2.1. Delimiting speaker groups based on their production of covariation patterns

Historically, approaches to identifying distinct groups of speakers focused on single linguistic features—mapping individual isoglosses and tracking quantitative variation in individual variables across macro-social categories (e.g., Labov, 1966; Trudgill, 1972). Since the mid-twentieth century, however, dialectology and (dia)lectometry have increasingly examined how language varieties are distinguished by their aggregate differences across a range of linguistic features (see Wieling & Nerbonne, 2015). Contemporary variationist sociolinguistics has also seen a growing interest in the relationships between sociolinguistic variables (e.g., Beaman & Guy, 2022).

Sociolinguistic work situated within the third wave in particular (e.g., Eckert, 2012; Hall-Lew et al., 2021) has emphasised how the co-occurrence of specific subsets of variables contribute to the expression of personal identity and construction of speaker styles, stances, and personae in speech production (e.g., Bucholtz, 2008; Pratt, 2019). Others have applied multivariate statistical methods to sociolinguistic data to delimit groups of speakers characterised by systematic patterns of covarying variables. For example, Horvath’s (1985) application of Principal Components Analysis (PCA) in her seminal work on ethnic variation revealed distinct Australian English sociolects in Sydney. These sociolects corresponded to age, ethnic and socioeconomic speaker groups characterised by their patterns of covarying face, fleece, price, mouth and goat variants. More recently, multivariate techniques have been applied to differentiate between speakers with more and less coherent use of a range of Swabian dialect features over time (Beaman & Sering, 2022) and identify the Bequian English linguistic signatures of villages in St. Vincent and the Grenadines (Meyerhoff & Klaere, 2017).

Multivariate analyses of New Zealand English data have also demonstrated that speakers of New Zealand English can be characterised in terms of covarying patterns in the normalised vowel space (Brand et al., 2021; Hurring et al., 2025). While the NZE vowel system is comparable to other standard, non-rhotic, English varieties such as RP and Australian English (Bauer et al., 2007, p. 98), NZE monophthongs have undergone extensive changes over the course of the twentieth and twenty-first centuries (see Figure 1). In particular, the short front vowels kit, dress, and trap have undergone a well-documented push-chain shift where the raising and fronting of trap prompted the raising and fronting of dress and centralisation of kit (e.g., Gordon et al., 2004; Hay et al., 2015). This has also been accompanied by a retraction and lowering of fleece, raising and fronting of nurse and thought, and fronting of goose (Brand et al., 2021; Maclagan et al., 2009, 2017). The back vowels lot, start and strut have each moved lower and/or backer in the vowel space to different degrees (Brand et al., 2021, p. 10).

Figure 1

Change over time for NZE monophthongs in Origins of New Zealand English (ONZE) corpus data (Gordon et al., 2007). Hurring et al. (2025, p. 7) used under CC BY 4.0.

Brand et al.’s (2021) application of PCA to data from the Origins of New Zealand English (ONZE) corpus (Gordon et al., 2007), representing speakers born 1864–1982, revealed that a speaker’s realization of a given monophthong within the set of ten NZE monophthongs in Figure 1 provides insight into their realisations of the other vowels in this set. The analysis specifically revealed three subsets of covarying vowels (vowel clusters), the first two of which are the focus of this paper. Brand et al. label the first subset as ‘the restructuring back vowels’ (speakers with lower, fronter thought have backer start and strut, henceforth the back-vowel configuration). The second subset captures a group of vowel changes in progress (trap, dress, fleece, kit, nurse, and lot), across which individual speakers are consistently leaders or laggers (henceforth the leader-lagger continuum).

Hurring et al. (2025) then implemented the same methodology as Brand et al. (2021) using the QuakeBox (QB), a more recent corpus of accounts of the 2010–2011 Canterbury earthquakes (Clark et al., 2016; Walsh et al., 2013). Speakers aged 18 to 85+ years old are represented in the corpus. Hurring et al. not only revealed similar clusters of vowels in QB as those in ONZE but documented relative stability of these clusters within recordings made of the same individuals eight years apart.² Hurring et al. therefore provide evidence for stable relationships between back vowels and between ongoing sound changes across NZE speakers and over the lifespan of individuals. Together, Hurring et al. and Brand et al. show that subsets of NZE monophthongs systematically covary in speech production.

Brand et al. (2021, p. 21) argue that the observed patterns of covariation cannot be solely explained as “the phonetic implementation of wholesale phonological features.” The leader-lagger continuum encompasses vowels involved in chain-shifts in the NZE vowel space (trap, dress, and kit) as well as vowels that are not, but are still undergoing change (e.g., nurse). Moreover, all NZE monophthongs have undergone change over time, but they are not all captured in a single cluster. The repeated observations of stable covarying vocalic patterns that are not clearly structurally linked brings us, then, to the question of whether the underlying explanation for these vowel clusters could be social (Brand et al., 2021, pp. 21–23).

The results of Hurring et al. (2025) further indicate that the observed clusters could be social by showing the same vowels continue to covary even after the apparent completion of some sound changes. For example, trap was one of the leader-lagger vowels in Brand et al.’s (2021) analysis of the older ONZE corpus, in which trap was also undergoing change as part of the front-vowel chain shift. In the contemporary QB corpus examined by Hurring et al. (2025), however, realisations of trap appear to be stable and are not significantly predicted by the age of the speaker. Nonetheless, its production continues to covary with the other leader-lagger vowels. While the initial covariation may have been partly structural, its persistence in contemporary NZE is consistent with acquiring a relatively stable social meaning.

One starting point for examining the potentially social basis of covarying NZE vowel clusters is their perceptual relevance and sociolinguistic salience: Do listeners associate different covariation patterns with different social characteristics?

2.2. The perception of speakers with different covariation patterns

How do patterns of covariation in speech translate in listener perception? The sensitivity of listeners to sociolinguistic variation in speech is well documented (see Campbell-Kibler, 2010; Drager, 2010; Thomas, 2002), but our knowledge of how listeners differentiate between speakers with distinct covariation patterns is comparatively limited. Despite the evidence for linguistic features working “synergistically in the perceptual processes” (Montgomery & Moore, 2018, p. 656), sociolinguistic perceptual research, like its research on production, has focussed on individual variables, often in highly controlled audio stimuli. Even perceptual work on personae has tended to focus on how social information affects listener categorisation of individual variables (e.g., D’Onofrio, 2015, 2018) or the social meanings (and, by extension, personae) associated with individual variables or multiple variables analysed independently (e.g., Becker, 2014; MacFarlane & Stuart-Smith, 2012; Pharao et al., 2014).

The available sociolinguistic perceptual research on covariation has mainly focused on how variants of multiple variables interact to affect speaker evaluations of male speakers’ masculinity and sexuality (e.g., Campbell-Kibler, 2011; Levon, 2007, 2014; Pharao & Maegaard, 2017). These studies demonstrate variant combinations can affect how speakers are perceived and evaluated by listeners, although co-occurring variants do not guarantee an additive perceptual effect. For example, Levon (2014, pp. 554–555) found combining elevated mean pitch and increased sibilance did not lead to an increase in perceived “gayness” beyond what either variable achieved independently. Separately, the literature on listener attitudes towards accents has shown standard or conservative English accents tend to carry more prestige than non-standard ones (Giles, 1970; Hiraga, 2005), although this effect can be mediated by listener age (e.g., Levon et al., 2021). Convergence research has also revealed listeners can associate ‘heard’ variants with the ‘unheard’ variants with which they (are expected to) typically co-occur. For example, Wade (2022) shows speakers converge on more ‘Southern’ realisations of price after exposure to Southern features other than monophthongal price in audio stimuli.

While there is evidence for New Zealand English variables carrying social meaning for NZE listeners (e.g., Bayard & Bartlett, 1996; Szakay, 2008; Szakay, 2012; Walker, 2007), there is limited evidence tying social judgements of NZE speakers to covarying sociolinguistic variables (though see Gordon, 1997). Sheard et al. (2025) took a first step to investigating whether the covarying NZE vowel patterns identified in Hurring et al. (2025) are perceptible to listeners, taking short clips from a subset of the monologues analysed in Hurring et al. and playing them to listeners in a pairwise similarity rating task. Listeners were asked to rate how similar the speakers sounded and given instructions that steered them toward rating the likely social characteristics of the speakers, as opposed to the content or superficial acoustic features. Sheard et al. used Multidimensional Scaling (MDS) to create a perceptual similarity space, and investigated how this space was related to a subset of linguistic variables: the Principal Components from the analysis in Hurring et al., representing the speaker’s position in the leader-lagger continuum and the back-vowel configuration, and speaker articulation rate and mean pitch.

Their analysis suggested that listeners systematically differentiated between speakers based on their pitch, speed, and position in the leader-lagger continuum, but not the back-vowel configuration. Figure 2 shows the interpretation of the perceptual space based on listeners’ pairwise ratings in Sheard et al. (2025), with labels for characteristics of distinct groups of speakers based on their speech production. The results indicate higher pitch speakers (top left corner, diamonds, circles) are perceptually distinct from speakers with lower mean pitch and/or slower speech rates (bottom, squares, and down-facing triangles). Within the lower-pitch (and/or slower) speakers, leaders (down-facing triangles) are perceptually distinct from laggers (squares). Fast leaders appear to be less perceptually distinct and overlap with each of the surrounding groups.

The results of perceptual research to date point, therefore, to listener awareness of certain covariation patterns and the capacity of co-occurring variants to affect listener evaluations. We also have evidence that the leader-lagger continuum is perceptible to listeners and differentiates between speakers of NZE from both the side of production and perception. We do not yet know, however, if NZE speakers with different patterns of vocalic covariation are associated with specific or different social meanings or characteristics. Is there a salient social difference between the speakers in different clusters shown in Figure 2?

Further, Sheard et al. (2025) show that some characteristics may be more salient than others–for example, speakers sharing high pitch are rated as similar more consistently than speakers sharing low pitch. The task also shows order effects, suggesting that the characteristics of the first stimulus heard may influence the dimension on which similarity is judged, consistent with work showing that similarity is not a symmetric concept (Tversky, 1977). Would the same clusters of speakers emerge if we did not require listeners to assess pairwise similarity but instead freely classify speakers into groups and listen to the speakers as many times as they liked? Or would such a task enable truly social effects to emerge more robustly?

Variationist and accent bias perceptual research tend to take a top-down methodological approach in which listeners evaluate speakers along a set of predetermined characteristics (e.g., professional, friendliness, employability) based on samples that researchers have pre-selected as representative of different speakers or accents. In our case, however, we have not yet established the social characteristics listeners might associate with either the leader-lagger continuum or back-vowel configuration. It is possible covarying vowel patterns carry associations with broader macro-social categories (e.g., rural or urban New Zealanders) or more specific, localised, personae (i.e., a specific kind of urban or rural New Zealander). Alternatively, it is possible they do not carry any social meaning at all or are not salient in the context of uncontrolled audio stimuli.

Figure 2

Interpretation of perceptual space produced by means of MDS in Sheard et al. (2025), with speakers grouped by shared production characteristics (used under CC BY 4.0).

As such, here we build on the foundations of the production and perception research on covarying NZE vowels by taking a bottom-up approach to listener perception of covariation that allows us to directly connect the perceptual grouping of speakers to the labels NZE listeners used to describe them. We conduct a free classification task, in which we asked participants to create groups of similar-sounding speakers. Crucially, we require participants to label these groups, providing insight into the types of judgements participants are making and the degree to which these are social or not.

2.3. Research questions

We address the following two research questions through the implementation of our free classification task:

1. A social interpretation of the relationship between covarying vowels: Brand et al. (2021) argue that the observed vowel clusters, particularly the leader-lagger continuum, likely reflect socio-stylistic rather than structural or phonological relationships between vowels. This explanation relies on these clusters carrying associated social meaning(s) which have not yet been tested for. As such, we ask whether speakers with different covarying vowel patterns are not only perceived to sound different to one another but are associated with distinct social characteristics. Are the previously documented perceptual differences between leaders and laggers socially meaningful? In other words, do listener judgements support a social interpretation of the relationships between covarying vowels, a more structural interpretation of these clusters as phonologically driven parallel sound changes, or a combination of the two?

2. The social meaning of speaker pitch and articulation rate: The participants in Sheard et al. (2025) consistently used these characteristics to rate speaker pairwise similarity, and speakers were differentiated within the MDS perceptual space based on both characteristics (in interaction with each other and–in the case of pitch–the leader-lagger vowels). Is this because they hear these characteristics as indexed to particular social meanings? Or is it simply because these characteristics are so salient that they dominate, even when the instructions steer participants toward social ratings?

3.1. Free classification task design

Free classification is a technique with comparatively limited implementation in the context of sociolinguistic perceptual research. Where it has been applied, the focus has usually been on listeners’ identification of regional accents of American English speakers (e.g., Clopper & Bradlow, 2009; Clopper & Pisoni, 2007). In free classification, participants are presented with multiple stimuli at the same time and asked to place the stimuli into groups, with the number and size of groups determined by the participant. It is a “perceptual sorting task in which listeners are asked to group stimuli according to similarity” (Lansford et al., 2014, p. 2). The “critical property” of the paradigm is that category labels and dimensions of contrast do not need to be specified in advance by the experimenter (Clopper, 2008, p. 575), making it inherently suited for our research question. It allows for clusters of speakers to be developed from the bottom up.

We use the same audio stimuli from the same 38 QuakeBox participants in Sheard et al. (2025) so that results of the free classification task are comparable to those of the pairwise rating task. These speakers are all Pākehā (New Zealand European) women, aged 46–55, who were analysed in Hurring et al. (2025) and consented to have audio of their QuakeBox monologue shared publicly. As detailed in Sheard et al., the clips are each up to 10 seconds in length, normalised to 70 dB for consistency, and contain at least half of the monophthongs analysed in Brand et al. (2021) and Hurring et al. Further details about the stimuli are available in Section 3 of the supplementary materials.

To enable the experiment to be conducted online, we used an adapted version of the Audio Tokens JavaScript toolbox developed by Donhauser and Klein (2023) integrated with jsPsych (de Leeuw, 2015), an open-source JavaScript framework for creating online experiments. This task specifically utilised the clustering setting from the Audio Tokens toolbox, which is effectively a free classification task where participants are presented with stimuli represented by coloured circles which are surrounded by a larger circle. Participants then drag and drop the stimuli into groups within the larger circle, with a black line appearing to connect stimuli placed in the same cluster. We also adapted the clustering setting of the toolbox so that a text box appeared on the page for each group a participant made (i.e., if a participant made four groups, there would be four text boxes, see Figure 3).³ Participants were required to write specific labels related to the groups they had created to complete the task. At the end of the task, participants completed a demographic questionnaire regarding their age, gender, ethnic background, level of education, occupation, and region they grew up in.

To familiarize participants with the interface and the drag and drop mechanism, they were given a practice task that used audio stimuli from musical instruments rather than voices to avoid priming them for the actual stimuli (e.g., using stimuli from different geographical varieties of English may have primed them to listen for regional differences in the actual task). To avoid overloading participants, the total number of stimuli were distributed randomly across three tasks (12 stimuli in the first of the tasks, 13 in the second and third).

Figure 3

The interface for the free classification task after the participant has made the speaker groups.

Participants were given the following instructions:

“Sometimes people speak similarly because they are friends, have similar occupations or personalities, or grew up in the same place. In this task, we want to know which speakers of New Zealand English you think sound similar to each other, and which speakers you think sound different to one another.

You will be given audio clips from different New Zealand women talking about their experiences of the Christchurch earthquakes.

We would like you to put the people who you think sound similar into groups by moving the coloured circles closer together. We are interested in who you think sound similar based on the way they talk, rather than the things they say.

Before this task, there will be a practice task using audio from musical instruments.

We estimate that both tasks will take about 20 minutes, but you may take as much or as little time as you need.”

3.2. Participants

Participants were recruited online using the same eligibility criteria as Sheard et al. (2025), with a total of 127 eligible participants completing the task.⁴ All participants are (a) over the age of 18, (b) a speaker of New Zealand English, and (c) born in New Zealand or living in New Zealand since the age of seven. Participants were paid a $10 e-voucher for completing the task. Based on the completed demographic questionnaires, 75% of participants are women, 18% are men, and the remaining 6% comprises non-binary genders (non-binary, agender, gender diverse, genderqueer). In terms of ethnicity, 75% of participants are Pākehā. The next largest ethnic group is made up of those who identified as both Māori and NZ European (10%), with 2% of participants identifying only as Māori. The remaining participants were of Asian background (NZ Chinese, Japanese, Indian, Nepali, and Filipino), or unspecified mixed ethnicity. Just over half the participants (53%) are aged 18–29, and a fifth aged 30–39 (22%) and 40–49 (18%). The remaining 8% are over 50 years old. Two-thirds of participants grew up on the South Island of New Zealand, predominantly in the Canterbury region (48%), with the remaining third growing up in the North Island. While participants grew up in both the South and North islands, all are currently living in the South Island. Almost half are currently living in Christchurch (48%), with 8% living in the Canterbury region but not in Christchurch and the remainder living in other areas of the South Island.

4.1. Identifying listener categorisation strategies

We classified the free text label responses from each participant on three levels, differing in the degree of detail. The initial label classifications are listed in the Specific classification column of Table 1. Where a participant used multiple categories within the same label, we classified the label according to the first social category in the label or, when participants did not use any social categorisation, the first listed classification. We then grouped specific labels into broader categories that captured the different levels of the categories (e.g., ‘young,’ ‘middle age’ and ‘old’ are all labels related to ‘Age’). Finally, we make a broad three-way, differentiation among labels related to ‘Social Factors’, labels related to ‘Features of Speech’, and labels not related to either social or speech characteristics.

It is clear from Table 1 that, despite our attempt to encourage participants to make groups of (perceptually) socially similar speakers, not all participants did so. Indeed, a full 50% of the labels related only to features of speech, with 40% of labels including social factors (see supplementary materials, Section 4.1). Listeners also tended to draw on multiple categorisation strategies across the three iterations of the task; the number of groups participants made in a single trial ranged from 1 to 8, with a mean of 3.6, while the number of broad label categories participants used in a single trial ranged from 1 to 3, with a mean of 1.7. Finally, as Figure 4 displays, while participants referred to a wide range of speech features and social factors in their labels, the most frequently used specific categories relate to perceived age, the typicality of their accent, their articulation and their speed.

Figure 4

The number of participants who used each specific label category.

From the distribution of specific and broad label categories, there is clear variability between the participants in how they approach the categorisation of ‘similar’ NZE speakers. There are also differences within individual participants. It is common for listeners to use not only one broad approach or another but a combination within the same iteration of the task. The comparatively high number of participants employing labels related to speed and pitch is consistent with other perceptual research documenting their salience to listeners and supports their role in perceptually differentiating between the same NZE speakers in the pairwise rating task (Sheard et al., 2025). Perceived clarity of articulation was also a highly common labelling strategy, as was perceived creakiness and whether a speaker had a ‘lisp.’⁵

4.2. Applying multidimensional scaling to determine dimensions of perceived social similarity

We then conducted our preregistered analysis and applied multidimensional scaling (MDS) in R (R Core Team, 2024) to the filtered results of the free classification task to create a perceptual similarity space. We filtered out participants who used the clip topic as a labelling strategy (n = 10) and who did not place audio stimuli into groups (n=1), due to these being clear markers of the task instructions not being followed. The results from the remaining participants were included (n = 116).

Multidimensional Scaling is a data-reduction technique that is highly suited to perceptual similarity data. It has a strong precedent of application in analyses of perceived speaker similarity in speech science and perceptual dialectology (e.g., Casserly, 2010; Clopper & Bradlow, 2009; Clopper et al., 2006; Gelfer, 1993; Kreiman et al., 1992; Matsumoto et al., 1973; McDougall, 2013; Murry & Singh, 1980). MDS converts measures of pairwise (dis)similarity between objects into distances within a low-dimensional space, where the relative positioning of each object reflects perceptual similarity (Borg & Groenen, 2005, p. 3). The dimensions of this perceptual space are interpreted as corresponding to the underlying cues driving listeners’ judgments of similarity. Importantly, unlike PCA, where the order of principal components reflects the comparative amount of variance in the data they explain (i.e., PC1 explains more than PC2), the order of the dimensions in MDS is arbitrary (i.e., Dimension 1 is not more important than Dimension 2).

As described in Sheard et al. (2025) and Section 4.2 of the supplementary materials, we used smacofSym from the smacof package (Mair et al., 2022) to apply a two-dimensional spline MDS analysis to a pairwise 38 x 38 dissimilarity matrix derived from the filtered results of the Free Classification task. As participants were not exposed to all speaker pairs, each matrix value represents the proportion of times a given pair of speakers was (not) placed into the same group. We chose to run a two-dimensional MDS analysis based on two permutation tests (Mair et al., 2022; Wilson Black & Brand, 2021) that did not indicate further dimensions were required.

One of the outputs of smacofSym is coordinates, or scores, for each stimulus in the input data. These scores can be mapped along an x (‘Dimension 1’/’D1’) and y (‘Dimension 2’/’D2’) axis, to create a perceptual similarity space. Stimuli situated closer to each other within this space were grouped together more often by listeners and are more perceptually similar. Stimuli situated further away from each other were grouped together less often by listeners and are less perceptually similar. It is the relative distance between stimuli that is important for the interpretation of an MDS space. The position of stimuli within the space is otherwise arbitrary; while it is possible that stimuli D1 or D2 scores would directly align with a given feature in production or with a different MDS space generated from other data, this is not guaranteed.

As we are here building on the results of Sheard et al. (2025), we can ensure that the MDS spaces are comparable by rotating our MDS scores to maximally align the two spaces (Figure 5). We have, therefore, applied Procrustes rotation (Oksanen et al., 2022) to the space (i.e., speakers’ D1 and D2 scores) to maximally align with speakers’ D1 and D2 scores from the MDS space in Sheard et al. (2025). The MDS spaces from the two tasks are now highly similar. The rotated D1 scores from the Free Classification MDS space and the D1 scores from the Pairwise Rating task are strongly correlated (r = 0.81, p < 0.05). The rotated Free Classification D2 scores also strongly correlate with the Pairwise Rating D2 scores (r = 0.73, p < 0.05). As such, the rotated scores in Figure 5B will be the object of the following analysis, and we explore the similarities between the two spaces further in Section 5.2.

Figure 5

(A) The MDS output from Sheard et al. (2025) based on pairwise ratings and (B) the Free Classification MDS output rotated to maximally align with the space in (A).

4.3. Principal Component Analysis to determine relationships between MDS dimensions and perceived speaker groups

We now have our dimensions of perceptual similarity from the MDS and have rotated the space so that D1 and D2 align with the MDS space in Sheard et al. (2025). The MDS does not tell us how listeners differentiate between speakers along these dimensions. When data on listener strategies are not available, the linguistic variables that differentiate between speakers from the side of production (i.e., speakers in one area of the space are faster than speakers in the opposite area of the space) are assumed to be perceptually meaningful to listeners (as in Sheard et al., 2025). The results from the Free Classification task, however, provide a means of identifying both (a) which speakers are most perceptually distinctive, and (b) which labels listeners used to describe them.

Here, we use the categories from Section 4.1. to identify which labels increase or decrease in use as speakers’ rotated D1/D2 scores increase or decrease by implementing a separate ‘Label’ PCA as described in Section 4.3 of the supplementary materials. We note that this PCA was not preregistered, and we did not preregister any analyses related to the label data. The input data for the Label PCA were speakers’ rotated D1/D2 scores, along with the proportion of times that each specific label category was used to describe each stimulus (see first column in Table 1, with categories used by <10% of total experiment participants excluded).⁶

Unlike MDS, PCA does not require researchers to specify the number of dimensions in advance, but we do need to decide how many Principal Components should be considered in the interpretation of its results. Following the approach to PCA outlined in Wilson Black et al. (2023), we used a permutation and bootstrapping procedure to compare the variance explained by each PC in bootstrapped versus randomised data. This comparison reveals that the variance explained by the Principal Components from PC3 onwards for the bootstrapped data does not explain more variance than would be explained by chance. As with the MDS, we decide to stick with two Principal Components (PCs), referred to here as Label PC1 and Label PC2. The output of PCA has two relevant components. First, each stimulus has a Label PC1 score and Label PC2 score. Second, a subset of the input variables (i.e., speakers’ MDS D1 and D2 scores and the label proportions) are loaded onto each Label PC. Each variable is loaded either positively or negatively. The value of a positively loaded variable increases as a speaker’s PC score increases, while the value of negatively loaded variable decreases as a speaker’s PC score increases.

Figure 6A and Figure 6B depict the variables loaded onto Label PC1 and Label PC2, respectively. A ‘+’ indicates that a variable is positively loaded on the PC while a ‘–’ indicates that variable is negatively loaded on the PC. The figure also represents the estimated randomised/permuted (blue) and bootstrapped (red) distribution of the loadings for each variable. If an index loading (+/– sign) falls above the 90% confidence band for the null distribution (i.e., above the blue line), then we consider it to meaningfully contribute to the PC.⁷ Rotated Dimension 1 and Dimension 2 scores are split across PC1 (Figure 6A) and PC2 (Figure 6B), respectively. D1 scores are positively loaded onto PC1, and D2 scores are negatively loaded onto PC2. This means that speakers with a positive Label PC1 or PC2 score will have a positive/higher D1 score and a negative/lower D2 score in the MDS space (and the reverse for speakers with negative Label PC scores).

Moving to listener labels, the following label classifications also meet our 90% confidence band level threshold for the Label PC1: NZ rural (accent), Low, Middle, and High Socioeconomic status (SES), Strong NZ (accent), Distinctive vowels and quiet, Clear articulation, and creaky. Notably, the NZ rural, Low SES, strong NZ, distinctive vowels and creaky labels are positively loaded (with a + sign in Figure 6A) while the High/Middle socioeconomic status, quiet and clear articulation are negatively loaded (with a – sign in Figure 6A). This means that listeners describe speakers with positive Label PC1 scores (and lower rotated D1 scores) less frequently with labels related to higher socioeconomic status and clear speech, and more frequently with labels related to low socioeconomic status, rural and stronger NZ accents and creak. We therefore refer to these speakers as perceptually low SES (noting that, in the NZ context, speakers from rural areas may sound distinct from urban speakers and have high incomes).

Figure 6

(A) Variable Loadings for Principal Component 1; (B) Variable Loadings for Principal Component 2; (C) Perceptually Low and High SES speakers whose PC1 scores are 0.5 (low alpha) or 1 (high alpha) standard deviation above (+) or below (–) the mean; and (D) Perceptually Slow/Low pitch and Fast/High pitch speakers whose PC2 scores are 0.5 (low alpha) or 1 (high alpha) standard deviation above (+) or below (–) the mean.

Correspondingly, listeners describe speakers with negative label PC1 scores (and higher rotated D1 scores) more frequently with labels related to higher socioeconomic status and clear speech, and less frequently by labels related to broader accents and creak. We therefore refer to such speakers as perceptually high SES. While there are some labels related to speech features (namely creak and clear articulation) on Label PC1, this PC is driven by labels related to perceived accent strength and speaker socioeconomic status.

Label PC2 in Figure 6B is instead driven by labels related to perceived speed and pitch; fast and slow speech rates, high, middle and low pitch meet our 90% confidence band level threshold for PC2, with labels related to Old, NZ Proper (accent) and flat tone also just meeting it. Slower speech rates and lower pitch are positively loaded onto Label PC2, meaning that speakers with positive Label PC2 scores (i.e., lower/negative rotated D2 scores) are described more frequently by listeners with labels related to slower speech and lower pitch and old age, and less frequently by labels related to fast speech and higher pitch. We therefore refer to such speakers as perceptually slow/low pitch. Correspondingly, listeners describe speakers with negative Label PC2 scores (i.e., higher/positive rotated D2 scores) more frequently with labels related to faster speech rates and high and medium pitch, and less frequently with the positively loaded labels. We therefore refer to these speakers as perceptually fast/high pitch.

Figure 6C highlights the perceptually high and low SES speaker groups within the rotated perceptual space in Figure 4B, and Figure 6D highlights the perceptually slow/low pitch and fast/high pitch speaker groups. Specifically, speakers whose label PC1 and PC2 scores are more than 0.5 (lighter shade) and 1 (darker shade) standard deviation above or below the mean PC score are encircled. Perceptually low SES speakers (+ Label PC1 and D1 scores, low SES labels) are represented by yellow squares. Perceptually high SES speakers (– Label PC1 and D1 scores, high SES labels) are correspondingly represented by purple circles. Similarly, perceptually slow/low pitch speakers (+ Label PC1 scores, – D1 scores, slow speech rate labels) and perceptually fast/high pitch speakers (– Label PC1 scores, + D1 scores, fast labels) are represented by blue upside-down triangles and red diamonds, respectively. In other words, listeners tended to classify speakers in the yellow groups (squares) as sounding more rural and lower socioeconomic status, speakers in the purple groups (circles) as higher socioeconomic status, speakers in the red groups (diamonds) as faster and higher pitch, and speakers in the blue groups (upside-down triangles) as slower and lower pitch, with the darker shapes representing the speakers with the highest and lowest Label PC1/PC2 scores (i.e., the most perceptually low SES/high SES/slow/fast).

To summarise, we first categorised listener labels. Second, we constructed a two-dimensional perceptual similarity space by means of MDS. The MDS space situates the speakers grouped together by listeners more frequently as closer together (and vice versa). We rotated the perceptual space so that it was maximally aligned with the space in Sheard et al. (2025). We then applied PCA to the label category proportions and rotated D1/D2 scores for each speaker. The label PCA captured a distinction between primarily socially-oriented (PC1) and speech-orientated (PC2) approaches to labelling speakers, with D1 scores loaded onto former and D2 scores onto the latter. Finally, using speaker Label PC1 and PC2 scores, we mapped four perceptual groups within the MDS space: lower SES, higher SES, slow (and low pitch), and fast (and high pitch) speakers. The results of the label PCA therefore point to a potential socially meaningful distinction between speakers who are differentiated along D1, which we will explore further in the next section.

4.4. Connecting perception to production: Mapping variables to perceptual space

Which linguistic variables characterise the perceptual groups established in the previous section? Replicating the procedure in Sheard et al. (2025) and described in Section 4.4 of the supplementary materials, we fit two regression trees with stimuli D1 and D2 scores as the dependent variables and the four independent variables below:

• Stimulus articulation rate, operationalised as the total number of canonical syllables in the stimulus divided by its total phonation time (based on the CELEX dictionary and forced-alignment of the QB monologue in LaBB-CAT)⁸

• Stimulus mean pitch, as extracted manually from Praat (Boersma, 2001)

• The speaker’s position on the leader-lagger continuum (i.e., How far ahead or behind they are in the changes for fleece, dress, kit, trap, nurse, strut, goose, as represented by their QB1 Principal Component 1 score from Hurring et al. (2025) – see Figure 7A)

• The speaker’s back-vowel configuration (i.e., the relative positioning of their start, thought and lot vowels in the vowel space, as represented by their QB1 Principal Component 2 score from Hurring et al. (2025) – see Figure 7B).

To be explicit, the independent variables in this analysis are the same measurements as those in Sheard et al. (2025); the analyses are identical except for the dependent variables from the new MDS space. As such, the independent variables are here also scaled across the 38 speakers, and speakers’ leader-lagger and back-vowel configuration scores from Hurring et al. (2025) are based on the monophthongs they produced across their entire QuakeBox monologue, not just within the stimuli. We also note that we preregistered a series of pairwise correlations between speakers’ D1 and D2 scores and the independent variables rather than regression trees. Section 5 of the supplementary materials reports the preregistered correlations along with the results of random forests assessing the stability of the importance of the independent variables in predicting D1 and D2 (Section 4.4).

Figure 7

(A) The Leader-Lagger continuum and (B) back-vowel configuration from Hurring et al. (2025; supplementary materials). Speakers with a high leader-lagger score are in the green areas of the distributions shown in (A) and thus are leaders in a set of ongoing sound changes. Speakers with a high back-vowel-configuration score are in the green areas of the distributions shown in (B).

Figure 8A and Figure 8B display the results of the regression trees predicting D1 and D2, respectively. For each node in a tree, we can see the estimated dependent variable and the proportion and number of observations represented in the node. The if-else statements from regression trees provide cut-off values for the relevant independent variables in predicting the dependent variable, making them a highly useful tool for identifying distinct groups of speakers based on measures in their speech production. We can then map the identified production groups onto our perceptual groups (e.g., are speakers above a particular leader-lagger score cut-off concentrated in the perceptually low SES group?).

The most important predictor of D1, of the four variables, is a speaker’s position in the leader-lagger continuum. Laggers are estimated to have a lower D1 score. There is a mediating effect of speed within the leaders where slow leaders, specifically, are estimated to have a low D1 score while fast leaders are not. As in Sheard et al. (2025), articulation rate is the most important predictor of D2; slower speakers are estimated to have a lower D2 score. Pitch also plays a mediating role where the speakers estimated to have the highest D2 scores are both fast and higher pitch. The results of the random forest procedures provide additional evidence for the importance of pitch, speed and the leader-lagger continuum in predicting D1 and D2. The analyses therefore indicate all three variables contribute to the structural relationships between speakers who are perceived to sound similar and different to each other.

How do these results relate to our four perceptual groups? Figure 8A maps the cutoff values in the regression tree predicting D1 onto the rotated D1 and D2 coordinates, with the laggers in purple circles and the slow and fast leaders (with a leader-lagger score below/above –0.26) in yellow squares and orange triangles, respectively. Figure 8A also shows where the leader and lagger groups are situated relative to the perceptually low SES and perceptually high SES speaker groups, as determined by speakers’ Label PC1 scores. In general, most of the slow leaders are members of the perceptually low SES group in yellow (with some fast leaders), while most members of the perceptual high SES group are laggers. In other words, perceptually low SES speakers tend to be slow leaders while perceptually high SES speakers tend to be laggers. There are, however, laggers not in the perceptually high SES group, and most fast leaders are not members of the perceptually low SES group.

Figure 8

(A) The regression tree predicting D1 and its output mapped onto the perceptual similarity space. Speakers whose label PC1 is <0.5 SD +/– the mean have a lower alpha. (B) The regression tree predicting D2 and its output mapped onto the perceptual similarity space. Speakers whose label PC2 is <0.5 SD +/– the mean have a lower alpha. (C) Overall interpretation of the perceptual space based on the label PCA and regression tree results.

Figure 8B maps the cutoff values in the regression tree predicting D2 onto the rotated D1 and D2 coordinates, with slow speakers (here, those with an articulation rate below –0.11, blue asterisks) concentrated in the bottom of the space, fast high pitch speakers (with a mean pitch above 0.15, red diamonds) concentrated in the top left of the space, and fast and low pitch speakers in between (orange triangles). Figure 8B also shows where the slow and fast speakers are situated relative to the perceptually slow/low pitch and perceptually fast/high pitch groups as determined by speakers’ Label PC2 scores. The fast and high pitch speakers are, indeed, concentrated in the perceptually fast group in red, while most of the slow speakers are concentrated in the perceptually slow group in blue. In other words, the perceptually slower and lower pitched speakers tend to be slower and lower pitched while perceptually faster and higher pitched speakers tend to be higher pitched and fast. Fast and low pitch speakers tend to occupy space in between these two perceptual areas.

We now build on the separate analyses of the two dimensions for an overall interpretation of the two-dimensional space in Figure 8C. Taking the cutoffs for speed, pitch, and the leader-lagger continuum from the regression trees, there are four main production groups characterised by differences in speaker production:

• Fast and/or higher pitched speakers (red diamonds)

• Slow and/or lower pitched speakers (blue asterisks)

• Laggers (purple circles)

• Slow leaders (yellow squares)

Notably, unlike the groups above, the fast leaders (orange triangles) are not consistently together; some fast leaders are grouped more often with the slow leaders, others with the fast and high pitch speakers, including fast laggers. This is consistent with Sheard et al. (2025) and provides further evidence for the fast leaders being less perceptually distinct than the other groups. We can now map the four main production groups onto our four perceptual groups characterised by differences in speaker perception:

• Faster/higher pitched speakers are concentrated in the perceptually fast/high pitch group (red diamonds, top)

• Slower/lower pitched speakers are concentrated in the perceptually slow/low pitch group (blue asterisks, bottom)

• Laggers are concentrated in the perceptually high SES group (purple circles, left)

• (Slow) leaders are concentrated in the perceptually low SES group (yellow squares, right).

Listeners appear, therefore, to perceive slower/lower pitch and faster/higher pitch speakers as such. The results also point to a socially-driven perceptual distinction between leaders with slower speech rates and laggers; the former are perceptually rural, lower SES and broader, while the latter are perceptually higher SES and with clearer articulation (despite there being both fast and slow speakers in this group).

The combined interpretation of the perceptual and production speaker groups in the MDS space also provides insight into some of the additional labels loaded onto PC1 and PC2 from the label PCA. For example, most speakers in the perceptually low SES group (with higher D1 scores) are not only slower but lower pitched speakers and are included in the wider perceptually slow/low pitch group. Faster and higher pitched speakers are, in contrast, concentrated in the top right (and therefore have both higher D2 and lower D1 scores). It is, therefore, reasonable for the creaky labels to be loaded with the perceptually low SES speakers because of this contrast (it remains possible that perceived creakiness is, in and of itself, associated with being more rural/lower SES). Similarly, most perceptually slow/low pitch speakers are concentrated in the bottom right of the space, with as many laggers as leaders within this group. It is understandable for the slower and lower pitched speakers to be associated more with labels related to older age.

To summarise, we applied regression trees to speakers’ rotated D1 and D2 scores to see if we could meaningfully interpret the perceptual groups based on their covarying vowel patterns, or their speed and pitch. The analysis indicates a general alignment between perceived and actual speaker speed and pitch. The perceptually low SES and high SES groups are also characterised by speakers with different covarying vowel patterns. Listeners perceive (slow) leaders in change as being lower SES, more rural, and having more distinctive NZ accents, while they perceive laggers as being of higher socioeconomic status. However, there is not a 1:1 relationship between a speaker’s position in the leader-lagger continuum and the associated social meanings. Some fast leaders are perceived to sound more like other slower leaders, others more like fast laggers. Fast and high laggers are also less differentiated along the social dimension than slow and fast leaders (i.e., all laggers are on the left of the space, along with faster and higher pitched leaders).

5.1. Acoustic and sociolinguistic salience in listener perception

5.2. Comparison with the pairwise rating task

Our experiment replicates the result from Sheard et al. (2025) in that leaders were distinct from laggers within a two-dimensional perceptual similarity space. The main point of contrast for D2 in both the Free Classification and Pairwise Rating tasks was speed and then pitch. Slow speakers are concentrated together, and, within the fast speakers, higher pitched speakers are distinct from lower pitched speakers.

However, there are also some differences in how the leader-lagger vowel effect was mediated by other variables. In our free classification task, leader-lagger vowels were represented in the first node of the regression tree predicting D1, followed by articulation rate in the second node. In Sheard et al. (2025) speaker pitch was the first node in the regression tree predicting D1, with leader-lagger vowels represented in the second node within the lower pitch speakers. While the correlations in Section 4.2 indicate the overall structure of the two spaces is similar in terms of which speakers are perceptually distinct from each other (see also Section 5 in the supplementary materials), pitch emerges as more important in the pairwise comparison space.

This may highlight methodological differences in terms of what listeners are doing in pairwise versus grouping tasks. Work outside of linguistics has shown that pairwise tasks can encourage more use of low-level acoustic features in assessing similarity of sounds (e.g., a doorbell, sandpaper, rain, and snoring), while grouping tasks encourage greater use of categorical features (Aldrich et al., 2009). Correspondingly, in our similarity rating task, listeners don’t know the range of variation in the dataset when they rate speaker pairwise similarity on a continuum. This may lend itself to concentrating more on the gradient surface characteristics of the voices (such as pitch). In the free classification task, participants are instead “instantly calibrated to the full ranges of the important stimulus dimensions” (Hout et al., 2013, p. 274), due to multiple stimuli being presented simultaneously. Being able to more fully consider the types of variation present across stimuli to make their decision may lend itself to less continuous approaches, such as their perception of whether the speaker is ‘rural,’ or ‘high SES,’ or ‘fast and high.’

The socially interpreted leader-lagger continuum may, therefore, be more readily apparent to listeners (although mediated by speech rate) in the more categorical free classification task. In the pairwise rating task, gradient surface acoustic cues are instead more salient, and pitch becomes the primary cue. Since sharing high pitch leads to high perceived pairwise similarity while sharing low pitch does not (Sheard et al., 2025), perhaps high pitch speakers are so salient in the pairwise task that the more social leader-lagger vowels can only emerge as relevant in the lower pitch speakers. Nonetheless, the key point is that, although the different task mechanics likely led to differences in the relative salience of cues for listeners, the leader-lagger vowels emerge as perceptually salient across both perceptual spaces.

5.3 Limitations and future directions

Here we discuss the study’s limitations and the future directions suggested by our results. First and foremost, the experiment design lends itself to higher variability in listener responses. As all production work on covarying vowels in NZE has been based on uncontrolled speech from the ONZE and QB corpora, using uncontrolled audio stimuli allowed us to compare speakers’ actual position in the leader-lagger continuum and back-vowel configuration directly to their perception by NZE listeners. Long and uncontrolled stimuli, however, introduce potential confounds that shorter, controlled stimuli would not. This includes high variability in paralinguistic features which, as discussed previously, are perceptually salient to listeners.

The uncontrolled stimuli also meant listeners were not evenly exposed to variables from both covarying vowel clusters. The back-vowel configuration, which captures covariation between thought, start, strut (F2), and lot, has been found to be a robust and stable cluster of covarying vowels in speech production in New Zealand English (Brand et al., 2021; Hurring et al., 2025). However, the results of Sheard et al. (2025) and the analysis presented here suggest that listeners are not systematically using the back vowel cluster in their evaluation of speakers. One possible explanation is that the back vowels are simply not salient, or at least not as (socially) salient, as the leader-lagger vowels, speech rate and pitch. This explanation aligns with evidence for New Zealanders associating the realisations of some vowels in the leader-lagger continuum, particularly kit, with the New Zealand accent (e.g., Bell, 1997), but not for any of the back vowels.

However, as Sheard et al. (2025) note, the back vowels’ apparent lack of perceptual relevance may instead be an artefact of the limited back-vowel data to which listeners were exposed. Listeners heard more than three times the number of leader-lagger vowels than they did tokens of the back vowels, assuming they listened to the audio stimuli in full. Future work could explore the perceptual relevance of the back-vowel configuration with more controlled stimuli (e.g., by repeating the experiment using stimuli with the same pitch and speech rate throughout) which only exposed listeners to variation in the back vowel cluster, not the leader-lagger vowels.

We may also have had more success in prompting specific social associations from listeners by introducing firmer task parameters (as in the free classification work on American regional dialects). Our primary intent, given the lack of established hypotheses and the potential for the labels researchers use to affect listener perception of stimuli (Niedzielski, 1999; Wade et al., 2023), was to avoid imposing our own intuitions or expectations onto listeners. But the free classification task may have been overly free, resulting with listeners tuning into non-segmental as much as segmental features. The question of whether imposing social categories on listeners would have helped or hindered in achieving our aims is an empirical one, and there remains substantial scope for developing our understanding of when and what listeners tune into socially when the experiment design does not force them to make such associations.

Finally, while we have shown that a speaker’s position in the leader-lagger continuum is associated with judgements linked to socioeconomic status, we have not shown that all the individual leader-lagger vowels are contributing equally or at all to the percept of socioeconomic status. Specific vowels may encode different NZE accents more than others and, correspondingly, carry more social meaning for listeners. Our focus on speakers’ leader-lagger scores in the analysis could, therefore, have obscured the relative importance of different leader-lagger vowels in listener judgements.

There is also the possibility of variables other than the leader-lagger vowels (but that may nonetheless correlate with speakers’ leader-lagger scores) contributing to the perceptual differentiation between speakers. For example, speakers in the perceptually high SES group are associated with clearer articulation more than the slower, perceptually low SES speakers, despite there being more fast than slow laggers in the former group. This points to the potential contributing role of consonant production or suprasegmental features, which the leader-lagger scores nonetheless successfully capture. We are left, then, with questions such as could listeners make the same ratings if they could only hear realizations of one of the leader-lagger vowels? And, if so, could they do it to equal degrees regardless of the identity vowel? Future work is necessary to disentangle the precise mechanism underpinning the association between the leader-lagger continuum and perceived socioeconomic status.