Glossa Psycholinguistics

2.1 Past work on prosody in pragmatics

The role of prosody in pragmatic disambiguation has been explored across various types of utterances and speech acts. In this section, we describe research focusing on indirect requests and irony specifically, as both are cases of pragmatic ambiguity. In each case, the same construction can be used for distinct meanings, and the question of which meaning a speaker intended may be informed by the prosodic cues present in the utterance. We therefore believe that such work can inform the hypotheses that we specifically test in our present research.

Past work on both English (Trott et al., 2019; Trott et al., 2023) and French (Ruytenbeek et al., 2023) has identified several acoustic cues that reliably predict whether or not a speaker is making a request. In each language, two sets of grammatical constructions were tested: modal interrogatives (e.g., Can you open that window? in English and the corresponding Tu peux ouvrir cette fenêtre? in French) and negative state remarks (e.g., My office is really hot). An important difference between the English modal interrogatives and the French modal interrogatives in these two studies is that the latter are characterized by non-inverted word order. This is why they were called “non-inverted interrogatives” by Ruytenbeek et al. (2023, p. 108). These non-inverted interrogatives have replaced, to a large extent, inverted interrogatives in contemporary spoken French (see Wall, 1985 for empirical evidence).

Modal interrogatives and negative state remarks differ in two important ways. First, although both constructions license a request interpretation, modal interrogatives are more likely to be interpreted as requests than negative state remarks, overall (Trott et al., 2023). Second, the literal speech act conveyed by each construction is also different. A negative state remark, interpreted literally, is an assertion; a modal interrogative, interpreted literally, is a yes/no question. That is, each construction has a different set of candidate interpretations: a negative state remark can be interpreted as a request or an assertion, while a modal interrogative can be interpreted as a request or a yes/no question about ability.

Accordingly, past work also identified distinct acoustic cues predicting the original intent of an utterance, as well as how it was interpreted.¹ Modal interrogatives were disambiguated by F0 slope, i.e., the degree to which pitch rose, fell, or stayed constant over time; this was true in both English (Trott et al., 2023) and French (Ruytenbeek et al., 2023). Specifically, a more positive F0 slope was associated with utterances intended (and interpreted) as yes/no questions, and less associated with the request interpretation. This is consistent with past work on English, suggesting that yes/no questions have a rising pitch contour (Banuazizi & Creswell, 1999; Hedberg et al., 2014). In contrast, negative state remarks were disambiguated by the number of voiced frames, i.e., the duration of the utterance. In both English (Trott et al., 2023) and French (Ruytenbeek et al., 2023), longer utterances were more likely to be intended (and interpreted) as requests.

Regarding the relationship between speech act information and the prosody of French interrogatives, Beyssade and Delais-Roussarie’s (2022) comparison between the information seeking and rhetorical question uses of polar interrogatives with est-ce que showed that the former reading can be associated with a rising, a high plateau or a rising-falling contour. In another study, Delais-Roussarie & Beyssade (2019) had provided evidence that the rhetorical question use of wh- (e.g., Qui mange des épinards? ‘Who eats spinach?’) and polar interrogatives (e.g., Il pleut? ‘It’s raining?’) results in increased utterance duration. In the same vein, Celle and Pélissier (2022) documented the prosody of the information seeking reading and the surprise question reading of Qu’est-ce que…? interrogatives, a construction that can be used to express criticism (Coveney and Dekhissi, 2017). They found that, as surprise questions, Qu’est-ce que…? interrogatives are prosodically differentiated using slower speech rate (i.e., longer mean duration of the first syllable); they are also less frequently produced with a rising final intonational contour.

There is also evidence that the comprehension of irony depends, in part, upon prosodic information. Several studies have shown that speakers of different languages, i.e., English, German and Spanish, use prosodic cues to signal ironic meanings and that comprehenders rely on the same cues when interpreting ironic utterances (e.g., Bryant, 2010; Cheang and Pell, 2008; Padilla, 2012; Rockwell, 2000; Scharrer et al., 2011). When it comes to French utterances, the cues that have been found to be reliably associated with ironic meanings are pitch raising or lowering, higher pitch variability, higher intensity, and increased utterance duration (for a discussion, see González-Fuente et al., 2016; Loevenbruck et al., 2013). For instance, Loevenbruck et al. (2013) investigated the prosody of ironic vs. literal utterances of one and the same sentence. They found that mean F0 and average duration are higher for ironic utterances of e.g., Il est en pleine forme (‘He is in great shape’) compared to their literal counterparts. They also reported on a slight final rise in F0 for ironic utterances. Their findings for mean F0 and average duration were replicated in the first experiment by González-Fuente et al. (2016), who aimed to identify the individual contribution of the features of pitch range expansion, syllable lengthening and intonational contours to the overall ironic meaning of French utterances of the type Nous sommes allées voir un film formidable (‘We went to see a fantastic film’). González-Fuente and colleagues also showed that, in the ironic utterances, the final word of the sentence was characterized by higher F0 variability.

Returning to the case of indirect requests, previous empirical work on the prosody of indirect requests has a major limitation: it cannot address the question of whether these distinct acoustic cues correspond directly to distinct speech acts, or whether the acoustic cues serve primarily to mark a deviation from the default (i.e., canonical) interpretation of an utterance.² This issue is discussed in greater detail in the section below.

2.2 How does prosody communicate intent?

As Trott et al. (2023, p. 137) note, a long-standing question in research on prosody and pragmatics is “whether particular prosodic features convey intent directly, or whether they function primarily as contrastive markers, which invite the listener to perform additional inference”. This question motivated past work on irony, e.g., on whether there is an “ironic tone of voice” (Attardo et al., 2003; Bryant and Fox Tree, 2005). It is also central to the question of prosody and indirect requests: does prosody convey a request meaning directly – i.e., is there a “requesting tone of voice” – or does it signal that a non-canonical speech act is intended? Here, we refer to these distinct hypotheses as the Direct Mapping view and the Non-canonical Marking view, respectively.

One preliminary answer to this question comes from observing the fact that distinct prosodic cues were predictive of the request interpretation across distinct grammatical constructions, i.e., modal interrogatives vs. negative state remarks (Ruytenbeek et al., 2023). This fact is inconsistent with the strongest form of the Direct Mapping view. However, this interpretation is complicated by the fact that the grammatical constructions tested (Ruytenbeek et al., 2023) differed along two dimensions. First, modal interrogatives are more likely to be interpreted as requests than negative state remarks; and second, modal interrogatives and negative state remarks have a distinct (though partially overlapping) set of candidate interpretations. Modal interrogatives are ambiguous between requests and yes/no questions, while negative state remarks are ambiguous between requests and assertions.

Thus, a weaker form of the Direct Mapping view might postulate that controlling for the set of candidate speech acts in question – e.g., a request vs. a yes/no question – the same set of acoustic cues will reliably and predictably discriminate between those speech acts. As noted in the section above, this weaker account is also partially consistent with past evidence, at least when it comes to modal interrogatives: the specific prosodic cues that predicted intent were also cues previously identified as conveying the yes/no question speech act, i.e., rising pitch contour (Ruytenbeek et al., 2023; Trott et al., 2023). A proponent of the Direct Mapping view might, therefore, claim that this weak version holds for modal interrogatives specifically.

Yet proponents of the Non-canonical Marking view might (justifiably) object that even this weaker conclusion is premature on the basis of the evidence. Namely, past work (Ruytenbeek et al., 2023; Trott et al., 2023) suffered from a confound between the salience of a request and which prosodic features were most plausibly associated with the candidate speech acts. For example, modal interrogatives have a higher probability of being interpreted as requests than as yes/no questions (Ruytenbeek et al., 2023; Trott et al., 2023). Thus, even though the prosodic features associated with the non-request speech act – a rising pitch contour – are, plausibly, meaningful, they might also merely signal a deviation from the canonical interpretation. Additionally, and unlike yes/no questions, the acoustic cues associated with the request speech act have less grounding in previous theoretical literature; this problem is further confounded by the fact that the request interpretation is the default for modal interrogatives. Evidence for even the weaker Direct Mapping view would need to demonstrate that those same features predict the intended speech act, regardless of whether the request or yes/no question meaning is the canonical one.

Here is a brief summary of the contrasting views and hypotheses outlined thus far:

• A.  H₁ (Direct Mapping view): For interrogatives specifically, there is a predictable, consistent mapping between sets of acoustic features and intended speech act and interpretation (i.e., request vs. yes/no question).

• B.  H₂ (Non-canonical Marking view): Even among interrogatives, the mapping between acoustic features and intended speech act will depend not only on the intended speech act, but also on which candidate interpretation is most salient (i.e., request vs. yes/no request).

Of course, it is important to note that additional nuance exists within this hypothesis space, particularly in terms of how prosodic cues might influence pragmatic inference “in the wild”. For example, it is possible that there is not a direct mapping between acoustic features and a comprehender’s interpretation (thereby negating H₁), and also that comprehenders do not merely rely on prosodic marking to search for plausible alternatives to a salient interpretation (thereby negating H₂); instead, or in addition, comprehenders might also deploy context-specific cues that raise or lower the probability of a given interpretation. For the purposes of the current work, we focus on the narrower question of which cues signal which intents for a controlled set of utterances without explicit discourse context (see the norming studies and explanation below).

It is unclear which hypothesis is best supported by existing evidence. Answering this question is crucial for theories of speech production and the role that prosody plays in signaling intent more generally: does the production process “select” a prosodic contour directly on the basis of the intended speech act (as implied by the Direct Mapping view), or does it generate marked features dependent on whether the intended speech act is the canonical one (as implied by the Non-canonical Marking view)? Notably, in the latter case, the question of what is “canonical” is by no means trivial; the most salient interpretation of a given construction could vary, depending on the situational context, the relationship between speaker and addressee, previous turns in the discourse, and more. We operationalize canonical in an admittedly narrower sense, in terms of the interpretation that comes to mind most easily in the absence of specific contextual cues supporting any particular interpretation (see the norming studies below).³

These possibilities may also have distinct implications for the role of audience design in the speech production process (Ferreira, 2019). Specifically, the Direct Mapping view raises the possibility that speakers could generate a disambiguating prosodic pattern more “automatically”, without having to weigh the possible interpretations of the utterance. The Non-canonical Marking view presents a more complicated situation. On one account, speakers may need to first identify the default interpretation of their utterance, then produce a prosodic pattern that would convey a departure from that interpretation, i.e., engaging in what Ferreira (2019) would describe as recurrent audience design. Another possibility, however, is that speakers still engage disambiguating prosodic cues relatively automatically, by relying on contextually-appropriate linguistic knowledge that identifies the relative salience of different speech acts for a given utterance in a given context (thus eliminating the need to “simulate” a comprehender’s likely interpretation).

The hypotheses also have important implications for language processing and comprehension. If acoustic cues convey speech acts directly, it suggests that – in principle – pragmatic inference could be performed by mapping specific cues onto stored representations, i.e., constructions that map between prosodic features and speech acts. But if acoustic cues primarily serve to mark a deviation from the default, it suggests that the role of prosody is primarily as the catalyst to a more complex inferencing process. These possibilities have different theoretical implications for how speech act information is stored and accessed, as well as for the time course of pragmatic inference.

2.3 Present work

The present work aims to adjudicate between these hypotheses for a restricted set of French constructions – namely, grammatical constructions whose candidate interpretations are either request or yes/no question. This includes modal interrogatives, which have already been tested in previous work on French (Ruytenbeek et al., 2023), and which were also included in the current work. It also includes non-modal interrogatives (Tu fermes cette fenêtre?), as well as a modified version of both types of interrogatives with the interrogative particle est-ce que (ESQ) (Est-ce que tu peux fermer cette fenêtre? and Est-ce que tu fermes cette fenêtre?). These are described in Table 1; note that the values in the p(request) column were obtained from the norming studies, which we describe below.

Importantly, each of these constructions plausibly licenses both a request and a yes/no question interpretation. But, equally importantly, the relative salience of those interpretations differs across the constructions. Past work (Ruytenbeek et al., 2023) has already shown that a request interpretation is most salient for the French modal interrogative construction. Further, the norming studies described below have confirmed that a request interpretation is less likely for constructions without a modal (e.g., Tu portes cette valise?); within the category of non-inverted interrogatives, the presence of est-ce que was associated with a lower probability of a request interpretation.

Thus, these constructions provide an ideal contrast for addressing the outstanding question from above, as H₁ and H₂ make distinct predictions.

• C. H₁ predicts that the same acoustic cues will predict speech act (request vs. yes/no question), regardless of the construction used.

• D. H₂ predicts that different acoustic cues will predict speech act (request vs. yes/no question), depending on which construction is used.

Put another way: if H₁ is correct, then the results for the new constructions tested should resemble past findings for the non-inverted modal interrogative construction, i.e., Tu peux VP? (Ruytenbeek et al., 2023). When intended as a yes/no question, non-modal interrogatives (with and without est-ce que) should have a rising pitch contour; when intended as a request, they should not.

But if H₂ is correct, then the results for non-modal interrogatives will look different from those found in past work (Ruytenbeek et al., 2023). Critically, there are multiple ways in which they could be different. One possibility is that non-modal interrogatives are not discriminated by any of the acoustic cues tested. Another possibility is that the same acoustic cues are predictive, but in different or even opposite ways: for example, it is possible that non-modal interrogatives intended as yes/no questions will not have a rising pitch contour, but that non-modal interrogatives intended as requests will (i.e., the opposite of the findings for modal interrogatives).

To sum up: the hypotheses make distinct empirical predictions about which prosodic cues will predict intent (and interpretation) across the grammatical constructions tested. This will allow us to adjudicate between the hypotheses on the basis of the study’s results. While the constructions tested represent only a narrow slice of the broader prosody/pragmatics interface, resolving this debate has important theoretical implications not only for speech production, but also for language processing and pragmatic inference.

3.1 Materials

The utterances of French sentences used as experimental stimuli in the perception study will focus on the various kinds of interrogatives illustrated in Appendix A. All code and data required to reproduce the analyses and figures in this paper can be found on GitHub.⁵ The design and analysis (except for the norming studies) were pre-registered ahead of data collection; the pre-registration can be found on OSF (https://osf.io/gb3fn).

3.2 Norming Studies

The goal of the norming studies was to establish the salience of distinct candidate interpretations for each utterance in our set of stimuli – independent of its prosody. This is crucial for testing the hypothesis outlined in the Non-canonical Marking view: namely, that the salience of distinct interpretations affects which cues are used to signal a speaker’s intent. Importantly, we operationalize salience here as how easily human comprehenders can bring a given interpretation to mind in the absence of explicit contextual cues favoring one interpretation or the other. (As noted earlier, participants may still construct plausible contextual scenarios when evaluating the salience of different interpretations, or engage more bottom-up, compositional processes; in either case, our goal is to establish the prior probability of a given sentence being interpreted as a request in the absence of prosodic cues.)

We conducted four separate norming studies, i.e., one for each construction illustrated in Appendix A. Following Trott et al. (2023), we expected to find a main effect of Grammatical Form: specifically, we expected that written versions of modal interrogatives would be more likely to be interpreted as requests than written versions of negative state remarks. More relevantly to the current work, we predicted that modal interrogatives would also be more likely to be interpreted as requests than non-modal interrogatives. Within each construction type, we did not have a specific prediction about the impact of est-ce que.

In each of the four norming studies, we recruited 320 native speakers of French (4 × 80) from France, Belgium or Switzerland, using Prolific as a recruiting platform. Of these participants, 145 self-identified as female (172 male; 1 other). The average self-reported age was 32 (standard deviation = 10; median = 31; range = 18–69). Each speaker received 2€ for participating.

In these norming studies, participants read a series of written sentences. After reading each sentence, they were asked to determine whether or not the sentence was a request by selecting the “yes” or “no” response option.

The goal of each norming study was to establish request probabilities for each of the four construction types in question. Each study included, in addition to the 48 sentences corresponding to one of the four construction types, 16 negative state remarks (My phone is broken, cf. Appendix B), and, as controls, 16 imperative sentences (Listen to this song), and 16 declarative sentences such as Paris is the capital of France. The key differences between the norming study were whether the target construction was a modal interrogative or a non-modal interrogative, and whether the target construction contained est-ce que.

3.2.1 Results

As depicted in Figure 1, there was a clear difference in the proportion of request interpretations across constructions containing a modal (peux) and those that did not contain a modal. Further, among constructions that did not contain a modal, the presence of est-ce que (ESQ) was associated with a lower proportion of request interpretations (M = 0.30). See Table 1 for the by-condition means.

Figure 1: Proportion of request interpretations across conditions. Participants were more likely to interpret items containing a Modal as requests. Further, for items without a Modal, the presence of ESQ decreased the probability of request interpretations, though the inverse pattern was observed for items with a Modal.

To analyze these results, we compared nested statistical models with log-likelihood ratio tests. Each model was a generalized linear mixed-effects model with a logit link, built using the lme4 package (Bates et al., 2015) in R (R Development Core Team, 2019). First, we constructed the “full” model, which had Request Interpretation (Yes vs. No) as a dependent variable, fixed effects of Modal (Yes vs. No) and ESQ (Yes vs. No), an interaction between Modal and ESQ, and random intercepts for Item and Participant (random slopes were not included).

The full model, including the ESQ:Modal interaction, explained significantly more variance than a model omitting only this interaction [χ²(1) = 14.5, p < .001]. Inspecting the parameter estimates of this model, we found that Request interpretations were more likely overall for Modal constructions [B = 9.1, SE = 1.15, p < .001], less likely for constructions with ESQ [B = –6.33, SE = 1.73, p < .001], and relatively more likely for constructions with both a Modal and ESQ [B = 7.44, SE = 1.99, p < .001]. These results are illustrated in Figure 1 below, and the coefficients are plotted in Figure 2.

Figure 2: Parameter estimates for a generalized linear mixed-effects model predicting Request Interpretation (Yes vs. No) as a function of whether the construction contains a Modal or an ESQ, as well as the interaction between the two; the full model also included random intercepts for individual Participants and Items.

3.3 Procedure

3.4 Methods

The perception study had a within-participant design, in which all participants are subjected to all manipulations (Modal × ESQ × Intent). Each participant was presented with all the indirect request utterances recorded by one of the 8 speakers. Thus, each participant heard 48 utterances in total (plus 20 control utterances); 35 participants were assigned to each of the 8 lists of stimuli corresponding to the 8 speakers in the production task.

3.5 Results

Here, we report the results for each study in turn. First, we focus on the analysis of acoustic features for the production study, which includes investigating which features are predictive of Meaning and quantifying the performance of a machine learning classifier equipped with each feature. Then, we ask how successfully humans identify the intended meaning of an utterance, and whether there is any overlap in the acoustic features that predict a speaker’s original intent and those that predict a comprehender’s interpretation. All analyses were pre-registered and carried out in R, unless otherwise indicated.

3.5.3 Perception Study

Following previous work (Ruytenbeek et al., 2023), we conducted two sets of analyses for the perception study. These analyses were designed to answer three separate, but related, questions: (i) Can listeners determine the intended interpretation of an utterance using its prosody?, (ii) Does their success depend on the type of construction in question?, and (iii) Which acoustic features are most associated with a listener’s pragmatic interpretation, and does the effect of each feature depend on the type of construction?

For the first question, we constructed a generalized linear mixed-effects model with a logit link, predicting a participant’s Interpretation (Request vs. Non-Request). We included a fixed effect of the speaker’s original Intent (Request vs. Non-Request), as well as random intercepts for item, speaker, and participant. As pre-registered, we also added by-participant random slopes for the effect of Intent. We compared this model to a model omitting only Intent and found that the full model did indeed explain additional variance [χ²(1) = 49.21, p < .001]. Specifically, items intended as requests were associated with a higher likelihood of being interpreted as a request [B = 0.35, SE = 0.04, p < .001]. Overall, 60.5% of Literal items were interpreted as requests, while 68.5% of Request items were interpreted as requests (see also Figure 3 below). Thus, regardless of the construction in question, a speaker’s original intent was predictive of a listener’s pragmatic interpretation, though the effect was small (mean accuracy was 54%). Consistent with the classifier results above, some speakers also produced more discriminable cues than others: speaker-wise accuracy ranged from 49.5% to 59.6%. Participants also varied considerably in their performance on the task, from 34.7% to 85.4% (median = 52.2%).

Figure 3: Rate of request interpretations by original speaker Intent (Literal vs. Request), as well as the constructional features of the item, i.e., whether it contained a modal or est-ce que.

We then added fixed effects of Modal and ESQ, as well as interactions between Modal:Intent and ESQ:Intent. Using a series of nested model comparisons, we asked whether each of the interactions explains more variance than a model containing only the main effects: namely, does the success with which a speaker conveys their intended interpretation through prosody depend on whether the construction contains a modal or est-ce que? A model containing both interaction terms explained more variance than a model omitting only the interaction between Modal:Intent [χ²(1) = 24.97, p < .001], as well as a model omitting only the interaction between ESQ:Intent [χ²(1) = 47.9, p < .001]. Specifically, items intended as Requests were less likely to be interpreted as such when they did not contain a modal [B = –0.64, SE = 0.13, p < .001], and items intended as Requests were also relatively less likely to be interpreted as such when they contained est-ce que [B = 0.77, SE = 0.07, p <.001]. Parameter estimates for each of the main effects and interaction terms from the full model are depicted in Figure 4.

Figure 4: Coefficient estimates for each of the main effects and interaction terms in the full model predicting Interpretation.

Further model comparisons allowed us to identify the independent effects of Modal and ESQ, i.e., whether listeners were more or less likely to interpret an utterance with a modal (or est-ce que) as a request. A model containing main effects of Intent, Modal, and ESQ explained more variance than a model omitting only Modal [χ²(1) = 4722.1, p < .001] and a model omitting only ESQ [χ²(1) = 482.4, p < .001]. Notably, and consistent with past work (Trott et al., 2023), the presence or absence of a modal was one of the strongest predictors of a participant’s interpretation: 89.1% of trials with a modal were interpreted as requests (vs. 40% without).

We then asked about the relationship between each acoustic feature and pragmatic interpretations. To test this, we built a model predicting Interpretation, containing fixed effects of each of the acoustic features in question (mean F0, F0 slope, and F0 Duration), as well as an interaction between each acoustic feature and Modal, and an interaction between each acoustic feature and ESQ. The model also contained random intercepts for Participant, Item, and Speaker. This model mirrored the model used in the production study, in that the goal was to ask: (i) whether interpretations varied by the presence of certain features (i.e., a main effect of each acoustic feature), and (ii) whether the relationship between a given acoustic feature and interpretations varied according to constructional features. We answered (ii) by comparing the full model to a model omitting only each interaction in question. We addressed (i) by comparing a model including the main effect of a given feature (as well as all other acoustic features and their interactions with modal and ESQ) to a model omitting that feature entirely. There were a total of nine likelihood ratio tests (six unique interaction terms, three main effects); we adjusted for multiple comparisons using Holm-Bonferroni corrections across all nine tests.

After adjusting for multiple comparisons, three of the six interactions tested remained significant: Mean F0:Modal [χ²(1) = 47.7, p < .001], Mean F0:ESQ [χ²(1) = 16.88, p < .001], and F0 Duration:ESQ [χ²(1) = 65.58, p < .001]. For items without ESQ, a higher Mean F0 was associated with a lower likelihood of a request interpretation [B = –0.27, SE = 0.07]; for items without a modal, a higher mean F0 was associated with a higher likelihood of a request interpretation [B = 0.48, SE = 0.07]. Finally, longer utterances were associated with a particularly reduced likelihood of request interpretations when they did not contain an ESQ [B = –0.59, SE = 0.07].

Further, two of the three main effects remained significant: F0 Slope [χ²(1) = 71.1, p < .001] and F0 Duration [χ²(1) = 16.98, p < .001]. To analyze the direction of these effects, we considered the coefficients from a model with just the main effects (as well as main effects of Modal and ESQ).¹⁰ Items with a more positive F0 Slope were less likely to be interpreted as requests overall [B = –0.14, SE = 0.04], and items with a longer duration were also less likely to be interpreted as requests overall [B = –0.17, SE = 0.04]. Differences in F0 slope are depicted in Figure 5. Both results were consistent with past work on the prosody of interrogative requests (Ruytenbeek et al., 2023; Trott et al., 2023), which found that both duration and F0 slope were negatively associated with request interpretations for yes/no interrogatives.

Figure 5: Mean z-scored F0 slope for utterances interpreted as either literal questions or requests. There were substantive main effects of both the presence of a modal and est-ce que. Beyond these effects, items interpreted as requests had a reduced F0 slope.

4.1 Limitations and future work

One limitation of the current work is that the analyses focused on coarse-grained acoustic features measured across the entirety of an utterance (e.g., Mean F0). It is plausible that a finer-grained analysis would reveal additional differences or nuances, perhaps with respect to the est-ce que phrase (cf. Druetta, 2002, 2003): for instance, past work on modal interrogatives (Trott et al., 2023) revealed increased duration of the modal verb (‘can’) when an utterance was intended as a yes/no question. Future work would also benefit from the use of established prosodic analysis frameworks, such as ToBI (Beckman et al., 2005). In addition to identifying subtle prosodic differences, such analyses would be critical for building a more precise mechanistic model of how listeners integrate prosodic cues in the act of language comprehension. Researchers might also make use of the growing body of machine learning toolkits and audio-language models (Baevski et al., 2020) to extract informative cues in a more “bottom-up” fashion.

Another limitation is that the elicitation of stimuli was not done in a naturalistic setting: although speakers were naïve and not given instructions about how to produce an utterance, they were explicitly told to produce an utterance as a request or a question. This raises the question of whether the same cues (or any cues at all) would be produced in more naturalistic valid settings – and, if such cues are produced, whether listeners benefit from them in similar ways. Work on prosody (and language production and comprehension more generally) necessarily involves trade-offs between ecological validity and experimental control. In our case, we emphasized control over the specific utterances produced and the context in which they were interpreted. Future work could explore more contextualized designs, ranging from embedding the existing recordings in a discourse context to analyzing the prosody of utterances produced “in the wild”.

Parametric manipulation of the context in which utterances are perceived also opens the door to additional theoretical questions. In many cases, listeners might need to navigate a variety of potential cues to an utterance’s meaning, such as constructional features (e.g., the presence of a modal), gesture (Kelly, 2001; Kelly et al., 1999; Kirk et al., 2011), or a speaker’s knowledge state (Trott and Bergen, 2020). These cues may not always point in the same direction. Which factors influence the privileging of certain cues over others, as well as the timing with which they are integrated into a model of an utterance’s meaning (Deliens et al., 2017)? Moreover, requests are, of course, a social action: beyond conveying a particular propositional meaning and illocutionary force, indirect requests likely serve other purposes as well, such as saving face (Brown and Levinson, 1987) or avoiding impoliteness (Ruytenbeek, 2019, 2020). Researchers could investigate how the prosodic features identified here influence the perceived (im)politeness of potential indirect requests, and how these judgments interact with situational features, such as the power differential between interlocutors (Brown and Levinson, 1987; Vergis et al., 2020) or a speaker’s entitlement to make the request (Craven and Potter, 2010; Curl and Drew, 2008; Trott and Rossano, 2020).