1.1 Phonological limits on phonetic variation

In natural spoken interaction, listeners extract both linguistic properties, i.e., lexical identity and phonological structure, as well as indexical information about talkers from the speech signal. Talkers range from the routinely familiar to those who are just being encountered for the first time. Listeners will have varying degrees of familiarity with a talker’s accent, i.e., speech characteristics shared within a particular community. In the extreme case of an unfamiliar talker whose accent is being heard for the first time, listeners may require time to ‘tune in’ to the talker’s particular speech patterns in order to rapidly and accurately identify the intended spoken words. This process is supported by a phonological system abstracted over a lifetime of experience with phonetic variation within and across talkers (Best, 2015; Johnson, 1997; Pierrehumbert, 2003).

There is little controversy about the long-standing view that the lexical stock of a language is predominantly constructed from a finite and abstract system of phonological contrasts (e.g., Bloomfield, 1926; Browman & Goldstein, 1989; Chomsky & Halle, 1968; Goldinger, 2007; Pierrehumbert, 2016). However, work in recent years has highlighted the plasticity of phonological categories. A given speaker’s productions can change over time in concert with changes in the ambient speech environment (Tamminga, MacKenzie, & Embick, 2017) including exposure to languages (Flege, Schirru, & MacKay, 2003; Fowler, Sramko, Ostry, Rowland, & Hallé, 2008; Sancier & Fowler, 1997), dialects (Nycz, 2013), roommates (Pardo, Gibbons, Suppes, & Krauss, 2012; Sonderegger, Bane, & Graff, 2017), television (Stuart-Smith, Pryce, Timmins, & Gunter, 2013) or broader societal change (Harrington, Palethorpe, & Watson, 2000). On shorter timescales, shifts in production can be observed within a single conversation in response to the interlocutor (Kim, Horton, & Bradlow, 2011; Pardo, 2006) or within a single experimental session through shadowing (Nielsen, 2011) or even subtler manipulations (Shaw, 2013). In perception, listeners can rapidly adapt their phonetic expectations for phonological categories to the speech patterns of particular talkers (e.g., Cutler, Eisner, McQueen, & Norris, 2010; Norris, McQueen, & Cutler, 2003), including talkers with accents other than that of the listener (Bradlow & Bent, 2008; Maye, Aslin, & Tanenhaus, 2008; Sumner, 2011). They also perceptually compensate for contextual constraints on speech articulation, such as viewing the speaker holding a pencil between their teeth (e.g., Kraljic & Samuel, 2006).

Understanding the interplay between the phonological system by which the lexicon is structured, on the one hand, and the extensive talker, accent, and token variability that is associated with the realization of particular phonological categories, on the other, presents a fundamental theoretical challenge. Addressing that challenge is central to a broader understanding of the phonological systems of human languages, how they are learned, and how they change over time. Specifically, a crucial yet still unanswered (indeed, virtually unexamined) question is: For a given phonological category, what limits of tolerance prevail before listeners’ fine-tuned abilities to deal with natural phonetic variation start to derail—i.e., before they begin to attribute the phonetic form they have just heard to a different phonological category in their own regional accent than the category intended by the speaker?

The concept of perceptual assimilation may be useful in identifying these limits. Perceptual assimilation was originally proposed to account for cross-language speech perception findings (Perceptual Assimilation Model [PAM]: e.g., Best, 1995; Best & Tyler, 2007), where it refers to listeners’ strong tendencies to perceive non-native phones as exemplars of the well-entrenched phonological contrasts of their native language. Perceptual assimilation refers to the specific role that phonological categories play in shaping speech perception. In the L2 (or non-native speech) context, perceptual assimilation accounts for difficulties in discriminating many non-native contrasts, as well as for non-native contrasts that pose little to no discrimination difficulties (e.g., Best & Hallé, 2010; Best, McRoberts, & Goodell, 2001; Best, McRoberts, & Sithole, 1988). Here, we posit that perceptual assimilation of natural phonetic variation operates within the native language, where it allows listeners to accommodate novel phonetic variants they encounter to the system of phonological categories they have abstracted from their prior experience. The mechanism of perceiving speech in terms of phonological categories (see also Holt & Lotto, 2010), while it can cause specific difficulties for L2, may support flexible perception in the native language in crucial ways. Listeners are reasonably well able to recognize the great majority of words of a novel talker who speaks a previously unencountered and phonetically disparate regional accent of the same language. This ability may depend on perceptual assimilation, i.e., that listeners tolerate a wide range of talker variability in the expression of phonological categories. The benefits of perceptual assimilation in this regard should be clearest when listeners are able to detect systematic relationships among the categories of the full phonological system, rather than being presented with only a few phonemes or contrasts. Perceptual flexibility, in other words, could facilitate recognition of the more stable networks of phonological contrast that are, and must be, maintained across the entire system (Chodroff & Wilson, 2018).

Regional variation in English vowels provides a paramount example of higher level constancy, or phonological systematicity, in the face of phonetic variation. In a wide-ranging survey of regional accent variation in English, Wells (1982) introduced the Lexical Set framework to express the phonological coherence of the lexicon across varieties of English as spoken around the world. There are large swathes of the lexicon that share the ‘same’ phonological vowel category even as the particular phonetic qualities of the vowel may vary substantially not only from talker to talker, but also geographically from region to region. We refer to the latter type of variation as regional accents. Because of this state of affairs, the KIT vowel, for example, refers to how the vowel in all words using the same stressed vowel as <kit> is phonetically realized within each particular accent. There are, of course, cases of vowel mergers and splits that have taken place historically in some but not other regional accents of English, i.e., phonological differences. But these processes are the exceptions (captured in Wells’ framework with either separate lexical sets, e.g., BATH for words that pattern variably across accents with either TRAP or PALM, or lexical-indexical differences, e.g., <tomato> as [təˈmɛɪɾoʊ] in American English, in which the bolded vowel belongs to the FACE lexical set, versus [təˈmɐːɾoʊ] in Australian English, in which the vowel belongs to the PALM set). Overwhelmingly, the network of phonological contrasts supported by the English lexicon is preserved across regional accents despite substantial phonetic variation across accents in vowel quality realization for a given lexical set.

Our aim in the present study was to exploit the natural variation in vowel quality found across the vowel systems of five regional accents of English in order to assess the stability of the perceptual system in the face of variation. Specifically, we asked whether the differences across English accents described in the sociophonetics literature have consequences for the perception of vowels in terms of phonological categories. We began by assessing how vowels in various accents are perceptually assimilated, as indexed by vowel categorization and goodness-of-fit ratings (i.e., a listener judgement of how well a particular token fits into their vowel category), across the entire vowel inventory, 20 vowels per accent (Experiment 1). We then assessed how pre-test exposure to a meaningful story told in a given accent may influence perception of vowels in that accent (Experiment 2). Our experiments provided listeners with exposure from multiple talkers of either their own English accent or one of the unfamiliar (or, at least, less familiar) regional accents, which we have carefully selected to exhibit a range of relevant vowel variation.

As our goal was to delineate system-wide behavior, our experimental paradigm falls between two experimental research traditions (see Cristià, Seidl, Vaughn, Schmale, Bradlow, & Floccia, 2012 for an overview of many of the relevant studies). One tradition targets specific speech segments or individual phonemic contrasts, and assesses phonetic adaptation at the level of sub-lexical units (e.g., Evans & Iverson, 2004, 2007; Kraljic & Samuel, 2005; Maye et al., 2008; Norris et al., 2003; Sumner, 2011). These experiments typically introduce synthesized variation to some relevant acoustic aspect of the target item, providing a focused test of the perceptual effects of just the manipulated aspect of only the targeted segment(s). Another tradition uses naturally produced, connected, meaningful speech and assesses how exposure impacts overall comprehension, using dependent measures such as reaction time or transcription accuracy, without a detailed analysis of the phonetic or phonological characteristics of the speech for which perception is impacted by accent exposure (e.g., Bradlow & Bent, 2008; Clarke & Garrett, 2004; Floccia, Butler, Goslin, & Ellis, 2009; Floccia, Goslin Girard, & Konopczynski, 2006). Experiments in these two traditions have not been designed to pinpoint system-wide adjustments to phonological expectations in response to natural phonetic variation. Our approach adopts key aspects of each research tradition and adds several crucial comparisons. We provide pre-test exposure to either the listeners’ familiar (native) accent or an unfamiliar accent, in the form of an engaging story (cf. Maye et al., 2008). However, unlike previous phonetic adaptation research, our materials are not synthetically manipulated. Instead, we make use of naturally produced variation present in multiple speakers of regional accents of English. To assess how exposure affects each vowel in the system, unlike prior research into exposure effects on comprehension of connected speech, i.e., involving real words, our experiment following pre-test accent exposure probed perception of the entire vowel inventory in spoken nonce words.

In two experiments, we investigated the effect of accent variation and of pre-test accent exposure on categorizations of English vowels as spoken in five accents (described in greater detail below): Australian English (the listeners’ native regional accent) and the four less familiar accents spoken in London (UK), New Zealand, Yorkshire (UK), and Newcastle (UK). In Experiment 1, we examined Australian listeners’ perceptual assimilation of vowels in nonce words produced in the five accents, as indexed by categorization and goodness ratings of the vowels after they had first attended to the pre-test exposure story told by Australian speakers. Experiment 2 assessed whether, and if so how, categorizations and goodness ratings of those same nonce word vowels spoken in each of the non-Australian accents are influenced by pre-test exposure to the story, but this time told by other speakers of that same unfamiliar accent.

1.2 Hypotheses

1.3 Selecting the accents for use in the study

One of the unique aspects of our design, in contrast to the previous research from which we motivated our predictions, is that we sought to evaluate ‘accent-level’ perceptual assimilation of vowels and perceptual adaptation to the pre-test exposure story passage, as opposed to talker-specific assimilation and adaptation patterns. To this end, we tested listeners’ vowel categorizations for each of five accents following pre-test exposure to a passage in the listener’s native accent (Experiment 1: perceptual assimilation) versus following exposure to multiple talkers of the unfamiliar accent (Experiment 2: exposure effects), none of whom were the same talkers heard in the test phase (vowels categorization task) of the experiment.

The baseline of comparison, i.e., our listeners’ native regional accent, is Australian English (Aus). The principal patterns of phonological realization associated with this variety are well-established (Cox, 2006; Cox & Palethorpe, 2007), and in recent years subject to increasingly fine-grained analyses (Billington, 2011; Cox & Palethorpe, 2012; Cox, Palethorpe, & Bentink, 2014). Aus, as well the other regional accents included in this study, are non-rhotic—/ɹ/ occurs only in syllable onsets. Non-rhotic accents tend to have more vowels than rhotic accents because the sequence of a vowel followed by a post-vocalic /ɹ/ in a rhotic variety often corresponds to a distinct vowel category in a non-rhotic variety. Selecting all non-rhotic varieties simplifies the mappings between accents in that a vowel in one accent can always be mapped to a vowel in the corresponding lexical item in another accent (cf. mapping vowels in one accent to /Vɹ/ sequences in another accent). To represent Aus for the present study, we recruited participants from the south-western suburbs of metropolitan Sydney. Four further non-rhotic regional accents were then selected, each displaying a different array of vowel realization disparities from Aus, and thus enabling the formulation of distinctive hypotheses about the responses of Aus listeners to phonetic properties of vowels that differ phonetically, or in a few cases systemically, from their own accent: Christchurch in New Zealand, and southeast London, Yorkshire, and Newcastle in the UK. These regional varieties have all been extensively analyzed in recent years (New Zealand: Hay, Maclagan, & Gordon, 2008; London: Fox, 2015, Kerswill, Torgersen, & Fox, 2008, Przedlacka, 2001, Ryfa, 2013, Tollfree, 1999; Yorkshire: Haddican, Foulkes, Hughes, & Richards, 2013; Newcastle: Watt & Allen, 2003, Watt & Milroy, 1999). Henceforth we refer to these four varieties respectively as NZ, Lon, Yrk, and Ncl.

We used existing descriptions to define a set of phonological-to-phonetic mappings between Aus and each comparator accent, organized in relation to Wells’ (1982) lexical sets. For each lexical set in each pairwise accent comparison with Aus, we first identified whether the comparison to the reference accent (Aus) yielded what Wells describes as a systemic difference, i.e., where the two varieties differ phonologically with respect to the number of contrasts in the system (for example, two varieties that differ by virtue of one variety contrasting NEAR and SQUARE vowels and the other not doing so). We also considered what Wells terms realizational differences, i.e., differences between accents based on the phonetic variants used across a lexical set. How these characterizations of production influence cross-accent perception in terms of phonological categories is largely unknown. To make concrete the predictions of our first hypothesis, we expressed the differences of the Wellsian taxonomy in terms of the vocabulary of perceptual assimilations in PAM, as introduced in the hypothesis section above. We considered realizational differences to be CG (Category Goodness) differences when listener and talker accents for a given lexical set vowel are small and unlikely to lead the listener to hear a different vowel from the target lexical set than the other-accent talker intended (e.g., Aus listeners may be more likely to hear the retracted variants of TRAP in the Ncl accent as perceptibly non-Aus exemplars of TRAP rather than as representing a different lexical set, such as PALM). In contrast, a CS difference is one where the listener may be led to hear the variant in the talker’s variety as belonging to a different lexical set from that which the talker intended (e.g., the raised variant of DRESS found in NZ is likely to be misheard by Aus listeners as KIT). We reiterate that our a priori characterization of accent differences in terms of perceptual assimilation types, e.g, CG, CS, etc., is based on published descriptions of the accents and also verified by our own recordings (see below). Whether or not perceptual assimilation of accent differences cuts along the same lines as the sociophonetics descriptions (H1) or tolerates a broader range of variation (H2) remains an empirical question, addressed by our first experiment.

As the basis for expressing accent differences as types of perceptual assimilation, as well as drawing on recent descriptive accounts of the accents, we also made reference to the vowel productions in real lexical items of the speakers who we recruited for the experiments. As part of the experimental protocol, each of the four speakers (two male, two female) who produced the nonce words for the vowel categorization task in each of our target accents produced 3–5 repetitions of isolated words (the keywords used in the categorization task, described below) for each of the 20 lexical sets which we were interested in analyzing in this study.¹ The keywords were: <bead, bid, bed, bad, bard, bud, rude, hood, bored, pod, paired, beard, bird, toured, paid, code, hide, boyd, proud>, respectively, for the lexical sets FLEECE, KIT, DRESS, TRAP, PALM (same vowel as START in all five accents), STRUT, GOOSE, FOOT, THOUGHT (same vowel as NORTH in these accents), LOT, SQUARE, NEAR, NURSE, CURE, FACE, GOAT, PRICE, CHOICE, and MOUTH. From these keyword recordings, we extracted Lobanov-normalized (Lobanov, 1971) frequencies of F1 and F2 and re-projected the values in Hz, to normalize for speaker (Adank, Smits, & Van Hout, 2004). Plotting these values by accent and vowel enabled us to add a visual dimension to our comparative evaluations of the accent pairs. Figure 1 shows plots of these words, divided into monophthongs (upper panels) and diphthongs (lower panels). For monophthongs, the ellipses show the 95% confidence intervals (CIs) across speakers of Lobanov-normalized F1/F2, measured at the vowel midpoints; for diphthongs, the arrows show the Lobanov-normalized F1/F2 values at 20% and 80% of vowel duration. The separation into monophthongs and diphthongs is based on classical designation. Note that which vowels are actually realized as diphthongs differs between Aus, Lon and NZ as a set, versus Ncl and Yrk as another set. Also note that each accent shows dense regions with numerous overlapping vowels, and sparser regions with much less overlap. These formant patterns are largely consistent with published descriptions on each accent (see citations above). We now summarize the findings of our pairwise accent comparisons, which will refer both to lexical sets, expressed in small caps, e.g., DRESS, and the recordings of the keywords displayed in Figure 1. In our discussion, we designate keywords with “<>”, e.g., <bed>. We have ordered the discussion from the most to the least similar vowel systems of the other accents relative to Aus.

Figure 1

Aus, Lon, NZ, Yrk, and Ncl nonce speakers’ keyword vowels (3–5 tokens per keyword per speaker × 2 males and 2 females per accent) for the nominal (classically-defined) monophthongs (top panel) versus diphthongs (lower panel) of English. Ellipses for monophthongs represent 95% confidence intervals around the Lobanov (1971) normalized values of F1/F2 at 50% of vowel duration; colored arrows for diphthongs represent change in formants from 20% of vowel duration to 80%.

3.1 Experiment 1: Perceptual assimilation of vowels across accents

3.1.1 Vowel perception in native Aus

Figure 2 shows Accuracy’ by vowel for the Aus baseline (A-A condition). Error bars indicate 95% confidence intervals. The vowels are ordered from highest Accuracy’ on the left side of the figure to lowest Accuracy’ on the right. The highest scores were in the 10s for the vowels of keywords <pod> and <bard>, corresponding to raw accuracies of 80% and 64% correct vowel choices, respectively; the three highest raw accuracies also included <bed> at 80% correct but its Accuracy’ was below 10, reflecting the high overall frequency of choosing that keyword across the experiment. The lowest Accuracy’ score was –9 for <toured>, with the lowest raw accuracy of 21%. Raw accuracy across vowels and conditions is provided in Appendix C. Confusion matrices are provided in Appendix E. Even the highest Accuracy’ (and raw accuracies) are not near ceiling levels, underscoring that this task is challenging for listeners even for native-accented vowels. We observe a positive mean Accuracy’ with error bars that do not overlap zero (see also leftmost column of Table 1), for seven Aus vowels (those of keywords <pod, bard, boyd, past, bid, bed, code>). This indicates that, for these seven vowels, participants chose the correct vowel more often than the sum of all alternatives. This was not the case for 12 other vowels, which have error bars that overlap zero. These vowels were just as likely to be categorized as other vowels as they were to be categorized as the target vowel. At the negative end of the continuum, Accuracy’ fell significantly below zero for one Aus vowel, <toured> (lexical set CURE), indicating it was significantly less likely to be identified as the CURE vowel than as some other vowel(s). As we discussed in the introduction, for many Australians this vowel may no longer constitute a separate category (i.e., CURE has undergone or is undergoing merger with THOUGHT/NORTH).

Figure 2

Mean Accuracy’ scores (error bars = 95% Confidence Intervals) obtained for categorization and goodness ratings of each vowel in each accent in Experiment 1 (Perceptual assimilation). The dotted gray lines in the non-Australian accent plots provide the reference levels of Accuracy’ found in the A-A condition, for ease of visual comparisons across accents.

Table 1

Linear mixed model results on Accuracy’ values for: a) A-A condition vowel coefficients; b-e) Accent * vowel coefficients. Other-accent vowels with significantly lower (negative) or higher (positive) Accuracy’ values than the corresponding Aus vowel are designated by boldface.

Keyword Vowel	(a) Australia			(b) London			(c) New Zealand			(d) Yorkshire			(e) Newcastle
	β	t		β	t		β	t		β	t		β	t
<pod>	8.00	4.07	**	–4.13	–1.38		–2.39	–0.79		–3.23	–1.03		–1.78	–0.58
<bard>	7.63	3.88	**	2.47	0.82		–0.69	–0.23		–4.15	–1.33		–9.56	–3.12	**
<boyd>	6.04	3.07	**	3.12	1.04		3.07	1.02		4.48	1.43		6.89	2.25
<past>	5.69	2.89	**	2.17	0.72		6.95	2.30		–19.79	–6.32	**	–15.97	–5.22	**
<bid>	5.35	2.72	**	–5.49	–1.83		–9.24	–3.06	**	0.60	0.19		–2.81	–0.92
<bed>	5.05	2.57	**	–0.17	–0.06		–16.29	–5.39	**	0.45	0.15		–5.26	–1.72
<code>	4.67	2.38	**	–4.73	–1.58		2.62	0.87		–11.47	–3.66	**	–12.04	–3.93	**
<paid>	2.47	1.25		2.19	0.73		–1.12	–0.37		–2.64	–0.84		–8.81	–2.88	**
<hide>	1.89	0.96		3.40	1.13		9.97	3.30	**	7.06	2.25		3.67	1.20
<bead>	0.79	0.40		6.86	2.28		3.01	1.00		5.58	1.78		1.98	0.65
<bad>	–1.26	–0.64		1.82	0.61		–0.34	–0.11		5.75	1.84		1.53	0.50
<proud>	–1.55	–0.79		4.16	1.39		0.27	0.09		–2.07	–0.66		3.40	1.11
<bored>	–2.40	–1.22		0.12	0.04		0.98	0.32		2.12	0.68		5.32	1.74
<bird>	–2.52	–1.28		–4.01	–1.33		0.05	0.02		2.79	0.89		6.41	2.09
<paired>	–3.52	–1.79		–0.68	–0.23		–5.15	–1.71		2.12	0.68		3.09	1.01
<beard>	–3.93	–2.00		–3.69	–1.23		–4.36	–1.44		3.60	1.15		3.74	1.22
<bud>	–3.94	–2.00		3.66	1.22		4.09	1.35		–5.81	–1.85		–0.68	–0.22
<hood>	–7.23	–3.67	**	–10.08	–3.36	**	2.94	0.97		2.35	0.75		5.08	1.66
<rude>	–8.38	–4.26	**	–0.77	–0.26		2.27	0.75		9.35	2.99	**	6.85	2.24

• Note: <toured> is absent from this table as it was the vowel excluded for the purpose of deviation coding in the models.

Table 1 summarizes the fixed effects in the mixed models. We have also included in the two lefthand columns the vowel coefficients for a mixed model of the A-A condition (excluding the effect of accent, in this case), so as to quantify the impressive range of Accuracy’ found even within the baseline condition. The columns for London, New Zealand, Yorkshire, and Newcastle show coefficients for the vowel*accent interaction term in the models described above. We considered t values ≥ |2.325| (df = 42, p < .025, two-tailed, 97.5% CI) to reflect significant differences in Accuracy’ values between a vowel in a given accent and the corresponding Aus vowel (these are indicated by double asterisk in the table).

3.2 Experiment 2: Accent exposure effects

To evaluate the effect of pre-test exposure accent, we compared vowel categorizations in the conditions in which listeners heard the exposure story in their native and most familiar accent (Aus) before completing the nonce vowel categorization task in one of the other (unfamiliar, or less familiar) accents against the conditions in which they heard the story in the same (unfamiliar) accent as the nonce stimuli, e.g., a comparison between A-L and L-L conditions. We thus evaluate whether accent-matched pre-test exposure facilitates performance in the vowel categorization task when the talkers in the two phases of the experiment are different.

Figure 3 summarizes the exposure effects on vowel Accuracy’ scores. For this comparison, the Aus-other accent conditions (shown in white bars), e.g., A-L, serve as the baseline for the matched exposure conditions (shown in grey bars), e.g., L-L. For reference, the top panel repeats the Aus baseline results (A-A condition) and the grey lines indicate those results in each panel.

Figure 3

Mean Accuracy’ scores (error bars = 95% Confidence Intervals) obtained for each nonce vowel in each accent in Experiment 2, after story exposure in Aus versus after story exposure in the same unfamiliar accent of the nonce words task. The dotted gray lines provide the reference levels of Accuracy’ found in the A-A condition (familiar accent), for ease of visual comparison across accent exposure conditions.

The main pattern we observe is that the matched exposure conditions (grey bars), e.g., L-L, closely resemble the unmatched conditions (white bars), e.g., A-L, for each accent. This pattern indicates that, by and large, pre-test exposure to the story in an unfamiliar accent did not influence patterns of perceptual assimilation of vowels produced by new talkers of that accent. We evaluated that observation statistically by fitting a series of three nested linear mixed effects models to the eight conditions relevant to assessing exposure effects (condition A-A was excluded), which are represented in the bottom four panels of Figure 3. The baseline model, defined in (i), contained vowel as the only fixed factor. The second model, the ‘accent’ model in (ii), added the accent of the test stimuli and the interaction between accent and vowel as predictors. The third model, the ‘exposure’ model in (iii), added exposure accent as a third factor along with all interactions among exposure, vowel, and accent, including the three-way interaction. All models contained a random intercept for subjects. Model comparisons, conducted via pairwise ANOVAs, are summarized in (iv) of Table 2. Adding accent to the model resulted in significant improvement over Baseline (vowel only). Moreover, the increased model complexity resulting from pairwise interactions between accent and vowel is justified by the lower AIC (Akaike Information Criterion) in the accent model than the baseline model. Adding exposure to the model together with interactions between exposure and the other factors, vowel and accent, increases the complexity of the model substantially, from 82 to 162 degrees of freedom. Yet, despite this substantial increase in degrees of freedom, there is only a marginal reduction in residual deviance in the exposure model. The differences between the accent and exposure models is not statistically significant. Moreover, the AIC of the exposure model is higher than the AIC of the accent model, indicating that the small increase in variance explained does not justify the additional complexity of the exposure model.

The model comparison thus confirms the trend shown in Figure 3. Pre-test accent exposure of the type that we provided our listeners did not have a significant effect on vowel categorization of nonce words (produced by new talkers of the same accent). Focusing on just those vowels that showed significant negative accent effects in Experiment 1 (9 out of 80 possibilities) does not render a different result. Some of these vowels show a small increase in Accuracy’ with exposure—Lon <hood> (FOOT) and NZ <bed> (DRESS) both show small improvements with exposure, as do Ncl vowels <bard> (START), <code> (GOAT), and <paid> (FACE). However, other vowels show no effect or effects in the opposite direction (negative effects of exposure)—the latter include NZ <bid> (KIT), Yrk <past> (BATH) and <code> (GOAT), and Ncl <past> (BATH). Thus, of the nine vowels that show significant negative effects of accent, five show small positive effects of exposure while four show small negative effects of exposure. When we consider exposure effects, the subset of vowels that showed significant accent effects echoes the broader trend in the data. Some combinations of vowel and accent show marginal increases and some show marginal decreases. The overall pattern cannot be distinguished from randomly distributed noise around a null effect of accent exposure.

From the literature on perceptual learning for speech, we had extrapolated several hypotheses about possible exposure effects in this task. One of these was that listeners would improve after pre-test exposure to the target accent because they learn phonetic patterns characteristic of the accent that generalize to new talkers and to nonce words. Prior perceptual learning studies have shown generalization across words within talkers (e.g., Norris et al., 2003) and generalization to new talkers (e.g., Kraljic & Samuel, 2006). Exposure to multiple talkers has also been shown to improve perceptual learning of a dialect, when the task is to identify that dialect (Clopper & Pisoni, 2004), and, of course, naturalistic exposure to an accent can lead to changes in vowel categorization decisions (Evans & Iverson, 2004). However, we are not aware of any positive effects of accent exposure involving the level of generalization in our task, i.e., across talkers and from words in context to non-words in vowel categorization. Our results thus suggest a limit on perceptual learning.

There are many possible reasons for the absence of learning in this paradigm. It is possible that the vowel categorization task was not sensitive enough—too abstract or too difficult—to pick up perceptual learning effects. Listeners may have shown more reliable positive exposure effects in a task that did not focus on the abstract phonological category, such as sentence transcription (Baese-Berk, Bradlow, & Wright, 2013; Bradlow & Bent, 2008) and sentence verification tasks (Adank et al., 2009). It could also be the case that multi-talker exposure does not aid perceptual learning of accent-based vowel differences (cf. Clopper & Pisoni, 2004). In addition, listeners might have benefitted if they had been provided with explicit labelling of the accent or instructions that the pre-test exposure accent and the nonce task accent were the same/different. Additionally, although we screened participants for exposure to the accents in our study via a questionnaire, we cannot rule out the possibility that listeners may have been ‘pre-adjusted’ to the accents due to prior exposure of which they were unaware. We also point out that since vowels carry indexical information about the talker, listeners may have attributed some of the vowel variation in the pre-test exposure story to the particular talkers, which may also have interfered with accent-based generalization to the nonce words of new talkers of that accent in the categorization task.

In the General Discussion, we return to these points and to some trends regarding how Accuracy’ varied for vowels and accents across the experiments. First, however, we consider whether the goodness ratings for correct/accurate responses alone might reflect some sensitivity to accent differences that was missed by our Accuracy’ index.

3.3 Goodness Ratings

In this section we analyze the goodness ratings, in order to establish whether there is any residual information in their patterning that cannot be detected by the Accuracy’ scores. Figures showing goodness ratings by vowel and condition are reported in Appendix D.

We fit a linear mixed effects regression model to the raw ratings of accurate responses. Random effects were participant (listener) and talker, and we tested a three-way interaction among accent, exposure, and stimulus word. ANOVA comparison with a model containing all two-way interactions reveals the three-way interaction to be highly significant (p < .0001).³ The model interaction is shown in Figure 4. Separate models fit to each accent show significant interactions between exposure and accent for Yrk, Lon, and NZ, and significant main effects of exposure for NZ (a negative effect) and Lon (a positive effect).

Figure 4

Significant interaction among stimulus word, condition and accent in the analysis of raw goodness ratings for correct (accurate) vowel categorizations alone. The x-axis labels correspond to the words provided to listeners as response categories for the nonce word stimuli.

The effect for the NZ accent is particularly salient, and differs clearly from the other three accents: Pre-test story exposure to NZ speakers leads Australian listeners to rate all NZ nonce vowels (except the NURSE vowel) as ‘less good’ than the same NZ nonce vowels following exposure to the story told by Aus speakers. If the goodness ratings were providing a channel of information about the perceptual mapping, then it is hard to understand why goodness ratings of correct responses would drop to the NZ voices after a period of exposure to the accent. Our interpretation of this result is that it does not tell us directly about the goodness of the mapping, but rather, reflects an attitudinal response to the NZ accent. Aus listeners are likely to encounter NZ speakers in the Sydney environment more often than they would typically encounter speakers of the UK accents. Therefore, when they heard the NZ accent in the exposure period, this provided sufficient information to identify the speakers as New Zealanders. We speculate that they then go on to infer that the voices in the perceptual categorization task are also New Zealanders. Past work has shown that Australian and New Zealand listeners readily recognize each other’s accents (Weatherall, Gallois, & Pittam, 1998). Moreover, Aus listeners perceive the way that New Zealanders pronounce vowels to be ‘less good’ than Aus vowel pronunciations, so this identification of the speakers as New Zealanders leads to a wholesale drop in vowel ratings. Australians’ greater degree of experience with New Zealand English compared to the other accents in the study goes along with some stereotypes and negative attitudinal responses. Australian material in the media and on the internet indicates an awareness of New Zealanders’ vowel pronunciations, e.g., “where ‘Jungle Bills’ is what you sing at Christmas time.” Although the results are somewhat controversial, there are studies suggesting that Australians do not hold the NZ accent in high regard (Bayard, 2001), including an open response survey that reveals a pattern of ‘reciprocated negativity’ between New Zealanders’ and Australians’ responses to each other’s accents (Garrett, Williams, & Evans, 2005, p. 223).

If this drop is not about perceptual adaptation but rather about attitude and experience, as we propose, we can make two further predictions. First, we might expect to see a wholesale drop in the non-accurate responses, too. And second, the effect might be mediated by the individual respondent’s own reported experience with New Zealand English.

To test the first prediction, we tested for an interaction between stimulus word and exposure within inaccurate responses to each accent. No interactions or main effects of exposure were found for the Lon or Ncl accent. For Yrk, there was a significant interaction between stimulus word and exposure, but no main effect of exposure. For NZ, there was both a significant interaction and a significant main effect (p < .02). That is, for the NZ accent, but not for the other accents, there is a significant overall drop in goodness, even for the incorrect answers.

To test the second prediction, we entered information about participants’ previous experience into our models. Our background questionnaires asked each listener to self-report any exposure they had to the target accents. This was designed to screen out participants with high levels of exposure, for example from family members or other cohabitants. However, the analyzed participants also contained some variation in their answers to this question. We went through the open responses and created a binary code, separating those reporting no exposure, or very little (e.g., a one-week holiday, or occasional interaction with customers), from those who reported greater exposure (e.g., regular interactions with friends or more distant relatives). There were too few participants with exposure to Yrk or Ncl accents, making analysis of this factor impractical for those accents, and the exposure pattern was imbalanced for the Lon accent (with relatively few having no exposure in the AL condition). However, the distribution for the NZ accent made analysis of this variable practical [17 low experience (10 in Z-Z), and 15 high experience (6 in Z-Z)].

A logistic mixed effects model of the NZ goodness ratings to accurate responses revealed a significant three-way interaction between stimulus word, experimental condition, and previous experience, as shown in Figure 5. When the exposure accent is Aus English, participants with greater previous experience to NZ English hear the vowels as better examples of their category than participants with less experience do. Both groups drop their goodness ratings following the NZ exposure passage, and the distribution of this drop across words differs somewhat for the two groups. Thus, it seems as though goodness ratings are tapping into something different from the categorization results, which limits their usefulness for sharpening our conclusions about perceptual assimilation.

Figure 5

Interaction plot showing significant interaction among stimulus, experiment condition, and previous experience in the raw goodness ratings for correct (accurate) vowel categorizations, for listeners responding to the NZ accent following the pre-test story in Aus versus in NZ. The labels of the x-axis correspond to the words provided to listeners as response categories for the nonce word stimuli.

4.1 Baseline performance on the vowel categorization task

The vowel categorization task was hard for listeners. Accuracy’ was well below ceiling for all vowels and accents. Thus, listeners did not approach perfect consistency in choosing the correct vowel from the grid for any of the vowels, even when presented in the native accent of the listeners. Part of this low accuracy is likely due to the less categorical nature of vowel perception as opposed to consonants (e.g., Pisoni, 1975; Repp, 1984). Indeed, we have observed sharper categorization boundaries in this type of task for consonants (Shaw, Best, Mulak, Docherty, Evans, Foulkes, & Hay, 2014). In general, it is known from other tasks, such as word reconstruction, that English listeners tolerate more variation in vowels than in consonants (Van Ooijen, 1996). Despite this, we find it unlikely that a native listener would fail to recognize, e.g., minimal pair words differentiated by vowels in their native language. The experimental task that we employed here, however, is much more difficult than minimal-pair word recognition, in that it provided the listener with only the bottom-up signal (lacking any lexical or phonotactic referent), and offered 19 possible responses (all of the monophthong and diphthong vowel categories in Australian English). For starters, there are very few phonotactic environments in English in which all vowels are attested, so the practical task of selecting amongst 19 vowels in a constant phonological frame, such as the one that we posed to the participants, rarely occurs ‘in the wild.’ Moreover, vowel identity is typically cued by the lexical items that contain the vowels. We designed our task to eliminate all lexical as well as phonotactic information (by placing the vowels in the phonotactically licit but rare carrier frame, /zVbə/), leaving the listener with only the phonetic signal to go on. The listeners’ job was made even more difficult by the task itself in that it requires an abstract meta-linguistic judgement about phonological categories, and goodness of fit to those categories. Moreover, the materials in the nonce categorization task drew from four different speakers of the target accent randomly presented (cf. a blocked design), and provided no phonetic context to adjust to the speaker beyond the /zVbə/ frame in which the target vowels were embedded. Under these circumstances, it is not surprising that vowel categorization is below ceiling. We interpret the low baseline performance in Aus (the A-A condition) as an indirect indication that other factors, including lexical item, phonotactic probabilities, and talker information likely contribute substantially to the ease with which we normally recognize contrastive vowels in words and larger meaningful utterances. Certainly the indeterminacy associated with the nearly pure bottom-up categorization our task required leaves plenty of room for other factors to yield higher accuracy in recognizing vowels in spoken words.

4.2 Effects of accent variation

Although we did indeed find some cases in which phonetic variation in vowels, as described in the literature and verified in our own recordings, influenced listener perception, the results show that vowel perception is largely consistent across accents. Phonetic differences across the accents for a number of vowels, including some that are stereotypical of particular accents, failed to result in significant miscategorizations above and beyond the miscategorization of those vowels in the native accent. For example, despite the overlap between FOOT and STRUT in Yrk, Australian listeners did no worse in categorizing and rating Yrk STRUT than they did on their own Aus STRUT vowel. From the standpoint of our first hypothesis, that accent differences map to perceptual differences, this is surprising. Rather, this result favors the alternative—that vowels with different realizations across accents are nevertheless assimilated to the same phonological category. Listeners appear to assimilate substantial variation, even variation on the magnitude that has been described as systemic differences across accents. Flexibility in the linguistic interpretation of the signal may, complementarily, also leave room for perception of indexical information, e.g., properties of the talker that are also available in the signal (Best, 2015). That is, non-commitment (or minimal commitment) to the linguistic identity of the signal, in the presence of variation and the absence of lexical information, may facilitate parsing phonetic variation in terms of talker-specific characteristics (see also Shaw, 2016). Alongside phonological categories, listeners may also be abstracting social categories of the talkers from phonetic variants (as argued in Szakay, Babel, & King, 2016), a process which may exact an encoding cost (Clopper, Tamati, & Pierrehumbert, 2016) but which, conversely, may provide benefits when the relevant characteristics of the talkers are known.

After correcting for biases in overall likelihood of selection of specific keywords we found that the pattern in Accuracy’ across vowels was fairly consistent across accents. This may indicate differences in perceptual assimilation strength across phonological categories, in this case among the vowels of English. Perceptual asymmetries are not unusual in the vowel perception literature (Polka & Bohn, 2003, 2011; Schwartz, Abry, Boë, Ménard, & Vallée, 2005), a fact that is expected if vowel categories have different perceptual strengths. There are also converging results from the phonology literature, where there is a growing awareness that phonological categories are more graded in both production and perception than have been typically assumed (Hall, 2009; Smolensky, Goldrick, & Mathis, 2014; Warner, Jongman, Sereno, & Kemps, 2004). Factors affecting the strength of perceptual assimilation (i.e., tolerance of variance) may include the predictability of the vowel in context (Shaw & Kawahara, 2017), functional load (e.g., Hockett, 1967), phonological neighborhood density (Munson & Solomon, 2004), or phonetic salience (Carlson, Granstöm, & Klatt, 1979; Escudero & Vasiliev, 2011; Rosner & Pickering, 1994), factors that we plan to pursue in future studies.

Amongst the backdrop of a general consistency in the patterning of Accuracy’ by vowel across accents, there were a small number of significant negative vowel*accent interactions. These are cases in which changing the accent leads to significantly lower Accuracy’ for a vowel. A total of nine such cases (out of a possible 80) were revealed in our data. These include a subset of the differences that we foregrounded in our review of the accents, but many of the other differences documented in the literature and present in our materials failed to result in significantly lowered Accuracy’. For example, the centralized <bid> (KIT) vowel and the raised <bed> (DRESS) vowel in NZ both led to lower Accuracy’, but the NZ NEAR-SQUARE merger (<beard>~<paired>) and raising of TRAP (<bad>) did not. Similarly, for both Yrk and Ncl, Accuracy’ was lower for BATH (<past>) and GOAT (<code>), as expected from our review of these accents in the introduction. Yet the lack of contrast between FOOT~STRUT in these two accents did not cause a significant decrease in Aus listeners’ Accuracy’ for either vowel in either accent. In these cases, Aus listeners showed perceptual flexibility, that is, tolerance of the variation. In particular, it is intriguing to note that both of the two systemic differences among the vowel systems of our selected accents (lack of FOOT-STRUT split in Yrk and Ncl; NEAR~SQUARE merger in NZ) failed to result in significant accent differences in Accuracy’. Familiarity with an accent might be at play here and is known to affect perception (Sumner & Samuel, 2009). Although we screened our participants for relevant exposure, it may be that these more stereotypical accent differences are known within our listener population despite lack of direct reported exposure, a hypothesis that could potentially be tested by drawing on methods from the folk linguistics literature (e.g., Niedzielski & Preston, 2000). However, we note that all of the vowels involved in the systemic differences we have described here showed fairly low Accuracy’ even in the baseline A-A condition. They are among the 10 Australian vowels that overlap with 0 Accuracy’, i.e., that within the context of this task naturally tolerate mutual variation, even by listeners of accents in which they are not merged. This is probably for reasons that are related to our discussion of perceptual strength, including properties of the lexicon that are likely to be shared to a large degree even across our geographically dispersed accents. The perceptual assimilation results may be suggestive—either independently or in tandem with other lexical factors—of which vowels are more/less likely to be susceptible to change over time. At the least, we speculate that toleration of phonetic variation, which varies from vowel to vowel, may contribute to the perceptual landscape that gives rise to accent differences diachronically.

Returning to our PAM-inspired characterization of assimilation patterns as either CS (category shifting) or CG (category goodness) differences, as laid out in the introduction, none of the CG accent differences we had foregrounded had a significant effect on Accuracy’. Recall that the CG differences are those that sociophonetic descriptions recognize as different phonetic realizations across accents. These realizations contain information about the regional accent of the talkers but the phonetic variation does not significantly influence categorization of the vowels. When it comes to mapping a variable phonetic form to a phonological category there is unity in the face of regional variation. The significant negative effects in our data were all CS differences and, as we have already indicated, not even all of the CS differences showed significant deviations in Accuracy’. Overall, perceptual flexibility seems to subsume much of the variation that has been discussed as distinguishing regional accents of English. From the standpoint of our two hypotheses on how perceptual assimilation relates to cross-accent differences, we interpret this result as greater support of the second hypothesis—as it predicts that a fair range of variation in relevant phonetic dimensions should be assimilated to the same phonological category in perception.

4.3 Resilience to exposure

In light of the observed accent effects, we now turn to the topic of exposure, and how relatively brief but immediate pre-test exposure to a less familiar accent impacted perceptual assimilation. Our exposure manipulation involved a meaningful story told round-robin by four speakers. Multiple talker exposure of this degree has been shown to support learning in the L2 perception literature for some contrasts (Bradlow, Pisoni, Akahane-Yamada, & Tohkura, 1997; Lively, Logan, & Pisoni, 1993; Wang, Spence, Jongman, & Sereno, 1999) and in some studies of foreign accent adaptation (Baese-Berk et al., 2013; Bradlow & Bent, 2008; Sumner, 2011). However, given the small number of vowels for which accent actually had a significant effect, the opportunities to make perceptual gains from exposure at the level of the accent were also relatively small. In numerical terms, the effect of exposure showed the greatest positive benefit for Ncl, which makes sense given that this accent was the most phonetically distinct from the listeners’ accent, based both on our review of the sociophonetic literature and on listener performance in Experiment 1. There may be no deeper reason for this except that the larger effect of accent leaves greater potential for improvement. Still, the effect of exposure was not significant overall, nor was the interaction between exposure and accent or the three-way interaction among accent*exposure*vowel. In short, exposure had little effect on Accuracy’ for categorization and rating of nonce vowels in any of the other accents in this experiment. There may be several reasons for the lack of improvement in vowel categorization following exposure.

We reiterate that the talkers who produced the nonce word items in the perceptual categorization experiment were different from the talkers who produced the exposure passage in all conditions. Some of the classic effects of perceptual learning for speech are now known to be talker-specific. For example, rapid adaptation of fricative categories following exposure to an ambiguous segment as reported in Norris et al. (2003) and in Kraljic and Samuel (2005) does not generalize to new speakers (Eisner & McQueen, 2005), although variation in stop consonants can (Kraljic & Samuel, 2006). Moreover, there is evidence that multiple talkers can be tracked over the course of the experiment (Kraljic & Samuel, 2007), further suggesting that adaptation may be talker-specific. Paradigms that provide more naturalistic exposure (as opposed to specific manipulation of a single contrast) have not reliably shown short-term adaptation (Floccia et al., 2009). Benefitting from exposure in our paradigm requires adapting to naturalistic speech at the level of the accent as opposed to adapting only to specific talkers. From the standpoint of the Ideal Adapter (Kleinschmidt & Jaeger, 2015), accent-based adaptation takes the form of generalization to new talkers and is predicted to occur as a function of similarity between talkers. An additional complication for listeners (and for the Ideal Adapter) is that, the exposure story and the categorization task stimuli were different not only in the specific talkers that produced them but also in style. The exposure passage involved running speech and a meaningful story; the nonce words were produced in isolation. There may have been accent level adaptation that simply did not generalize to the nonce words. Additionally, since nonce words provide no further information to ‘reinforce’ perceptual adjustments stemming from exposure, listeners may quickly rebound to baseline expectations for vowels in the nonce context. That is, the choice words representing lexical sets on the screen may have triggered long-term representations of the corresponding vowels as opposed to talker- or accent- specific adjustments.

Relative to studies that have documented rapid adaptation to talkers and/or accents, our study is unique in several critical ways. Instead of isolating a few phonetic dimensions and using artificial manipulation of the speech signal, our study featured natural variation found across talkers and accents of English and throughout its vowel system. The phonetic properties of our materials varied and co-varied on numerous dimensions, as they do in naturalistic speech contexts. We gave our listeners the opportunity to rapidly adapt, but only at the level of the accent, not at the level of the talker. In all conditions, the talkers in the test phase of our experiment were different from the talkers of the exposure story. Benefiting from the accent exposure in the story in the vowel categorization experiment would require abstracting phonetic properties of a given accent from the exposure story. The results of this experiment produced no evidence for such adaptation effects, and indeed, experimental evidence for accent-level adaptation to phonological categories, independent of the words that contain them, on such a short time scale is indeed slim—we have found no such cases in the literature. Listeners may naturally attribute phonetic variation, such as that represented in our materials, to properties of the specific speakers and only adapt to the accent at longer timescales than are feasible to observe in a laboratory experiment (cf. Evans & Iverson, 2007). More broadly, we surmise that non-commitment to the bottom-up phonological categorization of the acoustic-phonetic signal in terms of particular vowels should actually benefit lexical recognition when the message presents real words (recovery of the message) and, moreover, may support listeners’ ability to attribute phonetic variation to indexical properties of the talker rather than to linguistic properties of the utterance.

4.4 Sensitivity of vowel goodness ratings to attitudes about accents

As reported above, analyses of the goodness ratings alone, unlike the analyses of Accuracy’, revealed a significant effect of accent exposure on judgments about the quality of vowels, specifically for the NZ accent and to a smaller opposing degree for the Lon vowels. These two accents are more familiar to the Australian listeners than are the Yrk and Ncl (UK) accents, for which exposure yielded no clear effects on goodness ratings. Thus, it appears that the goodness ratings tap into different perceptual processes than do vowel categorization judgments. Given the significant negative effect of exposure on vowel goodness ratings for the NZ accent coupled with independent evidence that Australians tend to hold negative attitudes toward this accent, in contrast with the positive exposure effect on goodness judgments about vowels in the Lon accent, which is not generally held in the same low regard, we speculate that vowel goodness ratings may be more sensitive to prior experience with and/or attitudes about particular accents than is vowel categorization accuracy. Further research could help delineate the nature of this experiential and/or attitudinal effect on vowel quality judgments.