1. Introduction
This paper is concerned with the role of individual variability in the perception of the accentual contrast between H* and L + H* in British English and what this variability tells us about the contrast itself. The phonological distinction between H* and L + H* was posited, using these terms, by Pierrehumbert (Reference Pierrehumbert1980). Although Pierrehumbert’s intonation system focused on American English, its categories, including H* and L + H*, are often taken to apply to English in general, because of intonation’s ‘high degree of uniformity […] across most varieties of English’ (Ladd, Reference Ladd, Barnes and Shattuck-Hufnagel2022: 249; for a similar stance, see also Ladd, Reference Ladd2008, ch. 3; Gussenhoven, Reference Gussenhoven2016). Below we explain the nature of the contrast and review the empirical evidence for it.
Phonetically, both H* and L + H* involve a pitch peak aligned with the accented syllable. L + H* is realized as a sharp rise starting low in the speaker’s range, while H* starts from a higher point and rises more gradually (Brugos et al., Reference Brugos, Shattuck-Hufnagel and Veilleux2006, ch. 2.5). Production studies further suggest that the accentual peak is lower in H* than L + H* (Iskarous et al., Reference Iskarous, Steffman, Cole, Skarnitzl and Volín2023; Reference Iskarous, Cole and Steffman2024), but possible alignment differences remain uncertain: a comparison of Silverman and Pierrehumbert (Reference Silverman, Pierrehumbert, Kingston and Beckman1990) and Arvaniti and Garding (Reference Arvaniti, Garding, Cole and Hualde2007) suggests earlier peak alignment for L + H*, but Steffman et al. (Reference Steffman, Shattuck-Hufnagel and Cole2022) report the opposite. The pragmatics of the accents are addressed in Pierrehumbert and Hirschberg (Reference Pierrehumbert, Hirschberg, Cohen, Morgan and Pollack1990): H* signals to the listener to add the accented item to the mutual belief space, while L + H* signals that the accented item, and not some other relevant item, should be in the mutual belief space. Thus, both H* and L + H* signal speaker commitment, though for L + H* this commitment is accompanied by contrast. In Pierrehumbert and Hirschberg (Reference Pierrehumbert, Hirschberg, Cohen, Morgan and Pollack1990), contrast is not equated with correction but construed as any proposition conveying that an alternative proposition does not hold (cf. Krifka, Reference Krifka2008; Molnár, Reference Molnár, Hasselgård, Johansson, Behrens and Fabricius-Hansen2002). For instance, in (1), reproduced from Pierrehumbert and Hirschberg (Reference Pierrehumbert, Hirschberg, Cohen, Morgan and Pollack1990: 297), the proposition in B’s turn is contrastive because it selects one of the many possible reasons why the lamp under discussion stands up: thus, in B’s turn both weighs and ton would typically bear a L + H* accent, though It weighs a ton does not correct a previous commitment about the lamp (cf. Bartels & Kingston, Reference Bartels and Kingston1996). Following Pierrehumbert and Hirschberg (Reference Pierrehumbert, Hirschberg, Cohen, Morgan and Pollack1990), in the remainder of the paper, we will use contrast to refer to both corrective and contrastive uses of accentuation.
Although the distinction between H* and L + H* is often treated as undisputed, it is not unequivocally supported by empirical evidence. Production studies indicate that both H* and L + H* can be used with contrastive and non-contrastive functions to mark both foci and topics. This applies both to controlled speech, investigated by Metusalem and Ito (Reference Metusalem and Ito2008) using a Discourse Completion Task, and the spontaneous yet formal style investigated by Hedberg and Sosa (Reference Hedberg, Sosa, Lee, Gordon and Büring2007) and Im et al. (Reference Im, Cole and Baumann2023), who examined political debates and TED talks, respectively. Perception studies support these findings, showing that L + H* is more likely to lead to a contrastive interpretation but the use of H* does not preclude it. For instance, Watson et al. (Reference Watson, Tanenhaus and Gunlogson2008) used eye-tracking to investigate H* and L + H* and concluded that L + H* creates a strong bias for contrast, but H* is compatible with both new and contrastive referents. Stronger evidence for the contrastive use of L + H* comes from studies in which it is followed by deaccenting. Thus, in Ito and Speer’s (Reference Ito and Speer2008) eye-tracking study, the presence of a L + H* on the adjective in adjective–noun pairs led to faster processing of contrastive referents, provided the L + H* was followed by a deaccented noun. Similarly, Kurumada et al. (Reference Kurumada, Brown and Tanenhaus2012) found that in utterances like It looks like a zebra, H*s on the verb and following noun are interpreted affirmatively (it looks like a zebra, and it is), while a nuclear L + H* on the verb followed by noun deaccenting triggers a contrastive interpretation by evoking a negative alternative (it looks like a zebra, but it is not). Given such results which show variability and sensitivity to context, it is not surprising that H* and L + H* were the most frequent point of disagreement among early MAE_ToBI annotators (Syrdal & McGory, Reference Syrdal and McGory2000). For similar reasons, Brugos et al. (Reference Brugos, Shattuck-Hufnagel and Veilleux2006, ch. 2.5) caution that ‘both [H* and L + H*] can be used in a variety of contexts, and a specific context will not necessarily lead all speakers to select the same intonation contour’. The recognition of this variability, however, goes against the assumption that the two accents form distinct phonological categories, since their differences in form do not result in consistent differences in pragmatic interpretation.
The picture becomes more complicated if one considers recent studies on dialectal variation which cast doubt on the assumption that the intonational system of English is largely uniform. For American English, Burdin et al. (Reference Burdin, Holliday and Reed2018) report differences in both the frequency and realization of L + H* among speakers of Jewish English, African American English and Appalachian English (see also Holliday, Reference Holliday2021a, Reference Holliday2021b). Crucially, other studies suggest that some American English varieties may lack the H* ~ L + H* contrast altogether (see Arvaniti & Garding, Reference Arvaniti, Garding, Cole and Hualde2007, on Minnesota English; Kim & Arnhold, Reference Kim and Arnhold2024, on Canadian English).
The approach to the H* versus L + H* contrast in the literature on British English is even more varied. Studies that assume a uniform intonation system across English varieties adopt Pierrehumbert’s analysis, largely concluding that H* and L + H* are not distinct (e.g., Dilley et al., Reference Dilley, Ladd and Schepman2005; Ladd, Reference Ladd2008; Ladd & Morton, Reference Ladd and Morton1997; Ladd & Schepman, Reference Ladd and Schepman2003).Footnote 1 Ladd and Morton (Reference Ladd and Morton1997) found that stepwise increases in peak height, intended to create a continuum from H* to L + H*, were perceived gradiently. Ladd and Schepman (Reference Ladd and Schepman2003) report that consecutive high accents show a consistent F0 dip aligned with the onset of the second accent and that listeners use the location of this dip to determine syllable boundaries. Thus, they conclude that the F0 dip should be part of the representation of all high accents, a proposal that implies British English does not make a distinction between H* and L + H*. Taken together, these findings support the contention of Ladd (Reference Ladd2008, inter alia) that H* and L + H* are not distinct categories.
A final complication relates to individual variability. It is well established that individuals vary in how they process linguistic information in ways that relate to cognitive characteristics, such as memory and attention (for reviews, see Kidd et al., Reference Kidd, Donnelly and Christiansen2018; Yu & Zellou, Reference Yu and Zellou2019). Such effects are likely to be considerable for intonation, where both production and perception are probabilistic (Calhoun, Reference Calhoun2010; Kurumada & Roettger, Reference Kurumada and Roettger2022) and unconstrained by lexical meaning. If so, then differences between individuals could have a sizeable impact on how intonation contrasts are produced and perceived.
In the present study, we investigated three traits that could be sources of individual variability: musicality, autistic-like traits and empathy. We chose these traits because they have been linked to the processing of phonetic and pragmatic information, both of which are of relevance in the perception of pitch accents. Studies on music and language have linked musicality to the ability to discriminate and reproduce phonetic differences: individuals with greater musical ability are better able to detect pitch differences in speech (Schön et al., Reference Schön, Magne and Besson2004, on French; Cui & Kuang, Reference Cui and Kuang2019, on English) and imitate stress in L2 (Cason et al., Reference Cason, Marmursztejn, D’Imperio and Schön2020, on French learners of English). Autistic-like traits in neurotypical adults have also been linked to the processing of phonetic cues, in that more autistic-like traits correlate with weaker integration between phonetic cues and higher-order information (Stewart & Ota, Reference Stewart and Ota2008; Yu & Zellou, Reference Yu and Zellou2019). As Bishop et al. (Reference Bishop, Kuo and Kim2020) show, this sensitivity extends to prosody, in that individuals with fewer autistic-like traits are more attuned to the prosody-meaning mapping. Finally, more empathetic individuals show higher sensitivity toward pragmatic information, most likely as a result of their greater ability to understand what other people feel or think (Baron-Cohen & Wheelwright, Reference Baron-Cohen and Wheelwright2004). The role of empathy in processing pragmatics extends to intonation, with more empathetic individuals attending more to intonation information in order to extract meaning in both L1 and L2 (on L1, see Esteve-Gibert et al., Reference Esteve-Gibert, Schafer, Hemforth, Portes, Pozniak and D’Imperio2020; Orrico & D’Imperio, Reference Orrico and D’Imperio2020; on L2, see Casillas et al., Reference Casillas, Garrido-Pozú, Parrish, Arroyo, Rodríguez, Esposito, Chang, Gómez, Constantin-Dureci, Shao, Rascón and Taveras2023).
In sum, while the phonetic difference between H* and L + H* is undisputed, scholars disagree on the validity of the H* ~ L + H* contrast, in that a clear distinction in the accents’ functions is not supported by empirical evidence, thereby casting doubt on their forming a phonological contrast. The disagreements may relate to dialectal and individual differences, dimensions of variation that have sometimes been underestimated. To address both, here we focus on the processing of H* and L + H* in Standard Southern British English (henceforth SSBE), a variety on which there have been relatively few empirical studies, and additionally consider the role of individual variation. Thus, our results contribute to the debate on the status of H* and L + H* across English varieties and shed light on the role of individual variation in the processing of intonation.
Specifically, we examined H* and L + H* by adapting the Rapid Prosody Transcription (RPT) paradigm to our purposes. In typical RPT (Cole et al., Reference Cole, Mo and Hasegawa-Johnson2010), linguistically untrained participants listen to utterances and mark on their orthographic transcripts the words they perceive as prominent; prominence is subsequently investigated through post-hoc analysis of parameters expected to affect prominence ratings (e.g., Baumann & Winter, Reference Baumann and Winter2018; Bishop, Reference Bishop2016; Cole et al., Reference Cole, Mo and Hasegawa-Johnson2010; Im et al., Reference Im, Cole and Baumann2023). Our own aim in using RPT was not to investigate possible cues to prominence but to explore the phonological status of H* and L + H*. We chose RPT because it is an indirect and not openly metalinguistic task that does not require participants to make difficult judgments about the meaning of the accents, but allows us to assess how different the accents sound to them. For this reason, our study is not concerned with all the possible parameters that could affect the relative prominence of L + H* and H* or with the prominence of these accents relative to other accents in the stimuli.
We used the prominence ratings of H* and L + H* as a way of understanding how salient the difference between the two accents is for SSBE listeners. We reasoned that if H* and L + H* are distinct categories, this would be reflected in bimodal prominence distributions, with L + H*-accented words being consistently rated more prominent than H*-accented items. In contrast, if H* and L + H* are variants of one category, the differences between them would be less salient, and this would be reflected in substantially overlapping, and potentially unimodal, prominence distributions (cf. Boomershine et al., Reference Boomershine, Hall, Hume, Johnson, Lahiri, Avery, Dresher and Rice2008, for similar findings for segmental contrasts). As relative prominence can be related to function (cf. Im et al., Reference Im, Cole and Baumann2023), we further hypothesized that if L + H*s are more prominent than H*s regardless of their function, this would support Ladd’s (Reference Ladd2008) contention that L + H*s are emphatic versions of H*s and thus that L + H* and H* form a continuum (cf. Ladd & Morton, Reference Ladd and Morton1997). On the other hand, if prominence judgments take into account accent function (i.e., L + H*s are judged as more prominent only when they are also contrastive), this would suggest that phonetics and function are interpreted together, supporting the phonological separation of the two accents (Pierrehumbert & Hirschberg, Reference Pierrehumbert, Hirschberg, Cohen, Morgan and Pollack1990).
The above is the basis for the first study, reported in Section 2. The second study (Section 3) is a replication of the first, in which we additionally examined potential sources of individual variability in prominence assessment.
2. RPT study 1
2.1. Methods
2.1.1. Participants
Sixty participants, out of 85 recruited through Prolific (https://www.prolific.co/), completed the task. We report data from 47 of these after applying exclusion criteria based on language- or performance-related issues. Seven participants self-reported not being brought up in a monolingual SSBE household; the other six met one of the following exclusion criteria which indicated that they may not have given the task due attention: in more than 10% of the stimuli, they marked as prominent (i) all the words in utterances with up to five words, or (ii) more than 85% of the words in utterances with more than five words.
The 47 participants (19–47 years old, M = 33.77, SD = 7.7) were functional monolingual SSBE speakers; i.e., they had learned other languages through formal instruction only. Thirty were female, sixteen were male, and one was non-binary. None reported any history of speech or hearing disorders. Participants took approximately 30 minutes to complete the task and were remunerated for their participation.
2.1.2. Procedure
The study ran on Roleg, an online platform developed at Radboud University (https://www.roleg.nl/TaalExperiment/). It comprised 2 practice and 86 main trials, and included 4 self-paced breaks. In each trial, participants would first hear an utterance while seeing its transcript on screen; then they heard a second repetition during which they were asked to mark prominent words by clicking on a checkbox next to each word (Figure 1). We used typical RPT instructions (e.g., Cole et al., Reference Cole, Mo and Hasegawa-Johnson2010), asking participants to mark words they heard as ‘prominent, stressed, highlighted, important or emphasized’. They could select as many words as they saw fit but had to select at least one to proceed to the next trial (see https://osf.io/f7w9c/ for the full instructions). The transcripts shown to the participants did not include punctuation or capitalization, except that apostrophes were retained in contractions and possessives, and proper nouns and the pronoun I were capitalized.
Transcript of a sample stimulus used in the RPT task.
2.1.3. Stimuli
The stimuli were 86 utterances selected from the data of 5 female and 3 male SSBE speakers, aged 18–54 years (M = 29.25, SD = 12.28), who had been recorded for a production study that included both read and unscripted speech (Kim et al., Reference Kim, Hu, Gryllia, Orrico and Arvaniti2024); none was a professional talker. All stimuli were autonomous syntactic and prosodic entities; 22 were extracted from read speech and 64 from unscripted speech (see https://osf.io/f7w9c/ for additional information about the elicitation tasks). The total number of words in the 86 utterances was 879. The utterance length range was 3–24 words (M = 10.2; SD = 4.7), or 5–34 syllables (M = 13.3; SD = 6.5). The duration range was 0.52–6.8 s (M = 2.6, SD = 1.4).
Before the study, the stimuli were annotated both for the phonetic identity and pragmatics of the accents of interest here, namely 287 accents that had high or rising F0. The two annotations were done independently to avoid each influencing the other. This is explained in more detail below.
Phonetically, high and rising accents were categorized as H* or L + H*; the annotation was based on their F0 shape without taking into account their function in the utterance. Prosodic words were annotated as carrying a L + H* accent if they showed a deliberate F0 dip at the onset of the accented syllable. The dip had to be at a relatively low level in the utterance range (i.e., not the result of high and rising F0). Prosodic words with accented syllables that started with voiceless onsets were annotated as L + H*s if it could be ascertained that the preceding syllable deliberately ended in low F0; otherwise, the accent was annotated as H*. No distinction was made between downstepped !H* and non-downstepped H* (though see 2.1.4). Following MAE_ToBI conventions (Brugos et al., Reference Brugos, Shattuck-Hufnagel and Veilleux2006), and in the absence of guidelines specific to SSBE, accents in the absolute utterance-initial position were classified as H*s. Accents immediately followed by uptalk were not considered in the analysis because it would not be possible to separate the effect of the accent from that of uptalk in the assessment of prominence. Finally, accents other than H* and L + H* were not considered in the present analysis, as the aim was not to conduct a full prominence study.
Pragmatically, all lexical items in the stimuli were categorized as contrastive or non-contrastive based solely on the fully punctuated orthographic transcript of the utterances.Footnote 2 Items were deemed contrastive if, in context, they generated a small set of explicitly mentioned or easily inferred alternatives (Pierrehumbert & Hirschberg, Reference Pierrehumbert, Hirschberg, Cohen, Morgan and Pollack1990; Krifka, Reference Krifka2008, see section 1). This classification cuts across the categories new, given, topic and focus, as contrast is orthogonal to other information structure dimensions (cf. Molnár, Reference Molnár, Hasselgård, Johansson, Behrens and Fabricius-Hansen2002). As an illustration, in (2), under is marked as contrastive since it is one of the possible ways to go around the lilies; in (3) people and bags were marked as contrastive, as they were two entities of a parallel construction and therefore contrasting with each other. In addition, focus particles, such as just and only, and negative expressions (e.g., do not in I do not know) were also marked as contrastive. All other items were marked as non-contrastive. The pragmatic categorization was then matched with the phonetic categorization to give four categories: contrastive L + H*s, non-contrastive L + H*s, contrastive H*s, and non-contrastive H*s. Figure 2 shows the waveforms and F0 tracks of examples (2) and (3).
Waveforms and F0 tracks of two stimuli produced by two female speakers illustrating H* and L + H* accents with contrastive (C) or non-contrastive (NC) function.
The phonetic and pragmatic annotations of the stimuli were initially done by the last author following the criteria mentioned above. Additionally, the second author annotated the stimuli for pragmatics, and the third author did the same for phonetics. All annotators worked independently but followed the same criteria. Unweighted Cohen’s Kappa was calculated as a measure of reliability. For the phonetic annotation, the Kappa score was calculated considering whether a word was labeled as H*, as L + H*, or not labeled at all; the agreement was very high (0.85, C.I. = 0.81–0.89). For the pragmatic annotation, the Kappa score was calculated considering whether a word was labeled as contrastive or non-contrastive (only for words annotated as H* or L + H*); the agreement was substantial (0.7, C.I. = 0.61–0.79).
Table 1 shows the distribution of the accents across the four categories. Six accents, four H*s and two L + H*s, were removed from the analysis because Roleg reported aggregate responses for words that appeared more than once in a given utterance. Consequently, it was impossible to determine which instance of the word the participant had reacted to. This resulted in the analysis of 281 accents.
Accent distribution by phonetic and pragmatic classification
2.1.4. Analysis of the stimuli
The phonetic differences between the accent categories were examined by analyzing the F0, duration and amplitude of the accented syllables, to determine whether the phonetics and pragmatics of the accents varied independently, as per our annotation, thereby allowing for the independent investigation of each parameter’s contribution to prominence. We note that, due to the relatively small sample of data, these results are presented with caution.
Accented syllable F0 was extracted in Praat (Boersma & Weenink, Reference Boersma and Weenink2023), using the Python library parselmouth (Jadoul et al., Reference Jadoul, Thompson and de Boer2018) with an octave cost of 0.1; F0 values were taken every 0.005 seconds for female and 0.01 seconds for male speakers, with customized F0 ranges, determined by the F0 minima and maxima across the stimuli of each speaker. We next ran a series of Generalized Additive Mixed Models (GAMMs; Wood, Reference Wood2011; Wood, Reference Wood2017) in R (R Core Team, 2020) to test for F0 differences between the phonetic and pragmatic classifications of the accents. We selected the model with the best fit using the function compareML() from the R package itsadug (van Rij et al., Reference van Rij, Wieling, Baayen and van Rijn2022). This model included the parametric factors for Phonetics (H*, L + H*) and Pragmatics (Contrastive, Non-contrastive), smooth terms for Time by both Phonetics and Pragmatics and random smooths for Speaker over Time by both Phonetics and Pragmatics. Syllables categorized as carrying a L + H* showed a larger rise-fall movement and a later peak than H*s; in contrast, the pragmatic categorization into contrastive and non-contrastive accents did not yield substantial differences in F0 shape (see Figure 3). These results match those of Kim et al. (Reference Kim, Hu, Gryllia, Orrico and Arvaniti2024) who followed a similar procedure with a much larger corpus.
Visualization of GAMMs results: predicted F0 values (a) and estimated difference (c) as a function of phonetics, and pragmatics (b and d, respectively). The shaded areas refer to the 95% confidence interval. The difference is significant if zero is not included in the 95% confidence interval, as marked by the red lines.
Accented syllables were annotated in Praat using standard criteria of segmentation (Machač & Skarnitzl, Reference Machač and Skarnitzl2009). Their duration was extracted using a Praat script and z-scored by utterance. Durations were then analyzed using a Linear Mixed-effect Model (LMM) with Phonetics, Pragmatics and their interaction as fixed effects, and Speaker as random intercept. The model summary showed that neither the phonetic nor the pragmatic categories were differentiated by means of duration (βphonetics: L + H* = 0.21, t = 1.32, p = 0.19; βpragmatics: contrastive = 0.10, t = 0.69, p = 0.49; βphonetics: L + H* × pragmatics: contrastive = −0.11, t = −0.48, p = 0.63).
The Root Mean Square (RMS) amplitude of the accented syllable was extracted in Praat and then normalized by dividing the obtained value by the RMS of the whole utterance. RMS was analyzed in a linear model having Phonetics, Pragmatics and their interaction as dependent variables (random intercepts yielded convergence issues). Both Phonetics and Pragmatics were significant, but not their interaction (βphonetics: L + H* = 0.34, t = 4.89, p < 0.001; βpragmatics: contrastive = 0.13, t = 2.11, p = 0.035; βphonetics: LH × pragmatics: contrastive = −0.17, t = −1.72, p = 0.086).
Finally, we considered three more stimuli properties that might affect prominence ratings: downstepping, nuclearity and deaccenting of the material following the accent (cf. inter alia Turnbull et al., Reference Turnbull, Royer, Ito and Speer2017; Im et al., Reference Im, Cole and Baumann2023). As can be seen in Table 2, the distributions of these features across categories follow trends typical for English. A portion of the accents were downstepped and this applied mostly to H*s. In addition, there were more prenuclear than nuclear accents but there was no substantial difference between proportions of nuclear and prenuclear accents across categories. Finally, contrastive L + H*s were the accents most frequently followed by deaccenting.
Counts and percentages of accents in the stimuli that were downstepped, nuclear or followed by deaccenting
In short, differences in F0 were consistent with H* and L + H* categories. In addition, L + H*s and contrastive accents had higher amplitude than H*s and non-contrastive accents, respectively, while other characteristics of the accents did not follow a consistent pattern that could have affected the outcome of our main study. Finally, as the analysis indicated that the accent form and function were independent of one another, we concluded that the phonetics and pragmatics of the accents were not conflated by annotation and thus that the role of each on prominence assessment could be independently investigated.
2.1.5. Processing of responses and statistical analysis
Following previous RPT studies (Baumann & Winter, Reference Baumann and Winter2018; Cole et al., Reference Cole, Hualde, Smith, Eager, Mahrt and de Souza2019), we ran Generalized linear Mixed-effect Models (GLMMs) using the R package lme4 (Bates et al., Reference Bates, Mächler, Bolker and Walker2015), with the RPT response (word selected or not as prominent) as dependent variable. The fixed effects included Phonetics (H*, L + H*), Pragmatics (contrastive, non-contrastive) and their interaction. The random effects included random intercepts for Speaker, Listener and Item (accented word), and the by-Listener random slopes for Phonetics and Pragmatics.
Additionally, for each test item, we calculated prominence scores (p-scores), i.e., the percentage of participants who marked that item as prominent (Cole et al., Reference Cole, Mo and Hasegawa-Johnson2010; see also Cole & Shattuck-Hufnagel, Reference Cole and Shattuck-Hufnagel2016). We use p-scores primarily for result visualization.
2.2. Results
Both the distribution of p-scores (Figure 4) and the GLMM (Table 3) showed that L + H* accented words were rated prominent significantly more often than H* accented words, while contrastive words were rated prominent significantly more often than those with non-contrastive function. There was no interaction between Phonetics and Pragmatics. Contrastive items accented with L + H* were the most prominent, while H* accented non-contrastive items were the least prominent; both non-contrastive L + H*s and contrastive H*s had overlapping distributions of p-scores (see Figure 4c).
Density and box-whisker plots of p-scores as a function of Phonetics (a), Pragmatics (b), and their interaction (c).
Summary of the GLMM output for Study 1
2.2.1. Individual variability
Individual patterns were inspected using the by-Listener random slopes from the GLMM reported in Table 3 (see Drager & Hay, Reference Drager and Hay2012). The slope value for a specific variable indicates the extent to which a participant differentiated prominence as a function of that variable: the higher the value, the more the participant relied on it. Thus, the slopes can be used as a proxy of each participant’s relative reliance on phonetics and pragmatics. Figure 5 illustrates this point by showing the relationship between the slope values for Phonetics and Pragmatics. The cut-off for the groupings was determined using as reference the participant whose difference between the Phonetics and Pragmatics slopes was closest to 0: the participants within the 20th percentile around this reference point were taken to have a relatively balanced approach toward the two cues; participants below and above the 20th percentile were those deemed to have relied mainly on Pragmatics and Phonetics, respectively.
Random slope values for Pragmatics and Phonetics within the responses of individual participants (extracted from the model in Table 3). The panels show individuals grouped according to whether slopes for Pragmatics are higher than (left), about the same as (middle), or lower than those for Phonetics (right).
2.3. Interim discussion
Our main goal in this first study was to consider whether in SSBE, H* and L + H* are sufficiently distinct to receive different prominence ratings. Such a result would point to H* and L + H* being distinct categories, rather than forming a continuum, especially if their prominence separation requires a convergence of phonetic and pragmatic factors. Our results provide evidence that this was so.
The p-score distributions for the four subcategories created by crossing phonetics and pragmatics (Figure 4c) suggest that prominence assessment did not depend exclusively on either factor but on a combination of the two. Thus, our results agree with those of Turnbull et al. (Reference Turnbull, Royer, Ito and Speer2017), Cole et al. (Reference Cole, Hualde, Smith, Eager, Mahrt and de Souza2019), Bishop et al. (Reference Bishop, Kuo and Kim2020) and Im et al. (Reference Im, Cole and Baumann2023) on American English, and Baumann and Winter (Reference Baumann and Winter2018) on German, and add comparable information about SSBE.
In brief, contrastive L + H*s and non-contrastive H*s were the most and least likely subcategories to be selected as prominent, respectively, with their scores creating a bimodal distribution. Contrastive H*s and non-contrastive L + H*s, on the other hand, spanned the entire distribution of p-scores, indicating that their prominence assessment was at chance level and likely influenced by numerous factors beyond accent identity and information structure. As noted in 2.1.4, factors such as RMS amplitude, the presence of downstep, and tonal context, including the deaccenting of following words, and the status of an accent as nuclear or prenuclear, do affect prominence ratings and were present in our stimuli (see Table 2). However, it is also reasonable to assume, based on the p-score distributions, that their effect was largely limited to the two subcategories in which phonetics and pragmatics did not match.
In conclusion, the bimodal distribution of p-scores for contrastive L + H*s and non-contrastive H*s is not compatible with a view that H* and L + H* form a continuum (cf. Ladd & Morton, Reference Ladd and Morton1997). If that were the case, we would expect a unimodal and skewed distribution of p-scores.
However, the picture is more complex than the aggregate results would suggest. The individual responses showed that listeners varied when weighing pragmatic and phonetic cues to prominence: some prioritized the former, others the latter, while a third group relied on both approximately equally. These results echo Baumann and Winter (Reference Baumann and Winter2018) who found that some participants relied more on prosody and others on morphosyntactic properties when assessing prominence. Evidence for individual variability during RPT has also been reported by Bishop et al. (Reference Bishop, Kuo and Kim2020), who connected these differences to cognitive styles. This is explored in the second study.
3. RPT study 2
Our aim in conducting the second RPT study was twofold. First, we wished to examine whether the individual variation patterns would be replicated and if so, whether the aggregate findings were replicable. The second aim was to test whether musicality, empathy and autistic-like traits could be behind individual variability, as differences in these traits could affect the listeners’ sensitivity to pragmatic and phonetic cues. Differences in sensitivity can be related to long-standing disagreements regarding the phonological status and function of H* and L + H*. To this end, we investigated the link between RPT responses and participants’ empathy, autistic-like traits and musicality (see 3.1.2): we hypothesized that more empathetic listeners would be more sensitive to and therefore more likely to rely on pragmatics, while listeners with more autistic-like traits or higher musical abilities would be more sensitive to and therefore rely more on phonetics. The study was preregistered on the OSF (Open Science Framework) platform: https://osf.io/enrtj.
3.1. Methods
3.1.1. Participants
Eighty-five participants, recruited through Prolific, took part in the study. We report results for 82 of them (47 female; 19–50 years old, M: 33.8, SD: 8.6). We excluded two participants who met the exclusion criteria mentioned in section 2.1.1, and one participant whose answers were not registered due to technical issues with the platform. Participants were remunerated for their participation.
3.1.2. Measures of individual characteristics and participant responses
The Empathy Quotient test (EQ, Baron-Cohen & Wheelwright, Reference Baron-Cohen and Wheelwright2004) was used to assess whether empathy modulates sensitivity to pragmatic information during prominence assessment. EQ is a self-report questionnaire that measures cognitive and emotional empathy. It consists of 40 statements to which participants respond using a forced-choice 4-point Likert scale (strongly disagree, slightly disagree, slightly agree, strongly agree). Each statement is scored as 0 (non-empathic responses), 1 (somewhat emphatic responses), and 2 (most empathic responses); their sum gives a score ranging from 0 to 80. As noted in section 1, EQ had been previously used by Esteve-Gibert et al. (Reference Esteve-Gibert, Schafer, Hemforth, Portes, Pozniak and D’Imperio2020) and Orrico and D’Imperio (Reference Orrico and D’Imperio2020) to investigate differences in the processing of intonational meaning.
The Autism Quotient test (AQ, Baron-Cohen et al., Reference Baron-Cohen, Wheelwright, Skinner, Martin and Clubley2001) was used to assess whether autistic-like traits modulate sensitivity to phonetic detail during prominence assessment. AQ is not a diagnostic test for autism; rather, it positions neurotypical adults along a continuum measuring five traits associated with the Autism Spectrum Disorder: social skills, communicative skills, attention to detail, attention switching, and imagination. These five traits are measured by different AQ subscales: the questionnaire comprises 10 statements for each subscale (50 statements in total) to which participants respond using a forced-choice 4-point Likert scale (strongly disagree, slightly disagree, slightly agree, strongly agree); the score is calculated by assigning 0 (non-autistic-like response) or 1 (autistic-like response); the total score is the sum of all subscale scores and ranges from 0 to 50. The AQ has been used to examine individual variation connected to both phonetics (e.g., Stewart & Ota, Reference Stewart and Ota2008; Yu, Reference Yu2010; Yu et al., Reference Yu, Abrego-Collier and Sonderegger2013) and pragmatics (e.g., Bishop, Reference Bishop2016; Yang et al., Reference Yang, Minai and Fiorentino2018). Some studies have relied on AQ subscales; e.g., Yu et al. (Reference Yu, Abrego-Collier and Sonderegger2013) analyzed each subscale separately, while Bishop (Reference Bishop2016) and Bishop et al. (Reference Bishop, Kuo and Kim2020) used only AQ-Communication. Others have used the aggregate score for their main hypotheses, reporting information about the subscales, to variable extent, as a post-hoc analysis (e.g., Stewart & Ota, Reference Stewart and Ota2008; Yang et al., Reference Yang, Minai and Fiorentino2018; Yu, Reference Yu2010). Since no one subscale has been consistently linked to phonetics, we followed the latter approach and formulated our hypothesis considering the aggregate score. Finally, we note that AQ and EQ are negatively correlated with one another (Baron-Cohen & Wheelwright, Reference Baron-Cohen and Wheelwright2004). However, since here we were interested in the effect of AQ on sensitivity to phonetics and of EQ on sensitivity to pragmatics, the two tests do not tap into the same aspect of our study.
The Mini Profile of Music Perception Skills (Mini-PROMS, Law & Zentner, Reference Law and Zentner2012; Zentner & Strauss, Reference Zentner and Strauss2017) was used to assess whether musicality modulates sensitivity to phonetic detail during prominence assessment. The Mini-PROMS was chosen because it tests musical ability rather than musical training or love for music. It consists of four components: Melody (10 items), Metric Accent (10 items), Tempo (8 items), and Tuning (8 items). Participants listen to pairs of musical fragments and indicate whether they are the same or different using a 5-point Likert scale (definitely same, probably same, I do not know, probably different, definitely different). Correct answers receive 2 points (definitely same/different) or 1 point (probably same/different); incorrect and I do not know answers are awarded zero points. The score for each component is calculated as the sum of the points divided by 2; the total score ranges from 0 to 36 and is the sum of the component scores. Studies using Mini-PROMS as a predictor of the processing of prosody-related information include Foncubierta et al. (Reference Foncubierta, Machancoses, Buyse and Fonseca-Mora2020).
Participant responses to each of the above tests were normally distributed (see Figure 6) and covered most of the range of each test, from 16 to 70 for EQ (M = 43.15, SD = 12.33), from 1 to 42 for AQ (M = 19.18, SD = 9.91), and from 10.5 to 28 for MiniPROMS (M = 19.66, SD = 3.97). Cronbach’s alpha showed high reliability for all three tests: EQ = 0.95 (C.I. = 0.93–0.97); AQ = 0.93 (C.I. = 0.90–0.95); MiniPROMS = 0.97 (C.I. = 0.95–0.98). Pearson correlations between pairs of tests showed a negative correlation between EQ and AQ (r(80) = −.58, CI = −0.71, −0.41, p < .001), as expected, and no correlation between MiniPROMS and either EQ or AQ.
Score distributions for EQ (a), AQ (b) and MiniPROMS (c).
3.1.3. Stimuli and procedure
We employed the design and stimuli used in the first study; the platform issue concerning stimuli involving the same word had been resolved and thus all 287 accents were analyzed. Within a week of completing the RPT task, the participants also completed the AQ, EQ and MiniPROMS in random order chosen by themselves.
3.1.4. Recruitment and power analysis
We used a stop-go method to determine the sample needed to reach statistical power of 80% or higher (see https://osf.io/enrtj). Briefly, we paused the study after the first 25 participants and used their data to run simulations (for details see https://osf.io/f7w9c/). Following Vasishth and Gelman (Reference Vasishth and Gelman2021), we calculated power by simulating datasets with increasing number of participants, going from 20 to 120.
Figure 7 shows the results of the simulations for the interactions between EQ and Pragmatics, AQ and Phonetics and MiniPROMS and Phonetics. They indicate that with the present sample size of 82 participants, we reach more than 90% power for the interaction between EQ and Pragmatics, and MiniPROMS and Phonetics. Somewhat lower power was predicted for the interaction between AQ and Phonetics (approximately 70%). We decided to stop the study without reaching the 80% threshold for AQ, as the power analysis indicated a smaller effect size than MiniPROMS and EQ and thus a lesser role of AQ overall.
Power analysis output for the interaction AQ × Phonetics (a), EQ × Pragmatics (b) and MiniPROMS × Phonetics (c).
3.1.5. Statistical analysis
The RPT responses were analyzed in the same way as in the first study (see 2.1.5). In addition, we fitted three more GLMMs to test the effects of EQ, AQ, and MiniPROMS. For EQ, the model included Phonetics, Pragmatics, EQ score, and the interaction between EQ score and Pragmatics as fixed factors. For AQ, the model included Phonetics, Pragmatics, AQ score, and the interaction between AQ score and Phonetics as fixed factors. For MiniPROMS, the model included Phonetics, Pragmatics, MiniPROMS, and the interaction between MiniPROMS and Phonetics as fixed factors. All three models had Item, Subject and Speaker as random intercepts.
3.2. Results
3.2.1. Prominence ratings
As shown in Table 4, the output of the first GLMM replicated the results of the first study (see also Figure 8). The same applied to individual differences (see Figure 9).
Summary of the GLMM output for Study 2
Density and box-whisker plots of p-scores as a function of Phonetics (a), Pragmatics (b), and their interaction (c).
Random slope values for Pragmatics and Phonetics within the responses of individual participants (extracted from the model in Table 4). The panels show individuals grouped according to whether slopes for Pragmatics were higher than (left), similar to (middle), or lower than those for Phonetics (right).
3.2.2. The role of individual characteristics on prominence ratings
The interaction between EQ and Pragmatics was significant (see Table 5 for the model summary). As illustrated in Figure 10, contrastive accents were more likely to be selected as prominent by individuals with higher EQ than those with lower scores, while the lower prominence of non-contrastive accents did not change as a function of EQ.
Summary of the GLMM testing the effect of EQ on RPT responses
Probability of prominence selection as a function of EQ and Pragmatics. Shaded areas around the regression lines refer to 95% confidence intervals.
The interaction between AQ and Phonetics was also significant (see Table 6 for the model summary). As illustrated in Figure 11, H*s were more likely to be selected as prominent by participants with higher AQ than those with lower AQ, while the higher prominence of L + H* accents did not change as a function of AQ.
Summary of the GLMM testing the effect of AQ on RPT responses
Probability of prominence selection as a function of AQ and Phonetics. Shaded areas around the regression lines refer to 95% confidence intervals.
Finally, there was a significant interaction between MiniPROMS and Phonetics (see Table 7 for the model summary). As also shown in Figure 12, L + H*s were more likely to be selected by participants with higher MiniPROMS scores, while H*s were less likely to be selected as prominent, regardless of MiniPROMS score.
Summary of the GLMM testing the effect of MiniPROMS scores on RPT responses
Probability of prominence selection as a function of MiniPROMS scores and Phonetics. Shaded areas around the regression lines refer to 95% confidence intervals.
3.2.3. Interim discussion
The second study replicated both the aggregate results and the individual variation patterns of the first. In addition, it shed light on the sources of these individual differences.
The interaction between EQ and Pragmatics showed that more empathetic listeners were more likely to mark contrastive accents as prominent. This trend indicates that these individuals were more sensitive to pragmatic differences, as we had hypothesized. This result agrees with earlier studies on the role of empathy in the processing of pragmatics, whether empathy was directly measured using the EQ (Esteve-Gibert et al., Reference Esteve-Gibert, Schafer, Hemforth, Portes, Pozniak and D’Imperio2020; Orrico & D’Imperio, Reference Orrico and D’Imperio2020), or inferred from either the AQ total score, as in Yang et al. (Reference Yang, Minai and Fiorentino2018), or the AQ Communication subscale, as in Bishop (Reference Bishop2016).
Further, our results indicate that both musicality and autistic-like traits reflect sensitivity to phonetic information though, in different ways and to a different extent. The interaction of MiniPROMS scores and Phonetics strongly confirmed our prediction: participants scoring high on musicality were more likely to mark as prominent words accented with L+H* relative to those with low scores, indicating greater sensitivity to the differences between H* and L + H*. Our results agree with those of previous studies that musical abilities play an important role in the processing of phonetic information (Cui & Kuang, Reference Cui and Kuang2019; Schön et al., Reference Schön, Magne and Besson2004).
Finally, the interaction of AQ scores and Phonetics supported our hypothesis, in that it showed a link between AQ and sensitivity to phonetic detail, though not in the direction we had anticipated. Participants with higher AQ scores were more likely to mark H*-accented words as prominent relative to those with lower scores, suggesting that high AQ individuals were sensitive to small phonetic changes that may not be particularly salient to those with lower AQ. This finding supports previous research showing that the differences in the perception of phonetic information as a function of autistic-like traits may not depend on higher auditory sensitivity, but on a different way of processing higher-order information (Yu & Zellou, Reference Yu and Zellou2019).
In brief, our findings confirmed the existence of individual variability in RPT responses detected in the first study and showed that it is related to differences in cognitive styles. These differences mean that participants are more sensitive to either pragmatic or phonetic information. We contend that this sensitivity leads listeners to prioritize different aspects of information in the signal when assessing prominence.
4. General discussion
This paper addressed the long-standing debate concerning the phonological status of H* and L + H* accents in English; we investigated their relative salience, as reflected in p-scores, as a means of understanding their relationship in SSBE. We reasoned that if L + H* is an emphatic variant of H* (cf. Ladd & Morton, Reference Ladd and Morton1997) it should be rated more prominent than H* regardless of pragmatic function, while the p-scores of H*- and L + H*-accented items should form a unimodal distribution.
Our results, however, showed that both phonetics and pragmatics affected the responses. By and large, this result is compatible with other recent findings that showed phonetic properties, phonological status and information structure affect how prominence is assessed (c.f. Im et al., Reference Im, Cole and Baumann2023; Turnbull et al., Reference Turnbull, Royer, Ito and Speer2017). As a result of these factors, the p-scores of our four categories spanned the entire p-score range, indicating that some accented words were not selected by anyone and others were selected by all participants – an outcome that highlights the multiple influences on RPT responses.
Critically, the responses did not form a continuum interpretable as the result of additive contributions of phonetics and pragmatics toward increased prominence. Rather, the distribution of the p-scores was bimodal, indicating that H* and L + H* were processed as separate entities. In turn, this suggests that H* and L + H* were phonologically interpreted: listeners’ expectations were that form and function would be matching, leading to clear differences in p- scores, while mismatches created uncertainty. In this respect, the results support the view that the accents are not only phonetically distinct (cf. Kim et al., Reference Kim, Hu, Gryllia, Orrico and Arvaniti2024), but also that each phonetic realization is preferentially connected to a distinct information-structure related function.
However, this conclusion is modulated by variation in individual responses, which showed that some participants prioritized pragmatics over phonetics, others did the opposite, while a few participants weighed both cues (almost) equally. The variable response patterns indicate that the differences between H* and L + H* accents were not equally salient to all participants. Our second study suggested that this was due to individual differences associated with empathy, which modulated how salient pragmatic differences were, and musicality and autistic-like traits, both affecting salience based on accent phonetics.
We contend that these differences among individuals result in their being more or less sensitive to the overall distinction between H* and L + H*. For some individuals, one dimension is less salient than the other, while for others both dimensions carry similarly low or similarly high weight. Consequently, individuals may reach different generalizations regarding this accentual contrast: some acquire it because both the phonetic and pragmatic differences are salient to them; others do not because they fail to integrate the phonetic and pragmatic information into distinct accentual categories. Further, the replication of both the aggregate and individual results indicates a stable variation in the population, a status quo that may also lie behind the long-standing disagreements among linguists about the status of H* and L + H*.
In sum, while the aggregate results favor a contrast between H* and L + H*, the variable interpretation that speakers have of these accents suggests that some SSBE speakers may not acquire the distinction between H* and L + H* as a contrast (cf. Arvaniti & Garding, Reference Arvaniti, Garding, Cole and Hualde2007, on the lack of H* ~ L + H* in Minnesotan English and Kim & Arnhold, Reference Kim and Arnhold2024, for similar conclusions regarding Canadian English). In segmental phonology, similar cases are not unusual and are referred to as marginal contrasts (cf. Ladd, Reference Ladd, Goldstein, Whalen and Best2006; Scobbie, Reference Scobbie, Ramchand and Reiss2007). For example, while the Italian mid-vowels [e] ~ [ɛ] and [o] ~ [ɔ] are involved in minimal pairs in most Italian dialects, native speaker awareness of the two contrasts – i.e., of the mapping between each phonetic category and the lexicon – is low, suggesting they are likely to be ignored during lexical identification (Renwick & Ladd, Reference Renwick and Ladd2016). Hualde et al. (Reference Hualde, Luchkina and Eager2017) further argue that individual variability is an indicator of contrast marginality: those authors examined Canadian Raising and found that Chicago speakers distinguish minimal pairs such as writer [rʌɪɾɚ] ~ rider [raɪɾɚ] in their productions, but show individual variation in perceiving them. The characteristics of marginal segmental contrasts amply relate to the findings reported here about H* ~ L + H*. The link between phonetics and function in differentiating the accents is what characterizes them as contrastive. However, not all speakers are equally sensitive to the differences, therefore the contrast shows signs of marginality, at least in SSBE.
5. Conclusion
We tested the hypothesis that H* and L + H* form a continuum in SSBE and assessed this hypothesis by considering the extent to which the accents’ prominence, as reflected in RPT responses, differs based on their phonetic differences and pragmatic functions. Our results clearly showed that L + H* contrastive and H* non-contrastive accents have distinct p-score distributions, suggesting that the differences between the two accents are salient to SSBE listeners, so long as phonetics and pragmatics match. In turn, the presence of this distinction, coupled with the indistinguishable p-scores of H* contrastive and L + H* non-contrastive accents, supports the existence of prototypical uses of H* and L + H*. This is indicative of contrast. However, the presence of individual differences among our listeners suggests that the contrast is marginal and its presence in a given speaker’s grammar depends on individual cognitive traits, here empathy, musicality and autistic-like traits: combinations of these traits allow some individuals to integrate phonetic and pragmatic information to form categories, while others do not. We argue that the presence of these individual differences in how salient the distinction between H* and L + H* appears to individuals may be the key to understanding this long-debated issue: seeing the H* ~ L + H* contrast as marginal can explain the disagreements in the literature regarding these accents as stemming from the speakers’ variable use and understanding of the accents.
Data availability statement
The data and scripts are publicly available and can be found at https://osf.io/f7w9c/.
Acknowledgments
We thank Na Hu, Katherine Marcoux, Sofia Sialiaki and Cong Zhang for help with various aspects of the studies reported here, and Chris Cummins and Hannah Rodhe for their input and help with devising the pragmatic annotation scheme. Finally, we thank the editors and two anonymous reviewers for their helpful comments and suggestions.
Funding statement
This research is part of a project that has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (Grant agreement No. ERC-ADG-835263 to Amalia Arvaniti).
Competing interest
The author(s) declare none.
Open access
