1.1. Intonational meaning and the Autosegmental Metrical Theory

At the heart of the Autosegmental Metrical (AM) model of intonation (Pierrehumbert, 1980; see Ladd, 2008 for a review), there is the assumption that an intonation contour is a sequence of discrete tonal elements, i.e., tone targets, combining into either monotonal or bitonal pitch accents and edge tones (phrase accents and boundary tones). These elements originate from two discrete primitive level tones, high (H) and low tones (L). Studies conducted in the last 40 years within this framework have allowed the linguistic community to reach a good understanding of the phonology-phonetic interface at the prosodic level. Nevertheless, several controversies still remain unresolved, especially as far as the specific role of intonation in the expression of pragmatic meaning, what is referred to as ‘intonational meaning’ by Ladd (2008).

Probably the best known intonational meaning model is the one proposed by Pierrehumbert and Hirschberg (1990), in which the authors propose a strong tune compositionality. In this model, each single tone recoverable from the f0 contour would carry an inherent, context-independent meaning. In line with the tenets of AM, the model assigns specific meanings to the two primitive tones, H and L, and their contribution in pitch accent and edge tone combinations. The pragmatic meaning of an utterance is then obtained by adding the meanings of each tone composing the tune. Additionally, the authors posit a hierarchical phonological model in which the three structural elements composing a tune—pitch accents, phrase accents, and boundary tones—operate over a progressively higher domain dictated by the node they are attached to. Therefore, pitch accents would have scope over the lexical item they are attached to, phrase accents over the intermediate phrase, and boundary tones over the entire intonation phrase.

The strong compositionality of existing AM-based models shows nevertheless some serious drawbacks. Positing that each single tone possesses an inherent meaning makes it difficult to account for collocations of nuclear pitch accents and boundary tones (see Dainora, 2001). That is why several recent models of intonational meaning, couched within AM, have put forth a weak compositional approach, as Steedman’s (2007) model for English intonation and Portes and Beyssade’s (2015) account for French. We believe that both compositionality and the notion of ‘inherent’ meaning of an intonational contour need to be revised. On one side, intonational meaning can be viewed as gradual and not simply as discrete. Similar to the phonetics-phonology mapping implied in exemplar theory (Goldinger, 1998; Johnson 1997; 2006; Pierrehumbert 2001; 2002), pragmatic meaning can be thought of as being gradually implemented both in production and perception. This view would differ from a paradigm such as the one invoked by the Biological Codes theory proposed by Gussenhoven (2002), in which gradual acoustic cues would merely pertain to universal meanings, while linguistic intonation representations would only be discrete in nature (see Section 1.3 below).

1.2. The epistemic nature of intonational meaning

Very recent models argue that intonation is concerned with meaning related to the dialogical status of an utterance, such as its role in adding information to the Common Ground, its relation to private beliefs and the degree of commitment to a proposition, be it stated or evoked in the utterance (see Prieto, 2015, for a detailed account). A series of intonational meaning studies has hence connected specific intonational features to the role of carrying information relative to epistemic commitment on the part of the interlocutors. In some of these studies, the speaker’s use of intonation is assumed to carry the function of expressing that she is taking responsibility for the truth of a proposition and/or she is attributing such responsibility to the addressee (Beyssade & Marandin, 2006; Krifka, 2017). Among these studies, Bartels (1999) proposes that the use of the L- phrase accent in American English signals commitment on the part of the speaker relative to the truth of a relevant proposition in the discourse and, at the same time, to an instruction aimed at the addressee to publicly committing to it. A different perspective, though within a similar framework, has been taken by Steedman (2007) for English and, more recently, by Portes and Beyssade (2015) for French. Both of these models posit, for instance, that the role of boundary tones in declarative sentences is to variably attribute commitment to either the speaker (when falling) or the addressee (when rising).

Additionally, some authors also propose that pitch accents are responsible for determining whether or not the information contained in the utterance is predicted by the speaker to be either consensual or controversial. In this view, intonation is taken to convey an important role in dynamically changing and managing the Common Ground, and this by adding, accepting, or rejecting a proposition (Krifka, 2017). Some work has also uncovered a complex interaction between syntactic markers and illocutionary force (i.e., declarative versus interrogative modality) due to the presence of a specific intonational tune. For instance, in a study of declarative questions in American English, Gunlogson (2004) claims that rising nuclear contours in declaratives may indicate lack of commitment to the proposition on the part of the speaker, while rises in interrogative sentences might attribute commitment to the addressee. In this view, the meaning of declarative questions would be derived compositionally and in light of the direction of the terminal part of the contour. Specifically, a terminally falling declarative (statement tune) would signal speakers’ commitment to the proposition while a rising contour (yes-no question tune) would signal lack of the same (hence commitment being attributed to the addressee).

A great deal of research operating within the intonation-meaning framework mentioned above has also investigated the role of intonation in expressing epistemic bias in questions. Among various kinds of pragmatic bias, epistemic bias (Sudo, 2013) pertains to presupposed information about the content of the answer to the question on the part of the speaker. Within different speech acts, biased questions belong to the set of non-canonical questions, in that their function is not just a way to elicit information, but also to convey an epistemic bias (i.e., an expectation) about what that information might be (Dayal, 2016). While, according to standard semantics, a canonical yes-no question creates a balanced partition between the two polarities of a proposition, the role of epistemic bias is that of creating an unbalanced partition, with the speaker leaning towards either p (positive bias) or ¬p (negative bias). Several models have accounted for the role of bias as attributing question properties that are typical of assertions. Asher and Reese (2007), for example, analyze biased questions as complex speech acts, formed by both an assertion and a question part. A typical example of biased question in their analysis is represented by tag questions, which overtly include both a declarative and an interrogative component. Apart from tag questions (also treated by Ladd, 1981), bias appears to be an intrinsic property of negative polar questions (Ladd, 1981; Büring & Gunlogson, 2000; Asher & Reese, 2005; 2007; see also Goodhue & Wagner, 2018) and declarative questions.

From this short review, note that the role that intonation plays in determining the presence and the polarity of the bias (positive or negative) has emerged to the frontline. Analyses like the one proposed by Asher and Reese (2005; 2007) argue that the presence of an epistemic bias can be signaled by the use of intonational cues alone. Similar theoretical claims have been tested in recent experimental studies. Among them, Michelas, Portes, and Champagne-Lavau (2016) designed a production study aimed at investigating the role of intonation in encoding speaker commitment in French. Speakers were asked to produce questions in two pragmatic conditions, i.e., unbiased and negatively biased questions. Findings show that the role of pitch accents is crucial in differentiating across the two conditions, with H* nuclear accents signaling unbiased questions and H+!H* accents signaling negative bias, showing the contribution of intonation in determining the beliefs state of speakers. Portes, Beyssade, Michelas, Marandin, and Champagne-Lavau (2014) also investigated the dialogical function of intonation in French by specifically addressing the issue of how intonational contours express speaker commitment and ‘call on addressee,’ which is a way to anticipate the listener’s reaction to a speaker’s utterance. In particular, the authors found that intonational contours elicit specific reactions from the addressee, showing the role of intonation in determining whether information is agreed or not among conversational participants, and hence added to the Common Ground.

Further support for the role of intonation in conveying information regarding commitment and agreement between interlocutors has been provided by empirical work on other Romance varieties, such as Puerto Rican Spanish (Armstrong, 2012) and Catalan (Prieto & Borràs-Comes, 2018). In these studies, the authors show that question intonation can encode fine-grained distinctions related to the speaker’s own commitment to the proposition expressed through some form of epistemic stance about the conveyed content (Heritage, 2012). Prieto and Borràs-Comes (2018) designed a perception task in which native speakers of Catalan were asked to rate the acceptability of tune-context pairs. Four yes-no question contours were analyzed, a L* H% rise, a H+L* L% fall, a L+¡H* L% rise-fall, and a L+H* LH% rise-fall-rise, which were paired with six different contexts, resulting in different degrees of speaker commitment and agreement. Results show that intonation does indeed function as a device expressing commitment and agreement. Specifically, they show that the falling H+L* L% contour was rated as highly acceptable in high commitment contexts, while the L* H% rise correlates with low levels of agreement.

Intonational studies on both Catalan and Bari Italian have also pointed out the role of pitch direction (rises versus falls) in signaling different types of questions, such as the difference between confirmation and information-seeking questions. Grice and Savino (1997), for instance, showed that Bari Italian yes-no questions (questions about new information) are marked by the presence of a nuclear rising L+H* accent while confirmation-seeking questions (in which the information addressed is given) are generally marked by the presence of a falling nuclear H+L* accent. Hence pitch accent category and/or edge tone combinations appear to have a major role in signaling bias in questions.

1.3. Gradient intonational cues to meaning

One important issue concerning the nature of intonational meaning is the divide that is often argued to exist between gradient versus categorical elements of intonation. Within traditional AM approaches (Liberman & Pierrehumbert, 1984), pitch range modulations within the two phonological, L and H, primitive tones are argued to be used only to express paralinguistic meanings (such as emphasis, degree of speaker involvement, etc.). As a consequence, according to this view, a phonological tune can be expressed with very different pitch ranges without changing its pragmatic function. Moreover, pitch range can vary along two dimensions (see Ladd, 2008 for a discussion), i.e., span, which is the excursion between f0 peaks and valleys, and level, which relates to the register (or key) within which target f0 values are realized. Variations within these two components of intonation have been linked to the expression of paralinguistic meanings such as arousal, involvement of the speaker, and social dominance (see Ladd, 2014). Nevertheless, the fact that gradient cues associate with paralinguistic meanings and intonational categories with linguistic meaning is quite controversial. As pointed out by both Grice and Bauman (2007) and Prieto (2015), the separation between the two dimensions of cues and types of meanings is far from clear-cut.

A theory which attempts to address both the paralinguistic meaning and the grammatical meaning of intonation is the Biological Codes theory, proposed by Gussenhoven (2002). The theory is built on the assumption that the language-specificity nature of intonational meaning is the result of a process of grammaticalization of universally-valid phonetic-meaning mappings. Building on Ohala’s (1983) ethological view of fundamental frequency use (Frequency Code), Gussenhoven extends this view by postulating four Biological Codes from which two types of meanings are generated, one that is affective—signaling attributes of the speaker—and an informational, linguistic, meaning, coding information related to the message. In line with Ohala, the Frequency Code is intended as directly relating pitch height to the size of a speaker’s vocal folds, giving rise to meanings such as submissiveness and politeness (high pitch), versus impoliteness and confidence (low pitch). The informational interpretation of the Frequency Code relates to the opposition between finality and non-finality and certainty versus uncertainty, with the second attribute in both cases being associated with higher pitch. By the same token, assertiveness versus questioning would hence be related to f0 rises or increased peak height (see for instance Chen, Gussenhoven, & Rietveld, 2004). Note though that the Frequency Code predictions for politeness, being related to rising or high pitch, have recently been challenged by empirical results in a variety of languages (see Chikulaeva & D’Imperio, 2018 for Russian; Winter & Grawunder, 2012; and Brown, Winter, Idemaru, & Grawunder, 2014 for Korean).

The link between pitch span and linguistic, pragmatic meaning has already been highlighted by several studies in the past. As mentioned above, Hirschberg and Ward (1992) were the first to point out that that information relative to pitch span is actually instrumental for determining the difference between an incredulous or an uncertain reading of the rise-fall-rise contour in American English. Specifically, having ruled out the hypothesis of polysemy of this contour, they proposed that the two readings might represent a subset of the wider meaning labeled as ‘lack of speaker commitment.’ Moreover, recent evidence from Romance languages points to the fact that pitch span can indeed express linguistically structured meanings. Borràs-Comes et al. (2014), for instance, show that pitch span variation within the rising L+H* pitch accent can encode three different meanings in Catalan, i.e., neutral statement (lower peak), narrow-contrastive focus (medial peak), and counter-expectational yes-no question (higher peak).

Similarly, evidence for several varieties of Italian (Gili Fivela et al., 2015) points to the role of wider pitch span in the nuclear configuration region of wh-questions being used to encode counter-expectational meaning. The same type of phonetic cue appears to be employed in the Italian variety spoken in Este, Padua (a Northern variety) to signal incredulous echo wh-questions (Crocco & Badan, 2016). In Salerno Italian, which is the variety employed in the present study, this specific meaning appears to be brought about by the combination of a nuclear rising L*+H pitch accent with expanded pitch excursion (Orrico & D’Imperio, 2020). Interestingly, the meanings which are cross-linguistically attributed to expanded pitch span—i.e., incredulity, counter-expectation, negative bias—all imply some sort of rejection of the propositional content conveyed by the utterance.

Additionally, and more importantly, in all these cases it might be possible to talk about gradience in Bolinger’s (1989) sense, which implies a mapping between gradient phonetic cues and gradient meanings. Indeed, apart from paralinguistic meanings, also linguistic ones might be conceived in a gradient way. Heritage (2012), for example, argues that the same propositional content can be modulated by a language grammar on the basis of the relative knowledge possessed by the speaker and the addressee. This would lead to a continuum between unbiased yes-no questions, representing that the speaker is not committing to any proposition, and statements, representing an increased knowledge and commitment on the part of the speaker. In this view, biased questions would occupy an intermediate position within the continuum.

1.4. Individual variability in the tune-meaning mapping

It is well established that speech encodes indexical information about the speaker producing it, and this appears to be true also for intonation. In a study by Clopper and Smiljanic (2011), effects of both gender and regional dialect (Southern versus Midland American English) on intonation production were found. Specifically, the authors reported that female speakers of both dialects produce more L*+H accents (while males show a preference for H* in the same conditions). Crucially, though, they also found an inter-dialect difference in the frequency use of L- phrase accents for female speakers only, showing that the way different social variables interact with each other is a relevant aspect. An effect of gender has also been found in the production of high-rise terminals in London English by Levon (2016), a dialect in which men show a tendency to use high-rises in narratives to draw attention to interesting or new elements in the discourse. On the other hand, women appear to use this contour to maintain epistemic authority over the narrative.

Studies on intonational variability have unveiled that group- and speaker-specific strategies are also employed at the level of intonational encoding of pragmatic meanings. What is more, some of these studies have also highlighted the role of variability at the listener level. For example, Arvaniti, Baltazani, and Gryllia (2014) found that the pragmatic meaning of wh-question tunes in Greek is interpreted differently not only as a result of speaker gender, but also depending on the gender of the listener, with an increased tendency towards interpreting a wh-question as a statement with negative implicatures if the speaker is male and the listener is female. Additionally, work by Cangemi, Krüger, and Grice (2015) shows that different speakers use different prosodic strategies to encode three different types of focus in German (broad, narrow, and corrective) and, additionally, different listeners also use different strategies to encode them (though they do not elaborate on possible sources of such speaker- and listener-specific differences).

These studies point to the fact that both listener and speaker idiosyncrasies can affect the complex form-meaning mapping of intonation, contrary to what is traditionally maintained (e.g., Pierrehumbert & Hirschberg, 1990). Apart from indexical information, the body of studies dedicated to individual variability has also shown that two other factors appear to affect the linguistic behavior of speaker and listener, i.e., social variability and individual cognitive traits (Kidd, Donnelly, & Christiansen, 2018; Yu & Zellou, 2019). Exposure to different dialects, in terms of type and amount, is another important individual variable. Chen, Rattanasone, Cox, and Demuth’s (2017) eye-tracking study, for example, shows that Australian English listeners with and without early exposure to other English dialects show different behavior in reaction to mispronounced vowels. Specifically, listeners with exposure to other varieties appear to be more tolerant to deviant exemplars, which leads the authors to conclude that phonological categories in these individuals might be less rigid than those for individuals without exposure. More recently, Levy and Hanulíková (2019) report that variability in vowel production on the part of children is related to their levels of non-native dialectal exposure.

Only scarce evidence of the effect of exposure to other phonological systems can be found in intonation studies. In a study on tonal alignment acquisition in L2, Mennen (2004) reported that adult Dutch learners of Greek shift their tonal alignment patterns in prenuclear rising pitch accent towards that of native speakers of Greek, though not enough to obtain Greek-like values. Interestingly, the same learners would present alignment values in their L1 that would not be native-like either, pointing to a bi-directional effect of intonational convergence. Note that Gili Fivela et al. (2015) put forward the hypothesis that contact among regional dialects of Italian might be an important source for the amount of variability in intonational patterns. In a study on tonal alignment imitation, D’Imperio, Cavone, and Petrone (2014) report that short-term exposure to a different alignment in a question rise-fall contour produced by a speaker of a different variety of Italian can be immediately stored and reproduced.

Very recently, in a study on the perception of Corsican French intonation, Portes and German (2019) found that exposure to Continental French modulates the response to a regional pattern cueing either a question or a statement. This echoes findings by Walker, Szakay, and Cox (2019) in which socio-indexical segmental cues to Australian versus New Zealand English appear to be modulated by the amount of exposure to each variety. In fact, the authors find that the expected regional priming effect is actually weaker in their study than the one found in previous studies (cf. Hay & Drager, 2010). In a similar vein, Warren’s (2017) findings on sociophonetic priming in the perception of uptalk in New Zealand English point to a mutual interaction of dialectal exposure and different socio-indexical sources of priming.

From this short review, it emerges that individual variability in intonation research has mainly concentrated on socio-indexical and regional features. Only very recently individual cognitive skills have been argued to play a role in tone and intonation perception variability, both in L1 and L2. Among those skills, musicianship has been found to have a major role in terms of L2 acquisition in both children and adults. In a series of ERP and perception studies, general musical training (musicianship) has been linked to faster novel word learning by French participants, both in terms of form and meaning (Dittinger et al., 2016; Dittinger, Valizadeh, Jäncke, Besson, & Elmer, 2018). Among musical skills, a recent study has shown that rhythmical aptitude predicts performance in correct stress placement in French learners of L2 English (Cason, Marmursztejn, D’Imperio, & Schön, 2019). In addition to musicianship, other individual cognitive traits appear to affect prosodic cue processing. Jun and Bishop (2015), for instance, showed that individuals with a high level of autistic traits (measured through the Autistic Quotient) appear to be more sensitive to prosodic boundaries in cueing high or low attachment of a relative clause, showing that such cognitive factors might have an effect on syntactic parsing. In addition, these results are in line with findings relating higher degree of pragmatic skills and better discrimination and exploitation of pitch accent contrast in meaning processing (Colon & Bishop, 2015; Bishop, 2016). Along these lines, Estève-Gibert et al. (2018; 2020), in an Eye Tracking experiment on contrastive meaning processing in French, found that individuals who scored lower at the Empathy Quotient test (Baron-Cohen & Wheelwright, 2004) rely primarily on lexical disambiguating cues instead of early intonational ones for the purposes of homophone disambiguation.

As we know, the covariation of phonetic cues and socio-indexical information is a hallmark of exemplar models of speech perception (Pierrehumbert, 2001; German, Carlson, & Pierrehumbert, 2013). In this study we will extend our view to cognitive and dialectal exposure effects to predict the influence of gradual pitch span variability on epistemic bias identification. But before we detail the experimental design of this study, we will briefly review the specificities of Salerno Italian intonational phonology.

5.1. Exposure and Empathy Quotient interaction

Against our predictions for H2, our retained model did not show significant main effects of either Exposure or EQ (Exposure: F(1, 41.51): 1.21, p = .28 and EQ: F(1, 41.51): 2.59, p = .12). Nevertheless, and crucially, both individual factors interacted significantly with either PA or BT span step, showing that they do have a role in modulating the mapping of phonetic cues onto perceived bias.

A first clear effect was registered in how EQ scores interacted with BT steps. Note from Figure 7 that while the High EQ group differentiated across BT steps (red line in the plot), the Low EQ group failed to do so. The interaction between EQ score and Boundary tone was in fact significant (F(5, 59.13): 3.54, p = .007). Specifically, recall that Figure 5 above showed that boundary levels were a significant predictor of Certainty score ratings, with higher levels of BT being rated as conveying higher degrees of perceived Certainty (for a positive answer). Interestingly, this effect appears to be driven by high-empathy listeners, as revealed by the data in Figure 7 above. In fact, low-empathy listeners did not make any differentiation in the perception of bias across BT levels, given that none of the paired contrasts reached significance. The high-empathy group, on the other hand, clearly differentiated across boundary types, with significant contrasts between the lowest levels (BT1 and BT2) and the highest ones. The only exception is represented by the highest boundary level (BT6), which did not reach significance in interaction with any of the other levels. Table 3 shows the details of the multiple comparisons across boundary tone levels for the high EQ listeners group.

Figure 7

Line plot of the interaction between BT height steps and EQ score in the assignment of Certainty values.

The interaction between Exposure and PA levels was also significant, as predicted by H2 (F(2, 5788.43): 6.58, p < .001). Figure 8 reports a plot with the interaction. Post-hoc analyses showed that listeners who were exposed to non-native phonological systems could not reliably classify the three PA levels as carrying systematically different degrees of perceived bias. Specifically, the analysis of their ratings showed that they only reliably discriminated PA1 from the other two levels (PA2: Estimate: –7.89; SE: 1.16; z: –6.76; p < .00001; PA3: Estimate: –9.34; SE: 1.16; z: –7.99; p < .00001), while assigning non-significantly different degrees of Certainty to PA2 and PA3 (Estimate: –1.44; SE: 1.17; z: –1.22; p < .43). On the other hand, listeners who were only exposed to their native variety showed a consistent rating of the three different pitch accent levels as conveying different degrees of Certainty, with significant results for all pairwise comparisons (PA1-PA2: Estimate: –4.43; SE: 1.06; z: –4.14; p < .001; PA1-PA3: Estimate: –11.57; SE: 1.05; z: –10.93; p < .00001; Estimate: –7.14; SE: 1.06; z: –6.71; p < .00001).

Figure 8

Line plot of the interaction between PA height steps and Exposure in the assignment of Certainty values.

In line with the predictions, when considering the three-way interaction with EQ, the differences between the two Exposure groups appear to be even more prominent. The three-way interaction among PA, Exposure, and EQ in fact reached significance (F(2, 5788.43): 6.60, p < .001). The plot in Figure 9 shows the relationship among these three variables, while Table 4 reports the details of the pairwise comparisons.

Figure 9

Line plot of the interaction between PA height steps, Exposure, and EQ score in the assignment of Certainty values.

Clearly, listeners with exposure to non-native systems failed to associate different degrees of Certainty to the PA2-PA3 pair, regardless of their EQ levels. However, as shown in Figure 9, high-empathy listeners with no exposure to non-native dialects made a clear differentiation across the three PA levels, with the narrowest pitch span being associated with a negative bias and the widest pitch span to a positive bias. Low-empathy listeners within the same Exposure group, on the other hand, rated Certainty for all the PA levels, on average, above 50, though maintaining a significant contrast among them. Additionally, the plot also shows a tendency for low-empathy listeners to behave in a very similar fashion in the two Exposure conditions, while listeners with higher EQ scores showed a stronger contrast between the Exposure conditions. The interaction between Exposure and PA level only reached significance for PA2 (Estimate: –8.21; SE: 4.05; z: –2.02, p = .043).

Finally, in order to address the specific hypothesis that Exposure would affect the variability of responses, we compared coefficients of variation within Certainty score ratings given by the two Exposure groups. Results of the Asymptotic test yielded significant results (test statistics: 5.327, p = .021), though, contrary to predictions, non-exposed listeners yielded higher coefficients of variance (0.54) with respect to exposed listeners (0.51). However, visual inspection of the dispersion of response data suggests an interaction between Exposure and EQ. The boxplots in Figure 10 show that while a higher variation of responses was found only for the non-exposed High EQ listeners, a smaller variation was registered for low-empathy exposed listeners (Figure 10, right panel).

Figure 10

Boxplots of the distribution of the assignment of Certainty values for High EQ (left) and Low EQ (right) listeners in the two Exposure groups (x axis).

Comparisons of equality of variation distributions between the two EQ groups of the exposed listeners yielded indeed significant results (test statistics: 30.203; p < .0001), showing that EQ score modulated the variance of responses within the exposed group.

6.1. Summary of results

In a perception study, we investigated the role of pitch span within the nuclear configuration region of Salerno Italian yes-no questions in the recovery of a speaker’s epistemic bias. Two different base tunes (L+H* HL-L% and L+H* HL-H%), differing in the direction of the boundary tone (rising or falling) were stylized then resynthesized with manipulated pitch span (gradually larger f0 excursion between L and H targets) for both pitch accent and boundary tone region and were then presented to listeners, who evaluated each stimulus in terms of perceived speaker Certainty about a ‘yes’ response. Results show that pitch span has a crucial role in listeners’ identification of speakers’ epistemic bias and, in the specific case tested here, in evaluating whether the speaker expects a positive answer to a question.

The general effect that was detected is in line with our main hypothesis (H1), which is that a wider span in both rising pitch accent and rising boundary tone region would convey a higher degree of a speaker’s positive bias (hypothesis borne out of observation of SI production patterns), hence resulting in higher Certainty scores. What is more, the separate contributions of pitch accent span on one side and of boundary tone span on the other appear to have an additive effect, so that the highest Certainty scores were recorded for combinations of higher PA and BT steps. Additionally, the effect of variability at the listeners’ level was investigated and, crucially, results point towards a combined effect of both the cognitive (EQ) and our sociolinguistic (Exposure) variables, as predicted by H2. Specifically, the significance of these variables was found both in interaction with tonal cues and with each other, showing that they significantly and gradually affect the way epistemic bias is extracted from intonational cues. Additionally, we found a difference in the variation of responses between the two Exposure groups, which was crucially modulated by the EQ score. While this specific result does not directly confirm our predictions (higher response variation for exposed listeners), it provides evidence for the interaction between Exposure and EQ. In general, these results strongly suggest that epistemic bias identification is not homogeneous throughout the listeners’ community, and that individual cognitive skills interact with regional dialect exposure effects.

6.2. Gradience of cues and meaning

First, the findings reported here add to the literature on biased questions and intonation. A link had already been suggested also for other languages (e.g., Prieto & Borràs-Comes, 2018 for Catalan) and for other varieties of Italian (Savino & Grice, 2011 for Bari Italian) given that, as reported in the introduction, intonation is argued to have a prominent role in the production and perception of epistemic bias in yes-no questions. Our results contribute to this literature by uncovering the crucial contribution of gradual pitch span variation to perceived epistemic bias. As for the pitch accent region, Certainty scores were gradually affected by the step manipulation, suggesting that gradient pitch cues do not necessarily need to map onto discrete meaning categories. Hence, an important contribution of these findings is that gradual intonation cues map onto gradient meanings of pragmatic nature, beyond traditional phonetics/phonology mapping. As for the boundary manipulation, our results suggest a stronger mediation of the phonological level, in that only the lowest and narrower stimuli (BT1 and BT2) behaved differently from all the others.

As reported in the introduction, most studies of intonational meaning build on the assumption that specific intonational categories align with specific expression of speakers’ epistemic dispositions in a dichotomous way, e.g., Bartels’ (1999) L-/H- opposition mapping onto the presence/absence of an ‘assert morpheme’ or Gunlogson’s (2004) rise/fall opposition mapping onto presence/absence of speaker commitment. Nevertheless, recent contributions within Conversational Analysis have proposed that epistemic disposition can be better modeled in a gradient rather than a categorical way. Heritage (2012), for example, has proposed that English can encode several steps along a continuum reflecting gradient epistemic stance. For example, interrogatives (Are you married?), tag-questions (You’re married, aren’t you?), and declaratives (You are married.) would encode three different steps from an ‘unknowing’ to a ‘knowing’ epistemic stance relative to a specific matter, i.e., the proposition ‘You are married.’ Additionally, several linguistic strategies might be used to encode intermediate degrees of epistemic stance and, crucially, these strategies might not be interchangeable. Oshima (2017), for example, proposes that common strategies used in English to express a biased question (e.g., negative polarity and tags) carry information regarding the relative degree of bias. Specifically, tag questions might convey a higher degree of positive epistemic bias relative to negative polar questions. Our perception results can be interpreted in light of these proposals, though by assigning a stronger contribution to intonational cues, which alone would map onto several steps along a continuum of epistemic stance.

This provides further evidence against the traditional view of pitch range variation as only conveying paralinguistic meanings (as in Liberman & Pierrehumbert’s [1984] ‘free range hypothesis’). To the best of our knowledge, no intonational meaning model has attempted to formalize the contribution of gradual phonetic cues in the expression of meaning. The only model that assigns specific importance to relative pitch height/span information is Gussenhoven’s (2002; 2016) Biological Code theory, in which it is predicted that the rules modulating the meanings of intonational categories are yielded by grammaticalizations of universal meanings assigned to f0 (and intensity) modulations. Crucially, however, the gradient side of the form-meaning mapping in Gussenhoven’s theory is limited to the expression of paralinguistic meaning. For example, according to the Frequency Code, the paradigmatic opposition between high and low fundamental frequency (or rising and falling movements) would universally signal, respectively, an opposition between submissiveness and/or uncertainty on one side and confidence and/or finality on the other. On the grammatical side, this paralinguistic opposition would give rise to the opposition between questions and statements, modulated by language-specific constraints (see Chen et al., 2004). Similarly, according to this view, an expanded rise span would only signal non-linguistic meanings, such as polite attitude. The Effort Code, in a similar fashion, would predict that wider pitch movements would universally signal emphasis or surprise.

The findings reported here, however, do not easily fit into such a universal way of deriving the meaning of intonation. First of all, the gradience that we report is pragmatic in nature and therefore differs from Gussenhoven’s predictions that paralinguistic gradience is grammaticalized in distinct linguistic categories. Additionally, the oppositions found between the way high versus low pitch cues are mapped onto bias can hardly be explained as deriving from such universals. The clearest example is given by the Certainty ratings resulting from boundary tone steps. On the basis of predictions derived from the Frequency Code, we would have expected that higher BT steps would have yielded lower Certainty scores, while what we found was the opposite. As mentioned in the introduction, several predictions of the Frequency Code have not found support in recent experimental research investigating the role of pitch height in the expression of deference and politeness in Korean (Brown et al., 2014) and in Russian (Chikulaeva & D’Imperio, 2018). As for Korean, it appears that politeness is mainly conveyed by loudness variation and not pitch (Idemaru, Winter, & Brown, 2020), rendering the idea of universality of paralinguistic use of gradual pitch cues rather problematic. Also, incredulity meanings have often been linked to expanded pitch span (e.g., Hirschberg & Ward, 1992; Savino & Grice, 2011; Borràs-Comes et al., 2014), which could easily be interpreted as a grammaticalization of the Effort code. This is not in line with our findings, though, in that the link found here between narrower span in pitch accents and negative bias appears to go against the universality of intonational meaning predicted by this code. Therefore, our results point to the need to formulate language- and dialect-specific mappings between gradual acoustic-intonation cues and gradual pragmatic meanings.

The gradient relationship between gradient cues and meaning resulting from our study would be compatible with recent experimental evidence reported by Holliday and Villareal (2020), who test the social meaning of intonation. Specifically, the study investigates the link between both phonological elements (pitch accents) and phonetic detail (continuous pitch span and voice quality information) and social meaning related to speaker blackness. The findings of this study point towards meaningful variability of both phonological and phonetic information, with an incremental effect of continuous phonetic cues. Furthermore, the authors also find that the meaning of gradient cues is not shared across phonological categories. While for L+H* (phonologically considered more frequent in African American English than H*) gradient span contributed to assess a stronger social meaning related to blackness, the same was not found for H* (though results for phonological manipulation of the pitch accent did not show different listener behaviors). This suggests that phonological and phonetic properties of intonation work together in determining aspects of intonational meaning and, crucially, that variable meanings retrievable from intonational allomorphs might arise from modifications of the core meaning of that category.

One crucial difference, however, between the treatment proposed by Holliday and Villareal (2020) and our own is in the type of gradient meaning that maps onto the continuous cues. While Holliday and Villareal consider social meaning linked to the ethnicity of the speaker, we focused on dialectal variability. Nevertheless, a similar relationship between phonology and phonetics in the construction of meaning might be applicable to our data as well. While we did not specifically test for the relationship between phonological and phonetic elements (e.g., meaning of span modifications in two phonologically different pitch accents), we can speculate on the mediation effect provided by the phonological category.

Independent evidence for SI about the meaning of span modifications in different pitch accent categories has already been reported in Orrico and D’Imperio (2020) as far as wh-questions are concerned. Specifically, the study reports that a greater span in L*+H accents is linked to counter-expectational meanings (in terms of rejection of a previously uttered proposition), which has been analyzed as a lack of speaker commitment to a relevant proposition. On the other hand, the results reported in the present study appear to assign the opposite meaning to a greater span, which would suggest some degree of mediation of the phonological form. Similar results were reported by Chikulaeva and D’Imperio (2018) for the tonal encoding of politeness in Russian. Specifically, they report data from a production study in which f0 height in pitch accents positively correlated with the expression of politeness. Crucially, however, the same effect was not found for all pitch accents categories (e.g., different from H*, H+!H* accents were not found to modulate politeness as a function of f0 height), providing additional evidence of the interaction between phonology and phonetics of intonation in the expression of meaning. We, hence, suggest that while intonational phonological units might possess meanings, modifications at the sub-phonemic (more precisely sub-morphemic) level might be employed to modulate those core meanings. This treatment of intonational meaning and, above all, of the way phonological and phonetic elements cooperate to the construction of the global meaning is reminiscent of the proposal made by Gili Fivela (2008) for Pisa Italian (a central variety), in which it is argued that different ‘shades of meaning’ of a given phonological category can be conveyed by phonetic modifications within that category.

The meaningful character of phonetic modifications of a phonological category, however, does not appear to apply to all the structural cues of an intonational contour. As a matter of fact, such behavior of the phonology-phonetic interplay in the transmission of meaning emerges, in our data, only to pitch accents, while the same cannot be said for boundary tones. Boundary tone level was in fact found to bear intrinsic meanings related to the expression of question epistemic bias, since finer span modifications within the two broad phonological categories (L% and H%) did not appear to contribute to further characterize meaning. It is possible, though, that continuous modifications of boundary tones can map onto continuous meanings in the same way as pitch accents do, though for other dimensions of meanings which we did not explore here.

Finally, an effect of base tune was also found in our data. Specifically, the stimuli created on the basis of the rise-fall L+H* HL-L% base stimulus were assigned lower overall Certainty scores. This effect indicates that cues beyond pitch are also exploited to infer epistemic bias in SI. Note that the effect of Base also interacted with PA steps, hence determining how step height was employed for Certainty rating in that region. We do not know how f0 directionality might be related to spectral information or other non-pitch intonational cues, given that this is an area that has received less attention. However, these results are in line with findings suggesting that, for example, local duration cues play a role in the production of sentence modality in Neapolitan Italian (Cangemi & D’Imperio, 2015) adding to the complex relationship between the phonetics-phonology mapping on the one hand, and the phonology-pragmatic meaning mapping on the other.

Taken together, our results provide useful insight on the relationship between intonational cues and pragmatic meanings. First of all, continuous phonetic modifications within a phonological category are not randomly variable, rather they carry specific meanings of pragmatic nature. Such meanings, we posit, are not completely independent from the phonological category they belong to, but they represent modifications of the core meaning of that category. In other words, the phonetic space that an allophonic modification occupies is informative with reference to the strength of the meaning associated with the phonological category. Also, these modifications to the core meaning are incremental, which means that the specific meaning of a given phonological category is modified stepwise along a continuum of gradient modifications of the pitch accent form. In this way, gradient intonational cues at the sub-phonemic level map onto gradient pragmatic meanings.

6.3. Exposure and Empathy interaction

Crucially, our data suggest that the relationship between gradient cues and gradient meanings is not stable within a linguistic community. This is because the way in which each manipulation step is mapped onto a specific degree of bias varies also as a function of listener-specific characteristics at both social (exposure to other systems) and cognitive (degree of empathy) levels. These two types of variable would then operate over two distinct planes for the construction of specific form-meaning mappings.

Exposure—a variable of social nature—operates over the organization of the phonetic space in terms of width and shape of the category space (see also Clopper, 2014 for segmental models). Prolonged exposure to other phonological systems might cause the listener to widen the specific category and, in general, to develop more flexible category boundaries. In other words, listeners who undergo the process of long-term exposure to other dialects or languages are, generally speaking, exposed to much more variable input than those who do not, which is the reason why exposed listeners might attend more closely to fine phonetic information to assess the meaning of a question given the richness and variety in their input. This is reminiscent of findings reported by Chen et al. (2017), in which bi-dialectal listeners (Australian English listeners with at least one parent native of another English dialect) were found to have widened the phonetic space relative to vowel phonemes and, therefore, being much more tolerant to mispronunciations than mono-dialectal ones, though, in our case, the flexible boundaries are built for sub-phonemic categories, rather than phonological ones.

Empathy—a cognitive variable—operates, on the other hand, over the cue-meaning mapping of gradient cue-meaning itself or, in other words, it defines how fine-grained the mapping between intonational cues and meanings is. A listener with a higher degree of empathy would be much more attentive to fine phonetic details within the input and the way this detail is linked to specific (degrees of) meanings. In this way, a high-empathy listener is predicted to be able to access more steps along the cue-meaning continuum. This empathy-induced sensitivity would also explain why high-empathy listeners are much more affected by variable exposure to other systems. Empathy skills would then lead them to attend more closely to phonetic cues of the input, resulting in a faster and stronger perceptual change and causing an augmented sensitivity to category interference. The scenario described above is supported by an analysis reported in Figure 11. For the sake of simplicity, we will restrict this section of the discussion only to our pitch accent manipulation.

Figure 11

Density plot illustrating the difference in epistemic bias ratings for the three PA pitch accent span steps in the four different groups of listeners: non-exposed/High-EQ (top-left), non-exposed/Low-EQ (top-right), exposed/High-EQ (bottom-left), and exposed/Low-EQ (bottom-right). Different shades refer to different span steps within the L+H* pitch accent, with PA1 indicating the lowest span step and PA3 the highest. While showing that for all listeners gradient span cues can be associated with different degrees of meanings, the plot crucially shows the way in which both Exposure and EQ interact with the way in which intonation-meaning mapping gradient cues-gradient meanings mappings is shaped within the individual grammar.

The density plots reported in Figure 11 above show the effect of the interaction of our two variables (Exposure and EQ) on the span manipulation for the PA. The plot was created by using a simulated dataset, based on results obtained in the experiment reported in this paper, reporting therefore the way in which each group rated each pitch accent step. Specifically, the dataset was created by generating random data points, normally distributed around the means of each one of the four participant groups (non-exposed/High-EQ, non-exposed/Low-EQ, exposed/High-EQ, and exposed/Low-EQ). Standard deviation values were reduced relative to the results of the experiment. This was done in order to eliminate some of the noise, coming mainly deriving from the interaction between pitch accent steps and boundary tone steps. However, proportions for standard deviations across pitch accents steps and across the four different groups were the same as the actual dataset resulting from the experiment. As anticipated above, the plot refers to the three pitch accent steps (PA1, PA2, and PA3).

The first thing to be noticed from the plot above is that some degree of gradience is present for all the listeners independent of the specific sub-group they belong to. Recall that the gradient cue modifications for pitch accents did not involve a categorical shift at the phonological level, supporting the idea that it is hence possible to specify meaning of pragmatic nature at a sub-phonemic level. If we focus now on the non-exposed groups (top panels) we see that both high- and low-EQ listeners rated each step independently from each other (given the shape of the distributions), though with a difference in magnitude. Specifically, data for the (non-exposed) high-EQ group reveal the existence of a larger distance in terms of Certainty scores across the three steps. This might be taken as evidence of the existence of meaningful intermediate steps (e.g., between PA2 and PA3), which would point towards a more fine-grained mapping. On the other hand, the exposed groups (bottom panels), and in particular the high-empathy individuals, globally reduced the distance between the two highest steps (PA2 and PA3), indicating that their phonetic space might have been reorganized in relation to the meaning expressed, i.e., by enlarging category space they are specifying the highest degree of meaning as being expressed by a larger sub-phonemic category with reference to non-exposed listeners. In other words, being exposed to a different phonological system would not simply cause them to lose the ability to link gradual modifications of the pitch accent to different degrees of meaning, rather it would affect the way in which some of the manipulation steps are linked to gradual meanings. Similar to the non-exposed group, the role played by EQ is expressed in terms of the magnitude of the effect, with (exposed) high-EQ listeners showing a stronger merge of the two steps, hence pointing again towards a more developed sensitivity of these listeners to elements within the input (which in this case are predicted to be unreliable), leading to a stronger effect of Exposure in terms of native category boundary blurring.

These findings are in line with exemplar approaches to sociolinguistically induced segmental variability, such as those discussed in a study on different levels of exposure to either General or Northern American English (cf. Clopper 2014), proposing that the combined distributions of exposure to both local and standard varieties are expected to be more complex and potentially more overlapping (due to spread of phonemic distributions in the combined dialect language users). As noted by Clopper (2014), various classes of exemplar models can indeed account for simultaneous effects of facilitation, on one side, and interference, on the other, due to multiple levels of linguistic and indexical information which are simultaneously stored and represented. Moreover, exemplar-based models predict that both recency (short-term novel exposure) and frequency (long-term exposure) interfere with lexical processing so that individuals with exposure to lower levels of sociolinguistic and dialectal variation “should have relatively less variable distributions of phonological and/or lexical exemplars” (Clopper 2014, p. 71). Hence, the dialect exposure effect found in our study appears to mimic what has been already claimed in the segmental/lexical phonology literature, and could be accommodated by similar models designed for intonational phonology purposes.

As reported above, exemplar models build on the idea that detailed information is retrieved from linguistic input and is stored in memory together with contextual information. Exposure, therefore, is deemed to be a key factor in this process, also to navigate the uncertainty of the speech signal (as predicted by inference-based models). The role of individual EQ levels, though, cannot be subsumed by any of the existing models. In other words, extrapolating category distribution from Exposure (both short-term and long-term) cannot be taken to be a uniform process independent of individual cognitive skills. As noticed above, shape and the width of the density plots in Figure 11 are taken to mirror category distribution though not independently of individual empathy skills. Hence, both language change and immediate processing outcomes, at least as far as the form-meaning mapping is concerned, cannot be entirely predicted by current exemplar models. Though these models allow perception-production interference, they still do not predict category-formation and processing variability that would have a cognitive basis.

The effect of empathy discussed here might be considered as a variable that modulates the sensitivity of the listener to the information that is present in the input and her ability to consolidate the indexing of the variable linguistic form as a function of contextual information. It is not clear at this point, however, at what level this variable operates, whether it affects the mapping between portions of phonetic space with specific meanings (or contextual information), or the partition itself of the phonetic space. Future research should aim at addressing this issue.

6.4. The role of social and cognitive variables in phonological categorization: Which perspectives?

The exact picture of the mapping between prosodic cues and pragmatics is, given our results, rendered even more complex by considering the role of variability at the listener level. Our results show that both listeners’ exposure to other phonological systems and their EQ score have a significant impact on the way pitch span is exploited to infer epistemic bias. As for the role of Exposure, in our data this variable appeared to interact mainly with PA step manipulation. In fact, while listeners who were only exposed to their native dialect significantly linked the three different PA steps with different degrees of bias, exposed listeners showed a less gradual and more binary perception behavior (by providing different rating only to the lowest and the highest PA steps). While it is well documented that speech encodes socio-indexical information on the part of the speaker, these results point out the need to consider different kinds of individual variability also at the listener/perceiver level.

At a first look, our findings are easily accounted for by models arguing for the role of past experience (speech input) in phonological representation, as for instance exemplar-based models (Johnson, 1997; Goldinger, 1998; Pierrehumbert, 2001; Todd, Pierrehumbert, & Hay, 2019). The basic idea behind these models is that past linguistic experience is stored in an individual’s representation in a very detailed way. This theory has been fueled by experimental findings supporting the idea that mental representation of sound structure is continuously updated and enriched with phonetic details as an effect of exposure to new instances of a given category (Goldinger, 1998; 2000). This view is against the traditional treatment of mental representation of phonology in which continuous, and more crucially variable, phonetic details are treated like meaningless noise and filtered out by listeners in perception (e.g., Halle, 1985). These two positions are poles apart as far as abstract representation is concerned, with one position seeing the abstract representation as a by-product of the processing of similar sounds, while the other taking the abstract category as the only information that is actually stored.

However, more recent treatments have assumed a hybrid perspective, positing the importance of taking into account an enriched phonological representation (e.g., with community-specific information stored into memory) but, at the same time, by also stressing the importance of abstract representation, needed for instance to account for the perception and production of new words, which cannot be captured in early exemplar models (see Pierrehumbert, 2002; German et al., 2013 for arguments supporting this view). In other words, both abstract phonological information and gradual phonetic detail are stored into memory, and influence both production and perception of upcoming words and utterances.

Similar mechanisms appear to emerge for the intonation-meaning mapping. Recent evidence about this comes from recent work on social implicit priming in intonation reported in Portes and German (2019). As briefly reported in the introduction, the study reports on the way Corsican French listeners process the meaning of a rise-fall tune uttered by Continental French speakers, in terms of question or statement interpretation. The contour tested is a specific rise-fall that is grammatically legal in both varieties, though with different meaning, i.e., question in Corsican and statement in Continental French. Results of this study show that the priming effect, that is evoking either France or Corsica tune-modality mapping, is indeed significant, with listeners exposed to the Corsican prime showing higher question responses. Portes and German (2019) explicitly discuss these results framing them within an exemplar-based model predicting that, as an effect of exposure to the two varieties of French, different form-meaning mappings would be stored in memory, each associated with a specific variety. The role of the regional prime would therefore be that of activating one of the two mappings. Interestingly, the authors posit a great importance of the effect of relative exposure to the two varieties, which would have the role of modulating the strength of the priming effect. More specifically, they argue that a balanced exposure to the two varieties would result in the strongest effect of the prime manipulation. Even though our study does not allow us to make specific predictions about how different mappings between intonational cues and expression of epistemic bias in questions can be activated, our results are in line with the findings of previous studies in the sense that they clearly show that individuals can variably process intonational cues (e.g., the relationship between pitch span and identification of bias) as an effect of having been exposed to non-native phonological systems (a variable that the studies reported above did not directly manipulate).

An alternative model allowing for explaining how phonemic perception can change as an effect of variability in the input has been proposed by Kleinschmidt (2019), in which segmental perception is modeled, using Bayesian rules, as a process of inference under uncertainty. Specifically, the author claims that listeners manage to adapt to speaker-specific variability because each dimension of speaker variation has a given degree of informativity and of utility. These dimensions of variability might be more or less informative with regards to the distribution of cues used to produce a given category, while also being more or less useful for the listener to predict (given particular indexical information of the speaker) the mapping between phonetic cues and phonological categories. In other words, listeners possess prior expectations about linguistic variability, in terms of both what can be variable within the speech signal as well as the source of such variability (e.g., dialect or gender), hence exploiting such knowledge to navigate inconsistencies in the speech input. We also know that listeners can rapidly adapt to variable intonational input. For instance, Kurumada, Brown, and Tanenhaus (2012) and Tanenhaus, Kurumada, and Brown (2015) report that listeners, in order to navigate variability in the intonational input, are able to rapidly adapt to speaker-specific uses of intonation to encode pragmatic meanings. In other words, they are able to update their expectations about speakers’ intended meaning in specific contexts based on past experience. Note though that both inferential and adaptive approaches would need to explain perceiver-specific effects in light of our results.

Apart from the role of socio-indexical priming in determining intonational meaning mapping, also argued for uptalk use in New Zealand English by Warren (2017), the role of cognitive variables might modulate how the perceiver extracts epistemic stance. First of all, it appears that, as also suggested by the literature reviewed in Yu and Zellou (2019), not only socio-indexical information but also listeners’ cognitive skills have a role in shaping the way speech is perceived and processed. As for the effect of EQ quotient in our study, it appears that the degree of empathy mainly affects the way listeners interpret tone boundary information, with low EQ listeners failing to employ variable span in boundary tones to the purpose of determining epistemic bias, as opposed to high EQ individuals. In addition, our results also suggest that high EQ individuals might use a wider range of cues to interpret epistemic stance, relative to low EQ ones. This is in line with recent literature arguing for different interpretative strategies in individual listeners, such as the difference between pragmatic responders and semantic responders (Degen & Tanenhaus, 2015), and high-empathy levels being linked to greater (as well as earlier) use of pitch cues in recent Eyetracking data (Estève-Gibert et al., 2020) on intonation processing in French. However, different from the study reported in Estève-Gibert et al. (2020), in which listeners could also rely on later lexical disambiguation, here only intonational cues were available for the interpretation of the speaker’s intended meaning. A heightened sensitivity to pitch cues appears to be also true for individuals scoring low in the Autism Spectrum quotient (Jun & Bishop, 2015), which is subsumed by the cognitive component of the EQ.

In our experiment, remember that EQ was found to also have an effect in modulating the degree of the Exposure effect, with higher EQ predicting that the listener would be more affected by exposure to other phonological systems. In other words, while it can be reasonably assumed that exposure leads to the creation of a new form-meaning mapping having a specific contextual label, we should avoid assuming that the same type and length of exposure would have exactly the same effect on all individual listeners. The notion of an ideal listener navigating the many sources of production variability and contextual effects should hence be extended to include perceiver-specific processing styles. This would also apply to inferential Bayesian models accounting for the bidirectional role of socio-indexical and linguistic cues in segmental perception (Kleinschmidt, 2019). In fact, neither current versions of exemplar-based or inference-based models can account for the empathy interaction effects reported here.

Note that in early sociophonetic accounts of indexical variability, a dichotomy was introduced to distinguish between speaker versus ‘talker’ normalization (Johnson, 1990; Johnson et al., 1999), to underline the fact that phonemic categorization is guided by individual talker variability that goes beyond mere speaker (i.e., anatomically based) differentiation. Analogously, our data point to the role of an individual ‘perceiver,’ endowed with specific cognitive skills, shaping category retrieval of intonational form and meaning. In other words, indexically based talker variability is not sufficient to account for exemplar space structure in guiding phonological processing. Hence, perceiver-specific processes need to be integrated with more general listener processes in accounting for phonological processing. The notion of an ideal listener navigating the many sources of production variability and contextual effects should be accompanied by perceiver-specific processing styles. This applies also to inferential Bayesian models trying to account for the bidirectional role of socio-indexical and linguistic cues in segmental perception (Kleinschmidt, 2019).

Hence, independent of the specific perception model that is adopted, be it exemplar or inferential, our claim is that we cannot simply assume that listeners’ strategies in making sense of variability in the speech signal would be necessarily homogeneous across a linguistic community. Future work should take into account different sources of individual variability and its impact on the intonation-meaning mapping as well as on phonemic activation, possibly by postulating that cognitive, perceiver-based specificities would both modulate socio-indexical priming effects as well as allow or inhibit the impact of different kinds of dialectal exposure.