Although this study is concerned with context-specific /s/-retraction, there is of course a wealth of evidence highlighting the degree of within- and between-speaker variability in the production of sibilants more generally and the factors that condition these patterns of synchronic variation (Newman et al. 2001; Stuart-Smith 2007; 2020; Levon et al. 2017). This is particularly the case for /s/, in which variation has been shown to have taken on socio-indexical meaning: more fronted /s/ realisations are perceived as less masculine and more gay and, in production, are more frequent among male speakers who are gay or bisexual (Campbell-Kibler 2011; Podesva & Van Hofwegen 2014; Levon 2014).

However, it is important to separate the realisation of /s/ in /stɹ/ from the wider spectral variation in /s/ production. /stɹ/ has been studied as a sociolinguistic variable in its own right, separate from the wider patterns of socio-indexical variation in /s/ found in other environments (see e.g. Durian 2007; Gylfadottir 2015; Wilbanks 2017). There is also strong perceptual evidence to support this distinction: Phillips & Resnick (2018; 2019) find that a retracted /s/ in /stɹ/ clusters does not carry the same socio-indexical meaning of masculinity, toughness or heterosexuality that has been observed for retracted /s/ in other /sCɹ/ environments.

There is a general lack of large-scale studies of /s/-retraction that combine robust acoustic analysis with community-level data in order to investigate its status within a given speech community in detail, especially with a view to investigating change. Existing work is also variable with respect to the method of coding /s/-retraction. Some studies have adopted a binary classification, coding tokens of /s/ as either retracted or non-retracted (e.g. Janda & Joseph 2003; Bass 2009); while this finds some support from a study by Rutter (2011), who reports that a majority of retracted forms fall within a speaker’s normal range for [ʃ], Labov (2001) argues that, in Philadelphia English, there are at least four variants differing in how [ʃ]-like they are. A fine-grained acoustic measure of retraction is vital given that this process is also argued to occur, though to a much lesser degree, in other contexts such as pre-vocalic /sp, st, sk/ and /spɹ, skɹ/ (Labov 1984; Janda & Joseph 2003; Baker et al. 2011).

The few large-scale studies that have been conducted concern almost exclusively varieties of English spoken in North America (e.g. Gylfadottir 2015 in Philadelphia, PA and Wilbanks 2017 in Raleigh, NC) and all show evidence of apparent-time change with increasing /s/-retraction among younger speakers.1 However, the role of other social factors in explaining variation in retraction, such as gender and socioeconomic status, is less consistent. Durian (2006) and Bass (2009) describe “rapid anonymous” surveys, in which /s/ is auditorily coded, but the former finds a female lead for /s/-retraction in American English (Columbus, OH) and the latter a male lead in British English (Colchester). Larger-scale studies based on acoustic data are similarly variable, finding either that women are leading in this change (Wilbanks 2017; Stuart-Smith et al. 2019) or detecting no gender effect at all (Gylfadottir 2015). Additionally, while Labov (2001) finds an association between /s/-retraction and working-class speech in subjective evaluation tests, there is not yet any clear evidence of a relationship between /s/-retraction and social class on the basis of actual production data. Although we plan to address these topics in future work that considers the wider status of /s/-retraction in the speech community, and we do discuss these factors briefly in Section 4 since they are included in the model, the primary focus of this paper is on the mechanisms by which /s/ undergoes retraction in this environment.

As discussed in Section 1, there are two major proposals that aim to explain why /s/ retracts in these contexts. These differ from one another with respect to the locality of the triggering mechanism:

A fuller consideration of these competing explanations follows.

The first of these competing explanations, that retraction is caused directly by the /ɹ/ in these /stɹ/ clusters, was initially proposed by Shapiro (1995). This argument was motivated by the observation that /s/ retracts to a significantly lesser degree in /st/ clusters, e.g. steep, suggesting that the rhotic segment in /stɹ/ plays a crucial role and that the retraction process is therefore a case of assimilation “at a distance”. The presence of only low-level retraction in /st/ and other complex onsets has been widely attested and even varieties that are reportedly not undergoing the /stɹ/ change itself, such as Australian English, still show a slightly lower centroid frequency for /s/ in contexts such as /sp, st, sk/, described as the “phonetic pre-conditions” of the change in /stɹ/ (Stevens and Harrington 2016: 118).

While Shapiro (1995) relies on the inspection of secondary data and anecdotal evidence, later studies provide a more strongly quantitative and empirical support for these claims. An articulatory experiment by Baker et al. (2011), in which ultrasound tongue imaging is combined with simultaneous acoustic analysis, finds evidence of a coarticulatory bias towards retraction in other /sCɹ/ clusters in American English. In their study, /s/ is produced with a slight dampening of centre of gravity in all complex onset clusters, with a stronger lowering in centroid frequency observed when the complex onset contains /ɹ/, as in words such as spring and screech. Although the effect is registered most strongly in /stɹ/, which is clearly set apart from all other onset types, this result suggests that /ɹ/ has some long-distance lowering effect on the frequency of /s/ regardless of the intervening consonant.

Another line of argumentation provided by Baker et al. (2011) draws upon the inherent variability in the articulation of /ɹ/. It is well established that /ɹ/ can be produced with a range of lingual constriction types, ranging from more “tip-up” retroflex to more “tip-down” bunched articulations, often with little to no perceptible acoustic difference between them (Delattre & Freeman 1968; Twist et al. 2007). Baker et al. (2011) report that all speakers in their study used a bunched /ɹ/ variant in /stɹ/, but with variation within this category with respect to the similarity of the lingual shape in the articulations of /s/ and /ɹ/ in these clusters. Among speakers they classify as “non-retractors”, for whom retraction in /stɹ/ stems from gradient coarticulation rather than a distinct production target, the magnitude of retraction correlates with this inter-speaker variation in the tongue shape for /ɹ/ and specifically its similarity to the tongue shape for /s/.

It is also interesting to note the claim by Baker et al. (2011) that /stɹ/ clusters favour a bunched rather than retroflex tongue configuration for /ɹ/ as there is independent evidence (albeit from British, rather than American English) that bunched /ɹ/ is accompanied by more extreme lip protrusion relative to retroflex /ɹ/ (King & Ferragne 2020). This might suggest that any role of /ɹ/ may in part be due to regressive assimilation of labialisation and less so the lingual gesture, a possibility also raised by Janda & Joseph (2003: fn. 8) in a discussion of cross-linguistic differences in anticipatory labialisation.

One final piece of suggestive evidence regarding the role of /ɹ/ comes from a small-scale study of /s/-retraction in British English: Sollgan (2013) reports variability in tongue shape for /ɹ/ among speakers of Edinburgh English and, crucially, observes that alveolar realisations of /ɹ/ rarely co-occur with a retracted /s/ variant. Based on this, one might argue that there is a close relationship between the tongue shape of /s/ and of the rhotic segment in these onset clusters, mirroring the aforementioned observation in the Baker et al. (2011) study.

Competing accounts of /s/-retraction have proposed that affricated /tɹ/ clusters are responsible for the retraction of a preceding /s/. This argument was first made by Lawrence (2000), in response to the /ɹ/-centric explanation provided by Shapiro (1995) as detailed in the preceding subsection. He points out that when innovative [ʃ]-like variants are produced, they are always followed by an affricated /tɹ/ cluster and claims that the derivation follows a two-step process as follows: /stɹ/ → [stʃɹ] → [ʃtʃɹ] (Lawrence 2000: 83). Although this is based largely on anecdotal evidence, it is supported by several studies of /s/-retraction that note how participants who retract always have affricated /tɹ/ clusters and that /tɹ/-affrication predates /s/-retraction (Magloughlin & Wilbanks 2016; Smith et al. 2019). These results suggest a strong (though not necessarily unidirectional) link between these two processes: speakers may affricate the /tɹ/ cluster without also retracting the /s/ but they do not retract /s/ without affricating the following /t/. Affrication has also played a central role in explanations of /s/-retraction in other, lesser-studied varieties of English, such as Trinidadian English where it has been described as the catalyst for this change in /s/ centroid frequency (Ahlers & Meer 2019).

The affrication of such clusters, e.g. in words like train and try (similarly in its voiced counterpart /dɹ/, e.g. drink, dry) is well-accepted in descriptions of English, dating back to the relatively early reference by Jones (1956: §270) that “[t]he Southern English tr and dr seem to be intermediate between single affricates and sequences of two distinct sounds”. Its status is noted by textbooks on English pronunciation and highlighted for foreign learners of English (Cruttenden 2014; Lindsey 2019). Additional support for affrication of /tɹ/ clusters comes from children’s spellings, e.g. try as CHRIE and dragon as JRAGIN (O’Neil 2013: 222).

The extent to which /tɹ/-affrication is a stable phenomenon or worthy of study as a change in progress is not clear in existing descriptions and this is clouded by the lack of empirical studies, particularly when compared to /stɹ/. Wells (2011), at least, asserts its stability in Southern British English: “I don’t believe there is any such phonological change in progress.” This claim is potentially contradicted by Lindsey’s (2019: 61–2) recent description of Standard Southern British English in a dedicated chapter entitled A New chrend. He describes traditional Received Pronunciation as having had retracted [t̠ɹ] and [d̠ɹ] in words such as train and drain but modern-day Standard Southern British English as having [tʃɹ] and [dʒɹ]. Lindsey describes this as a phonological process of simplification and regularisation, indicating that a change has taken place. Jones’s mid-twentieth century observations could be interpreted as evidence in either direction: on the one hand, the existence of long-standing /tɹ/-affrication could be taken as evidence for its stability; on the other, Jones’s use of the description “intermediate” could support an argument whereby affrication was formerly something like [t̠^ʃɹ] and is now more like [tʃɹ]. At least in other varieties of English, such as North America (Smith 2013: 71), Australia (Stevens & Harrington 2016: 119) and New Zealand (Gordon et al. 2004: 612), all of whom describe the process as some kind of sound change (albeit without empirical data to demonstrate this). It is possible, however, that the British Isles are ahead on this trend.

Wells (2011) suggests a minimal pair as a diagnostic: sentry and the condensed form of century (i.e. with no intermediate schwa). If the pair are the same, it indicates phonological neutralisation. If not (which Wells claims is the common tendency), he asserts that -tr- goes with sentry and requires a broad phonetic transcription of [tɹ], not [tʃɹ].

Regrettably, as indicated, there is next to no empirical work demonstrating whether /tɹ/-affrication is a long-established stable phenomenon or, alternatively, a change in progress representing either an advancement from [tɹ] > [tʃɹ] or a gradient shift along a continuum. The major exception to this is Magloughlin’s (2018) thesis on /tɹ/-affrication in American English (see also Schwartz 2021 for perceptual tests). Magloughlin demonstrates that affrication from [tɹ] to [tʃɹ] is a phonetically gradual change in progress in Raleigh, NC well underway by the middle of the twentieth century. She presents articulatory data from ultrasound tongue imaging, supporting a situation whereby earlier generations displayed an epiphenomenal gradient effect of coarticulatory affrication, which was the diachronic precursor to a fully phonologised and subsequently stabilised realisation over time, in line with the predictions of the life cycle of phonological processes (Bermúdez-Otero 2015) (although Magloughlin does not herself reference this framework). She provides evidence from gestures such as lip rounding generalising to other positions in the word where the coarticulatory effect is not present, in accordance with the life cycle prediction of rule generalisation. Magloughlin proposes two types of speakers: those for whom /tɹ/ is identical to [tʃɹ] and those who have some kind of different affricate. Although this research illuminates the situation for North American varieties, and British English varieties may arguably be way ahead on the trajectory of the completion of change, Magloughlin’s findings demonstrate that within a given speech community, /tɹ/-affrication could be a sound change, even if Wells is right concerning present-day British English and it is stable.2

/tɹ/ clusters are certainly affricated in Manchester, although we cannot comment here on whether there is full affrication to [tʃɹ] or some kind of in -between variant. Three of the four authors are native Mancunians and report complete homophony between sentry and compressed century, perhaps suggesting a more advanced situation than that reported by Wells (2011).

An attractive aspect of the affricate-based explanation, as opposed to assimilation to /ɹ/, is that it would also capture the behaviour of /s/-retraction in /stj/ and /stʃ/ clusters, for which this study provides the first acoustic evidence in a large-scale investigation of the speech community. This is because /t/ also affricates before /j/ in a process often called “yod-coalescence”, e.g. tune [tʃʉːn], although there is not yet a detailed quantitative/acoustic study of this phenomenon.

/tj/-coalescence is discussed by Wells (1997), who claims that it was initially found in unstressed syllables, e.g. perpetual, before spreading to stressed contexts, e.g. tune, in the late twentieth century. Hannisdal (2006) provides a good overview of this widespread change in British English, with a focus on Received Pronunciation, by comparing various pronunciation dictionaries over the course of the twentieth century and the extent to which they list /tʃ/ as a possible variant for words historically containing /tj/. Hannisdal notes that coalesced forms are absent in the first edition of the English Pronouncing Dictionary (Jones 1917) but begin to appear sporadically by the later 16^th edition (Jones 2003). In other such dictionaries published since the turn of the century, affricated variants are listed more consistently, e.g. in the Longman Pronunciation Dictionary (Wells 2000) and the Oxford Dictionary of Pronunciation for Current English (Upton et al. 2001). Increases in /tj/-coalescence have also been attested within individual speech communities, including the likes of Ipswich and the Fens in East Anglia (Britain et al. 2008), where it co-exists with variable yod-dropping, and across various locales in the East Midlands (Braber & Flynn 2016).

As with the discussion of /tɹ/-affrication earlier and the fact that speakers have to affricate /tɹ/ for /s/-retraction to be licensed, it is also the case that significant retraction of /s/ in /stj/ clusters is similarly limited to instances where the /tj/ cluster itself does actually undergo coalescence.

Although the coalesced form is now incredibly widespread across varieties of British English, some speakers do still retain the /tj/ pronunciation. However, this resistance to /tj/-coalescence is largely restricted to speakers of “conservative RP”, who see such realisations as a less formal variant, particularly in word-initial onsets of stressed syllables (Upton 2008: 229; see also Ramsaran 1990 and the Daily Telegraph quote reported by Kerswill 2001: 12 describing coalescence as an “insidious degradation of spoken English”). Anecdotally, coalescence is certainly predominant in Manchester English, applying consistently across the lexicon and throughout the whole social scale.

As outlined in Section 1, investigations of /s/-retraction have almost all focused exclusively on the /stɹ/ context. There is, however, some evidence from smaller-scale studies of how the affricate derived from /tj/-coalescence influences a preceding /s/ in words such as student. Retraction in this environment is discussed briefly by Glain (2014) in a study of what he terms “instances of contemporary palatalisation”, referring to the increasing occurrence of palato-alveolar sounds such as /ʃ, tʃ, ʒ, dʒ/ in certain segmental contexts in British English. However, it should be noted that Glain draws no causal link between affrication and retraction, suggesting instead that it is caused directly by the /ɹ/ and /j/ in these clusters. A retracted /stj/ variant is also mentioned briefly in the Longman Pronunciation Dictionary (Wells 2000: 50), where it is interesting to note that the derivation is given as /stj/ → [stʃ] → [ʃtʃ] but for a word like strong it is given as /stɹ/ → [ʃtɹ] with no equivalent affrication.

Finally, retraction in /stj/ has been attested in New Zealand English. In his response to Shapiro (1995), Lawrence (2000) explicitly mentions the behaviour of /stj/ sequences but only cites examples in word-medial position (e.g. moisture) and across word boundaries (e.g. last year). Like Shapiro (1995), this paper is also largely based on anecdotal evidence rather than a robust quantitative investigation. A study of this type was conducted by Warren (2006) using elicitation recordings from the New Zealand Spoken English Database (Warren 2002), though the /stj/ results are based solely on the words student and Stewart and tokens were coded only auditory in a binary fashion (retracted vs non-retracted) rather than measured acoustically. The results suggest a strong gender divide with male speakers much more likely to retract in /stj/ relative to female speakers (42% vs 14%), with this stark divide leading to an overall significant difference in the rate of retraction between /stɹ/ and /stj/ among these speakers of New Zealand English. The behaviour of /s/-retraction in this study is also further complicated by the way in which the process interacts with variable /t/-deletion in the same clusters.

Overall, there is some evidence of retraction in /stj/ across varieties of English. However, the severe lack of large-scale acoustic analyses, particularly compared to the widely-studied retraction in /stɹ/, means that we know very little about its exact synchronic and diachronic behaviour and the extent to which these two environments of retraction behave in a similar fashion.

This study is based on a sample of 118 speakers (61 male, 57 female) who grew up in Manchester from the age of 3 or younger, with at least one local parent, stratified by age, gender, social class and ethnicity. For the purposes of the study, Manchester is defined as the urbanised area within the M60 ring-road motorway, including neighbourhoods immediately south of the M60, such as Sale, Wythenshawe, Northenden, Cheadle and Stockport.

The informants’ ages at interview range from 16 to 87. Social class is operationalised in terms of occupational levels as occupation has been shown to be the best single indicator of socio-economic status for the purposes explaining linguistic variation, both in the US (Labov 2001) and in the UK (Baranowski & Turton 2018). Following Baranowski (2017), there are five occupational levels ranging from lower-working for unskilled workers to upper-middle class for occupations such as university professors and high-level managers and administrators. The assignment to a particular social class is based the occupational history of a speaker rather than just the last job they held; children are assigned the social class of the parents. The coding of ethnicity is based on speakers’ self-identification, with 118 white British, 18 Pakistani and 13 Black Caribbean informants in the wider corpus, though in this study we focus on the white British speakers due to the relatively smaller samples in other groups. A more detailed breakdown by age and social class is given in Table 1.

The informants were recorded during sociolinguistic interviews, conducted mostly between 2011 and 2018 with only a small number of speakers recorded earlier than this. Some informants were recorded in their homes, some at their place of work and others on a university campus. All were recruited through a “friend of a friend” approach (after Tagliamonte 2006: 21–2). The interviews focused on eliciting narratives of personal experience, which are known to approximate speakers’ vernaculars (Labov 1984; Tagliamonte 2006). The interviews were supplemented with two formal elicitations: word-list reading and minimal pairs for a number of vocalic and consonantal contrasts. However, as these elicitation tasks do not contain sufficient tokens of /s/ in the target environments, the results in this paper are based solely on spontaneous speech. The recordings were forced-aligned using FAVE (Rosenfelder et al. 2014), the online Forced-Alignment and Vowel Extraction suite developed at the University of Pennsylvania, in order to produce a time-aligned transcription for efficient and automated extraction of the relevant acoustic measures. This process of acoustic data extraction and processing is described in the following section.

Studies of /s/-retraction almost always characterise the fricative quality using its centroid frequency, commonly referred to as its centre of gravity (CoG), which has been shown to correlate with the size of the anterior cavity in sibilant production: more [s]-like realisations have a higher CoG and more [ʃ]-like realisations have a lower CoG (Jongman et al. 2000). Although other spectral moments such as standard deviation, skew and kurtosis have also been shown to be relevant in quantifying place of articulation along the front–back plane (Koenig et al. 2013), the analysis in this paper utilises CoG due to its widespread use in other sociolinguistic studies of this variable (see e.g. Baker et al. 2011; Wilbanks 2017; Stuart-Smith et al. 2019).

All interviews in the corpus were recorded at a 44.1 kHz sampling rate with a Sony PCM-M10 recorder and Audio-Technica ATR3350 lavaliere microphone, though some were later downsampled to 22 kHz for the forced-alignment process; because of this, all interviews were downsampled to this same 22 kHz rate at the acoustic analysis stage in order to ensure comparability between speakers with respect to centroid frequencies. In addition to this, all recordings were low pass filtered at 750 Hz to remove spectral information corresponding to residual voicing and environmental noise, as is common in sociophonetic studies of this variable (see e.g. Gylfadottir 2015; Wilbanks 2017).

Acoustic measurements were extracted using a custom Praat script (Boersma & Weenink 2018) and the speakr package (Coretta 2021) in R (R Core Team 2018) to automate extraction of spectral values for the whole corpus. As well as capturing all tokens in the two environments of interest, namely /stɹ/ and /stj/, the script also extracted acoustic measurements for underlying /s/ and /ʃ/ segments in all other environments, as well as the fricated part of /tʃ/ segments, in order to establish speaker- and community-wide norms for the wider sibilant space. CoG was measured at the midpoint of these fricatives and at five equally-spaced points across their duration based on the time-aligned intervals as determined by the forced alignment process. The script also extracted a smoothed transformation of the entire spectral profile. The analysis in this paper will focus on the static midpoint measurements for ease of comparison with earlier studies, but in future work we plan to look at the wider spectral profile of these tokens for greater insight into the dynamics of /s/-retraction.

In addition to these primary acoustic measures, other linguistic information was extracted including segment duration, position in the word (initial vs medial vs final), surrounding phonological environment (the segments immediately adjacent to the fricative and, in the case of /stɹ/ words, the vowel following this consonant cluster) and lexical frequency (measured on the Zipf scale and based on frequency counts in the SUBTLEX-UK corpus, see van Heuven et al. 2014).

The script extracted acoustic measurements of just over 100,000 sibilant tokens but, due to the automated nature of forced-alignment and the resulting potential for erroneous measurements, the data extraction was followed by a data-cleaning process in R based on the practices followed in earlier work (see e.g. Stuart-Smith et al. 2019 on the automated analysis in SPADE). This involved the removal of statistical outliers, identified as tokens with a CoG value outside 1.5x the interquartile range for a particular speaker’s production of either /s/ or /ʃ/ (N = 4,327), tokens with a duration less than 50 ms (N = 10,385) and tokens with a CoG or peak frequency less than 2,400 Hz (N = 13,445). The final dataset contains 72,986 tokens, 1,019 of which are /stɹ/, 236 of which are /stj/ and 82 of which are /stʃ/.

Following work by Gylfadottir (2015) and Ahlers & Meer (2019), the CoG measure was normalised and scaled into z-scores based on each speaker’s individual distribution. This removes individual differences in absolute frequency range (e.g. due to physiological differences in vocal tract size) while maintaining the relative relationships between individual tokens, thus allowing for more reliable comparisons between speakers when investigating the community-wide behaviour of /s/-retraction. This is particularly important in the case of /s/-retraction, where interest lies primarily in the relationship between categories, the retracted /s/ and a speaker’s own sibilant range for /s/–/ʃ/, rather than the absolute frequency values of these tokens. The result is a z-score where higher/positive values correspond to a higher centroid frequency, i.e. more [s]-like, lower/negative values correspond to lower centroid frequencies, i.e. more [ʃ]-like, and values close to 0 represent the midpoint of a speaker’s sibilant frequency range. The results presented in the following section are based primarily on these normalised measures.

This paper is the first to report a comparable apparent-time change also occurring in the /stj/ (e.g. student) and /stʃ/ (e.g. mischief) environments. Crucially, these three contexts appear to behave as one with respect to their current rate of retraction and there are no significant differences between their rates of change in this time period either. In Section 2, we outlined the two major proposals that have been put forward concerning the mechanisms behind /s/-retraction, specifically whether it is caused directly through a long-distance effect of /ɹ/ (see e.g. Shapiro 1995; Baker et al. 2011) or alternatively through contact with an adjacent affricated /t/ (see e.g. Lawrence 2000). Returning now to this question, the new evidence presented here lends strong support to the latter proposal. While it is not impossible that retraction is triggered by three distinct mechanisms in /stɹ/, /stj/ and /stʃ/, the fact that their behaviour is so similar and that they change in tandem suggests that we should instead appeal to a single unifying explanation that invokes affrication as the cause of retraction.

Of course, under a frequentist statistics framework centred around null hypothesis significance testing, it is perfectly possible that the lack of significant difference between these three environments is in part due to issues with sample size and statistical power. Indeed, while the three environments are broadly similar among the youngest speakers, visual inspection of the frequency distributions from Figure 2 indicated that /stʃ/ was perhaps slightly ahead of /stɹ/ and /stj/ at an earlier point of the change. While it would be difficult to account for a difference in the opposite direction, there is in fact an obvious reason why at one point /stʃ/ may have been ahead of the other two environments in this change: the affrication of /t/ before /ɹ/ and /j/ was (and to some extent still is) variable and, as a result, the change in these environments could be underestimated due to the inclusion of older speakers who may not actually take part in this /t/-affrication process at all. The change will naturally appear to be further ahead in /stʃ/ than in /stɹ/ and /stj/ while affrication remains optional in the latter two compared to its obligatory presence in the former, at least for older speakers who acquired the dialect at an earlier stage while affrication in /tɹ/ and /tj/ was still an ongoing change. The inclusion of “underlying” /stʃ/ words such as mischief in this analysis brings with it an obvious benefit, namely it allows us to isolate the effects of /s/-retraction from the interacting changes affecting the adjacent clusters, since these words have—and have always had—an affricate present underlyingly. A follow-up study should be conducted with a specific focus on apparent-time change in /tɹ/-affrication and /tj/-coalescence. This would shed light on whether these two changes have been taking place alongside /s/-retraction (and as such the three sound changes have been working in tandem) or, alternatively, that they have in fact been stable for a long time in this speech community, laying the foundations for /s/-retraction from an early stage.

It is important to note that the results may not necessarily be generalisable to all instances in which /s/-retraction has developed and propagated through a speech community. While this change appears to have been initiated in a number of varieties at a roughly similar time, it is possible that they are all independently triggered by a range of different mechanisms. This is especially pertinent given the fact that most varieties of American English are yod-dropping: given the absence of /j/, they demonstrate no comparable affrication or retraction in /stj/ words like student and stupid and, as such, the distribution of retracted /s/ tokens is quite different. Children acquiring the language are exposed to fewer instances of retracted /s/ before an affricate and therefore might not receive the kind of input required to reanalyse the segmental conditions of /s/-retraction in this way.5

However, even setting aside our data for a moment, there are still questions that can be raised regarding the proposal put forward by Baker et al. (2011). Most notably, if retraction is caused directly by assimilation in tongue shape to /ɹ/, why is retraction consistently so further advanced in /stɹ/ relative to /spɹ/ and /skɹ/ in all varieties of English affected by this change? This is particularly surprising for /spɹ/, where the intervening consonant has a labial place of articulation and as such the tongue has no distinct target to hit between the /s/ and /ɹ/. In a study of /s/-retraction in Philadelphia English, Gylfadottir (2015) raises a similar concern and in doing so also argues against this being a case of “assimilation at a distance” to the non-local /ɹ/ in these clusters. This question could be further illuminated with articulatory data on this covert variability in rhotic production in different segmental environments and across a range of speech communities but, in the absence of this, it remains the case that an explanation in terms of adjacency to /tʃ/ more closely matches the data.

Alongside the disagreement over why /s/-retraction takes place in these environments, there are also conflicting reports over the origins of /s/-retraction, specifically with respect to the initiation of change in /stɹ/ in the other environments that demonstrate some low-level retraction.

Janda & Joseph (2003) argue that retraction started in /stɹ/, where today it is registered most strongly and later “spread” to the other contexts such as /sp, st, sk/. While this kind of rule generalisation is entirely possible, in which the target environment of the change is reanalysed to encompass a wider range of segmental contexts, this claim is only based on the fact that retraction is “sporadic” and less advanced in these contexts. In other words, it is based on neither real-time nor apparent-time data from which change can be observed or inferred. As such, there are obvious limitations to relying on these differences in effect magnitude from synchronic data as a tool for estimating the initiation of change.

Instead, it is entirely plausible that retraction began in all complex onset clusters simultaneously, before advancing at a faster rate of change in /stɹ/—and /stj/, as demonstrated in this paper for Manchester English—leading to the clear separation of contexts we observe in many varieties of English spoken today. Indeed, this has been suggested by Stevens & Harrington (2016) for Australian English, in which low-level retraction is observed in /sp, st, sk/ and even /stɹ/. This has been described as the “phonetic pre-conditions” for the more advanced retraction in /stɹ/ that has developed in most varieties of English, suggesting a trajectory of change that sees retraction staring off in all complex onset clusters first before advancing more rapidly in /stɹ/.

In our data, we observe that /st/ is significantly more retracted than pre-vocalic /s/ but that it actually shows very little change over time. Moreover, the minor change we do find is actually in the opposite direction, with younger speakers producing more /s/-like tokens with a higher CoG in /st/ relative to older speakers. This would not be the direction of change we expect if retraction in /st/ represented a later stage of the wider retraction process after spreading to more contexts.

This increase in the CoG of /st/ in apparent time mirrors the same change observed in our data for pre-vocalic /s/, which is also produced with a higher CoG and therefore a “hissier” quality, among younger speakers. Conversely, these same speakers produce a “hushier” /ʃ/ with a small but significant decrease in its CoG in apparent time. Taken together, these results illustrate an apparent expansion of the sibilant space over time leading to a greater acoustic contrast between /s/ and /ʃ/ for younger speakers. Interestingly, a similar result appears in work conducted by Wilbanks (2017) on the variety of American English spoken in Raleigh, NC although there it is restricted to male speakers.

There are two possible interpretations of this finding: (i) that it represents a change in progress involving more distinctive productions of /s/ and /ʃ/ targets and therefore an expansion of the fricative space over time or (ii) that it actually represents the physiological effects of ageing and a reduction in the acoustic /s/–/ʃ/ contrast within speakers’ own productions as they age. There are also two possible pathways by which the latter might occur, both of which find support in previous literature (Matthies et al. 1994; Perkell et al. 2004; Koch & Janse 2015). An articulation-centric explanation might foreground the effects of reduced motor control in old age and by consequence the greater difficulty in producing the precise articulations of /s/ and /ʃ/, which already involve a combination of gestures such as tongue grooving and lip rounding in addition to the midsagittal tongue shape (Rutter 2011). There is an alternative explanation routed through perception and the feedback loop: speakers with hearing loss, which disproportionately affects older adults (Bowl & Dawson 2019), would suffer reduced auditory feedback and may as a result impact their own production of these segments. This would likely be registered most strongly in sibilant production, and particularly /s/ itself, since hearing loss primarily affects the higher frequency range where more of the energy in /s/ is concentrated.

Further work should be conducted in other speech communities to establish how geographically widespread this phenomenon is beyond the disparate locales of Manchester and North Carolina (Wilbanks 2017) and also to shed more light on the exact status of this change. The difficulty in identifying its direction—a community-level increase in sibilant contrast over time or an individual-level decrease as speakers age—is a natural limitation of relying on apparent-time data to infer language change. As such, future work would ideally draw upon a longitudinal or real-time approach in order to tease apart the various temporal dimensions of birth year, age at interview and time of interview (see Fruehwald 2017).

Regardless of the direction of change, the fact that pre-vocalic /s/ and /ʃ/ are not themselves stable highlights the importance of interpreting the degree of change in /s/-retraction with respect to the wider sibilant space itself rather than in isolation.