this study seem to suggest that prosodic factors are better predictors of /r/-sandhi usage than segmental ones, at least in QEII’s sample.

The results also show that grammatical factors seem to play little or no role in the production of linking /r/. The hypothesis formulated in this respect, that is, that phonological factors (TU boundary, stress, neighbouring /r/, vowel quality) would have a greater effect than grammatical factors (syntactic boundaries and word class) on /r/-sandhi is confirmed. In this respect, linking /r/ may be avoided when syntactic boundaries coincide with TU boundaries but less likely when they do not. As to the lexical vs. grammatical nature of the linking and linked words, the results obtained are in line with the absence of an effect described by Hay & Sudbury (Reference Hay and Andrea2005) or the little effect found by Pavlík (Reference Pavlík2016). It may be the case that the perception that function words could favour /r/-sandhi can simply be an effect of prosodic stress and the use of schwa in function words, which are typically unstressed. It seems, therefore, that prosodic structure predicts /r/-sandhi better than syntactic structure or grammatical information.

The fact that prosodic factors are better predictors of /r/-sandhi usage than segmental ones in the QXB corpus underscores the relevance of prosodic phrasal context in /r/-sandhi usage. This context may also explain rate of speech and coarticulation effects. As discussed below regarding the phonological modelling of /r/-sandhi, the latter is often modelled as a segmental process. However, the phenomenon can also be viewed as a type of gradient hyper- and hypo-articulation (H&H) or overshoot/undershoot alternation. It was mentioned above (Sections 2.3, 4.2.2) that /r/ displays gradient features in its realisation. Moreover, speech rate might also determine /r/-sandhi production. It has been pointed out, for example, that linking /r/ is less common at both rather markedly rapid and slow rates (Lewis Reference Lewis1975). Pavlík (Reference Pavlík2016), for example, found a small effect of speech rate on /r/-sandhi production, with a decrease in linking /r/ as speech rate increased. This could be due to the fact that fast speech is characterised by high articulator velocity and gestural overlap. A reviewer suggests that this might be indicative of lack of (post-lexical) phonetic planning in more citation-like, slower speech with frequent pauses vs. undershoot in faster speech. QEII’s Christmas broadcasts seem to be delivered at neither a noticeably slow nor a particularly fast speech rate (i.e. average tempo). Rate of speech was not investigated in the current study. However, future studies should look at the possible influence of rate of speech in the same speaker as well as the possible gradient timing/gestural correlations.

Turning now to the usage-based variables, the results indicate a clear effect of frequency on the use of linking /r/. Yet they reveal that collocational frequency is more relevant than the individual lexical or vowel quality frequencies of the words or vowels, respectively, at either side of /r/-sandhi contexts. The results indicate that QEII is less likely to produce /r/ when the lexical frequency of the linked word is high. This coincides with the results obtained by Hay & Sudbury (Reference Hay and Andrea2005). However, the fact that the lexical frequency of the linking word did not have an influence on linking /r/, and that the frequency of the linked word had only a minimal effect (only in the first analysis and in the QXB corpus) indicates that individual lexical frequencies are probably not an important predictor of /r/-sandhi usage.

The above-mentioned finding is in line with that of a study on the influence of a range of phonological, morphological and lexical frequency variables on the use of t/d-deletion in Canadian English by Walker (Reference Walker2012). He found that higher lexical frequency correlated with higher deletion rates. Yet, when several factors were taken into account, the effect of lexical frequency was non-significant. Given the existence of a body of work that suggests a major role for frequency in phonology (e.g. Bybee Reference Bybee2001), Walker pointed out that frequency effects might lie beyond the word, suggesting collocations as a possible locus for those effects. The current study confirms this: collocational frequency had the highest predictive power, with higher collocational frequencies favouring linking /r/ in both the BNC and the QXB corpora. This may not be surprising given that, as discussed by Biber and colleagues (Biber, Conrad & Cortes Reference Biber, Conrad and Cortes2004), frequent collocations are often used with some pragmatic or discourse function (e.g. thematisation or construction of referential themes such as the year’s events and QEII’s wishes for the following year (e.g. this year I …, the year ahead …). Thus, they help speakers (in this case QEII) perform discourse functions. Moreover, they play an important part in the speaker’s communicative repertoire and may lexicalise – or entrench – phenomena such as /r/-sandhi, giving rise to higher rates, as is the case, for example, of word-internal linking /r/ in compounds. From the point of view of exemplar theory, frequent collocations with a potential /r/-sandhi context have more chances of being produced with /r/. If /r/ is actually produced relatively frequently, speakers’ mental representation will contain many memories of that collocation produced with /r/. When those memories are averaged over for a production target, in turn, a robust /r/ will result (Hay & Maclagan Reference Hay, Margaret, Preston and Nancy2010, Reference Hay and Margaret2012, Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014). Alternatively, the higher frequency of linking /r/ in frequent collocations can be interpreted ‘as a remnant of the historical /r/ which continues to resist the /r/-loss in British English’, which is still in progress (Pavlík Reference Pavlík2016: 112). As Bybee (Reference Bybee2001, Reference Bybee2006) points out, high frequencies encourage phonetic change but they render items more conservative in the face of grammatical or analogical change. As a case in point, most verbs that are still irregular in contemporary English (e.g. say, go, get, etc.) have a high lexical frequency, the latter strengthening their representation in memory and helping them resist the pressure from analogical change.

In addition to lexical collocation frequency, and to the best of our knowledge, this is the first study that describes the role of vowel quality collocational frequency, showing that higher frequencies had a positive effect on the rate of linking /r/. The most frequent vowel quality collocations were mid central + mid central (e.g. /ə.ə/), mid central + close front (e.g. /ə.ɪ/), mid central + open central/back (e.g. /ə.ɑː/), mid central + close-mid/open back (e.g. /ə.ɔː/), open back + mid central (e.g. /ɑː.ə/), close-mid back + mid central (e.g. /ɔː.ə/), and mid central + mid/open front (e.g. /ə.e/). As can be seen, mid central vowels (particularly schwa and centring diphthongs) are involved in vowel quality collocations favouring linking /r/. Collocations that did not contain mid central vowels had no favouring effect on linking /r/. Finally, it should be pointed out that vowel type frequency did not have an effect on the use of linking /r/, which is consistent with the hypothesis that collocational frequency is more relevant than individual vowel frequencies.

The findings in relation to usage-based variables provide arguments for the debate regarding the phonological modelling of /r/-sandhi in non-rhotic English. Most traditional analyses assume that no underlying coda /r/ is present word-finally and that /r/-sandhi arises as a function of /r/-insertion (e.g. Vennemann Reference Vennemann1972, Johansson Reference Johansson1973). Other approaches assume that underlying coda /r/ is present but that /r/-deletion operates before a consonant or pause (e.g. Mohanan Reference Mohanan1986, Harris Reference Harris1994). Yet other accounts claim that both rules (/r/-insertion and /r/-deletion) coexist in the synchronic phonology of varieties (e.g. Anttila & Cho Reference Anttila and Cho1998, Kijak Reference Kijak2009). All these approaches aim at constructing synchronic models of competence and exclude considerations of function and usage, so they cannot predict inter-/intra-speaker variation.

An alternative view is to consider lexical representations as being modelled as exemplars of remembered sequences (e.g. Bybee Reference Bybee2001) and that surface forms are not generated from underlying representations. Within this more functional, usage-based framework, every linguistic item in the lexicon consists of a cloud of phonetic exemplars containing detailed information about the context in which they are used. In this view, /r/-sandhi tokens would not be the result of deletion or insertion rules, but rather the result of how entrenched or lexicalised /r/ is in specific utterances (Hay & Sudbury Reference Hay and Andrea2005, Mompean & Mompeán-Guillamón Reference Mompean and Mompean-Guillamón2009, Hay & Maclagan Reference Hay and Margaret2012, Sóskuthy Reference Sóskuthy2013, Pavlík Reference Pavlík2016). The more frequent a collocation including a potential /r/-sandhi is, the more likely it is that /r/ will become lexicalised in it and thus be produced. The usage-based factors described in this study can be interpreted as evidence in favour of a usage-based framework. However, this interpretation does not rule out the possibility of a hybrid account including abstract representations, exemplars and phonetic implementation processes as suggested by some researchers (e.g. Cox et al. Reference Cox, Palethorpe, Buckley and Bentink2014).

More generally, a usage-based interpretation of the findings also allows accommodating not only inter-group but also inter-speaker variation within the same group, given that language users differ in the quality and quantity of input they receive, leading to overlapping grammars that nevertheless differ in some respects. A functional, usage-based interpretation can also accommodate differences in the relative weight of some predictors when comparing individuals and group-level trends and, possibly, any effects of the individual’s agency on his/her production. One interesting conclusion from the study of the factors conditioning /r/-sandhi usage in QEII is that the multiple variables contributing to variation in a given phenomenon – which studies at the group-level typically show – can also be seen at the individual level. Yet the weight of relevant variables may be different across individuals given various idiosyncrasies as well as their different linguistic experience.