1.1. Function of disambiguating prosody: For the speaker or for the interlocutor

The function of (disambiguating) prosody concerns, in short, the question whether prosody is produced mainly for the interlocutors or for the speakers themselves. This goes in line with the question in how far the prosodic realization of an utterance is dependent or independent from the actual situation in which it is being produced. If there is a rather direct link between syntax and prosody, disambiguating prosody should be ‘automatically’ present in any case–independent of the situation. However, if prosody is less automatically connected to the structural properties of the utterance, but used in a more controlled way by the speaker to support the interlocutor’s parsing of an ambiguous utterance, then the use of prosody may vary more depending on the situation and/or properties of the interlocutor. The latter assumption can be subsumed under models of ‘situational dependence’ and the former under models of ‘situational independence.’

Situationally dependent models, on the one hand, assume that prosodic realizations depend on the actual communicative situation. Prosodic cues are only necessary, and therefore expected, if the speaker is aware of the ambiguity and the possible misunderstanding of the interlocutor and if the context does not provide other, non-prosodic disambiguating cues. Those other cues can be linguistic or non-linguistic. Models assuming situational dependence of prosodic realizations predict that speakers use prosody differently when addressing interlocutors with different needs or, more generally, that speakers use prosody differently in different communicative or contextual situations. Situational dependence supports the view that prosody is realized for the interlocutor–to help them derive the intended meaning. In that sense, the speech planning mechanism would be required to foreshadow for any stage of the upcoming speech whether it is in fact ambiguous and lacks disambiguating cues of any kind in order to evaluate the necessity for disambiguation (Speer et al., 2011, p. 87f.). Furthermore, in a strict interpretation of context dependence of prosodic cues, their occurrence would then be more likely in situations which do not provide any disambiguating information and, thus, they should appear rather inconsistent and infrequent (Speer et al., 2011, p. 36f.). This inconsistency, however, would render them unreliable for perception (see e.g., Kraljic & Brennan, 2005, p. 196).

Situationally independent models, on the other hand, assume that prosodic realizations are largely independent from actual interlocutors or the communicative/contextual situation. Under such accounts prosodic cues are produced automatically and their realization is affected by grammatical factors such as phrase structure, information status, or phonological length (Kraljic & Brennan, 2005; Speer et al., 2011, p. 37). In this view, prosody is not primarily realized for the interlocutor, but more automatically ‘for’ the speaker. Since prosodic cues are interpreted as depending on linguistic factors, their occurrence should be rather common and frequent, which would make them reliable for perception (Kraljic & Brennan, 2005; Speer et al., 2011). In the following we introduce some exemplar studies which support either the situationally dependent or the situationally independent account.

Allbritton, McKoon, and Ratcliff (1996) addressed the issue of situational in/dependence by testing whether untrained, naïve speakers (versus trained speakers) would spontaneously use prosody to resolve syntactic ambiguities in various kinds of sentence types. The speakers were instructed to read aloud “as if you were telling someone a story that you wanted them to understand” (Allbritton et al., 1996, p. 716). It turned out that most naïve and trained speakers did not prosodically disambiguate most of the sentences. Only if the instruction made them aware of the ambiguity and asked them explicitly to produce two different versions, trained speakers used prosody for disambiguation. This can be interpreted as a finding supporting situational dependence. The authors concluded that either the role of prosodic cues for conveying the underlying syntactic structure is limited or laboratory recordings cannot be generalized to real-world settings (Allbritton et al., 1996, p. 732). Applying a more real-world setting, namely a game-like interactive referential communication task, Snedeker and Trueswell (2003) confirmed the hypothesis that the relation between syntax and prosody is mediated by the context. In their study, naïve participants produced clear prosodic groupings of attachment ambiguities (“Tap the frog with the flower”) only in situations in which the context did not provide sufficient information to situationally disambiguate the two possible meanings. The authors concluded that “speakers produce [prosodic cues] primarily when they appear to be necessary for clear communication” (Snedeker & Trueswell, 2003, p. 128). Based on these two example studies, one could conclude that the use of prosody for disambiguation depends on the awareness of the speaker about the ambiguities and/or on whether the actual context made both readings plausible. Prosody is thus mainly used for the interlocutor (cf. audience design, Bell, 1984).

In contrast, others found evidence in favour of situational independence: Using a co-operative interactive game-board task, similar to Snedeker and Trueswell (2003), Schafer, Speer, Warren, and White (2000) and Speer et al. (2011) found no evidence for a dependency of prosodic cues on situational disambiguation or discourse factors using global attachment ambiguities and temporal closure ambiguities. The speakers produced prosodic cues independent of the communicative situation (even in only locally ambiguous sentences), but still with some flexibility or variability in the choice of cues (Speer et al., 2011). This was also confirmed in another interactive game-like study design by Kraljic and Brennan (2005), who also found overall limited effects of the context. In their interactive setting involving attachment ambiguities (“Put the dog in the basket on the star”), speakers produced clear prosodic cues for disambiguation, irrespective of the needs of their interlocutors (i.e., regardless of whether the contextual setting provided disambiguating information or not) and irrespective of whether an interlocutor was present or not. With respect to the function of prosody, they conclude that prosodic marking emerges from the level of planning and articulation, that is, prosody is not produced dependent on the situation but rather automatically and situationally independent. Similarly, for coordinate name sequences, Wagner (2005) found that prosodic boundaries are produced independent of the context and independent of the need of the interlocutors to comprehend, which implicates that prosody is mainly used ‘for’ the speakers themselves, in an automatic manner.

In sum, there is evidence supporting either of the two accounts on the function of prosody. The differential findings might be related to task differences (e.g., instruction, presence of an interlocutor, degree of interaction between speaker and interlocutor, potential for misunderstandings, awareness of the ambiguities) or the complexity or length of the to-be-produced structures (e.g., longer utterances in Speer et al., 2011 than in Snedeker & Trueswell, 2003). For a detailed discussion on these differences see Kraljic and Brennan (2005), Snedeker and Trueswell (2003), and Speer et al. (2011). For the option of intermediate positions between situational dependence and independence see also Speer et al. (2011, p. 37f.).

1.2. Individual variability in prosody production

We now turn from the group level to the individual level and to the question of whether individuals vary in their prosodic realizations. Variability between or within speakers is interesting for two reasons. First, if all speakers reliably use disambiguating prosody to distinguish between coordinates without and with grouping (example [2] versus [3]), this would indicate a close link between syntax and prosody (e.g., Nespor & Vogel, 1986) and situational independence (e.g., Speer et al., 2011). If, on top of the disambiguation, we would also find variability across speakers in how they realize prosodic boundaries, this would add evidence that the link between syntax and phonology is relational rather than categorical (e.g., Clifton, Carlson, & Frazier, 2002; Wagner, 2005). Second, if speakers do not reliably disambiguate the different syntactic structures (example [2] versus [3]) or show within-speaker variability in different contextual settings, this would speak in favour of situationally dependent models–and against a close prosody-syntax link. So far, the issue of inter-individual variability concerning prosodic boundary production in coordinates has been explored only scarcely (one exception, for German, being the work by Petrone et al., 2017, or, for English, the findings by Allbritton et al., 1996, and Lehiste, 1973), and variability induced by different situational contexts has, to the best of our knowledge, not been studied yet.

Regarding variability between speakers (i.e., inter-individual variability), Petrone et al. (2017) found that their speakers differed in how the prosodic boundary was realized in coordinates with internal grouping (i.e., they found multiple types of prosodic boundaries): Only two out of 12 speakers consistently used the same f0-contour, namely a rise. Although production of a rise was also the predominant contour in six further participants, these additionally employed a high plateau. Another three speakers varied between rise, high plateau, and final fall to different degrees and one speaker produced either rises or falls. Using similar three-name sequences, Lehiste (1973) reported that two (English) speakers differed in how they used durational cues for disambiguation (insertion of a pause versus lengthening of the coordinating element).

With respect to variability within speakers (i.e., intra-individual variability) induced by contextual settings, specifically concerning the type of interlocutor, previous research has focused on differences in prosodic realizations when children, elderly adults, or non-native speakers are being addressed in comparison to young adult native speakers. Most studies take the speech addressed to an adult native speaker of the language under investigation as a baseline for comparisons. For easier reading, we will refrain from mentioning this adult baseline in the following. For example, for attachment disambiguation, Kempe et al. (2010) reported lengthened vowels when English-speaking adults addressed two–four-year-old real or imaginary children and, in addition, found longer pause durations. Other studies investigated intra-individual variability in prosodic information per se (i.e., not focusing on disambiguating prosody): Biersack, Kempe, and Knapton (2005) reported an increased pitch range and higher f0-maxima as well as longer durations due to the lengthening of vowels in semi-spontaneous speech addressed to a two-year-old imaginary child in English. DePaulo and Coleman (1986) also reported longer pauses in spontaneous English speech addressing a six-year-old child. When it comes to prosodic cues in speech addressing a non-native interlocutor, results are inconclusive: While one study involving English speakers found no differences (DePaulo & Coleman, 1986), another one found a lowered speech rate due to lengthened pauses (Biersack et al., 2005), and Smith (2007) reported an increased f0-range and segmental modifications leading to a more emphatic style in French. Regarding prosodic cues when addressing elderly interlocutors in English, Kemper, Vandeputte, Rice, Cheung, and Gubarchuk (1995) reported a slower speech rate due to prolonged vowels and more frequent pauses in spontaneous speech of a map task with a physically present interlocutor. Although expected, they did not find exaggerated pitch ranges. For German, Thimm, Rademacher, and Kruse (1998) also reported more pauses as well as more variation in intonation in spoken explanations of an alarm clock when a positively stereotyped elderly person was addressed as opposed to a young adult.

As an alternative to the experimental manipulation of type of the (imaginary or real) interlocutor, some studies varied the contextual setting via the presence or absence of noise. Speech production in noisy environments leads to increased f0-values and f0-range, increased signal amplitude, increased word or segment durations, and spectral changes such as smaller spectral slope (Davis, Kim, Grauwinkel, & Mixdorff, 2006; Folk & Schiel, 2011; Garnier, Bailly, Dohen, Welby, & Lœvenbruck, 2006; Jessen, Köster, & Gfroerer, 2003; Junqua, 1993, 1996; Landgraf, Schmidt, Köhler-Kaeß, Niebuhr, & John, 2017; Lu & Cooke, 2008; Summers, Pisoni, Bernacki, Pedlow, & Stokes, 1988; Varadarajan & Hansen, 2006; Zollinger & Brumm, 2011). These noise-dependent changes are summarized under the term Lombard speech, tracing back to Étienne Lombard who first described the noise-dependent increase in speech amplitude for French (Lombard, 1911; as cited in Zollinger & Brumm, 2011). Lombard speech is also described as a source of inter- and intra-speaker variability (Jessen et al., 2003; Junqua, 1993; Stanton, Jamieson, & Allen, 1988). For a recent review on the neural mechanisms of the Lombard effect in humans and animals see (Luo, Hage, & Moss, 2018).

1.3. Prosody of coordinates (in German)

As our study specifically investigates the prosody of coordinates, we will briefly review some relevant models on prosodic phrasing in coordinates which have been proposed in the past. We will focus on the Proximity/Similarity model (Kentner & Féry, 2013) since it has been tested with German speakers in similar structures as we use in the current study.

With respect to the question how coordinates are prosodically phrased in English, Taglicht (1998) formulated the ‘Coordination Constraint.’ It specifies that the same hierarchical level of intonational boundaries must be applied to all elements at the same syntactic level. Watson and Gibson (2004) argue in their ‘Left hand side/Right hand side Boundary hypothesis’ that the likelihood of the presence of a prosodic boundary depends on the size of the preceding and following constituents because of processing demands: For larger constituents, the speaker needs more refractory time to recover from the preceding constituent and more time to plan the upcoming constituent, respectively. Wagner (2005, 2010) demonstrates that, in coordinate structures, the relative strength of the boundary reflects the level of embedding at the syntactic level and thereby confirms the close match between prosody and syntax in coordinate structures. According to Wagner (2010, p. 186) more deeply embedded constituents “are separated from each other by weaker boundaries than constituents that are less deeply embedded” and “constituents separated by relatively weaker boundaries are perceived as grouping together.” In a similar vein, Kentner and Féry (2013) developed a model that aims to account for both, processing demands, depending on the constituent size or complexity, and demands of the syntactic structure, depending on the depth of syntactic embedding. Their so-called Proximity/Similarity model assumes two principles that “interact to shape the prosody of syntactic structures” (Kentner & Féry, 2013, p. 283) such as coordinated name sequences. Proximity is related to the syntactic constituent structure and states that “adjacent elements which are syntactically grouped together into one constituent should be realized in close proximity” (Kentner & Féry, 2013, p. 282). Similarity is related to the depth of syntactic embedding and refers to the idea that “constituents at the same level of embedding should be realized in a similar way, that is, they should be similar in pitch and duration, irrespective of their inherent complexity”; this principle is comparable to the models of Taglicht (1998) and Wagner (2005, 2010) which assume that elements at the same syntactic level are prosodically matched.

For the structures used in the present study ([2] and [3]), the Proximity/Similarity model makes the following predictions: In coordinates with internal grouping, the principle of Proximity predicts a weakening of the prosodic cues–compared to a coordinate without internal grouping–at an element x if the neighbouring element to the right is part of the same group as x (cf. the first name, i.e., Moni, in [3]). Reversely, Anti-Proximity predicts a strengthening of a prosodic boundary if the right-adjacent element of x does not form a group with x (cf. the second name, i.e., Lilli, in [3]). The Similarity principle predicts that a simplex element at the same level of embedding as a complex constituent is, for instance, lengthened to adjust its duration to the length of a complex constituent. This would be relevant for groupings in which the first name is followed by a complex sequence on the right (cf. Moni und [Lilli und Manu]). As the current study will focus on coordinates with an internal grouping of the first two names (as in [3]), the Similarity principle will not be discussed any further. In coordinates without internal grouping (considered the baseline form; see [2]), all names are expected to be separated by boundaries of the same strength.

With boundary cue weakening, Kentner and Féry (2013) refer to the use of lower pitch and shorter durations on the first grouped element compared to a non-grouped baseline, while a strengthened boundary at the right edge of the grouped element is expressed by a higher boundary tone and longer durations. On the final element of the coordinates (cf. Manu in [2] and [3]), Kentner and Féry (2013) observed neutralization in duration and f0-movement. The findings of increased duration of the word preceding the boundary and a possible pause along with higher pitch at the prosodic boundary have been confirmed in a further study on elicited coordinate productions in German (Petrone et al., 2017) and are in line with results on prosodic marking of syntactic boundaries in spontaneous German speech (Peters et al., 2005). What is still unclear with respect to the Proximity/Similarity account is whether its assumptions also hold situationally independent. Until now, variations in prosodic phrasing of coordinates within speakers across different situations/interlocutors have not been explored sufficiently. In addition, it is unclear to which extent there is variability across speakers and, specifically, whether the speakers differ in how they use and combine the different prosodic cues to mark the prosodic boundaries.

Therefore, in our study, we use coordinate name sequences ([2] and [3]) to replicate the findings of Kentner and Féry (2013) under different contextual settings, that is, in different situations. At the same time, due to the focus on (2) and (3), our data will not be sufficient to adjudicate among the models briefly introduced above in this section. Thus, the current study will not directly contribute to the question as to whether–or to which extent–processing demands or the level of syntactic embedding drive prosodic realizations. Instead, the main focus of our study is on inter- and intra-individual variability and its limits in prosodic boundary production, the relation of the different prosodic boundary cues to one another, and on the situational in/dependence of prosodic phrasing.

In summary, the functional role of disambiguating prosody or its situational in/dependence has been studied by means of the presence or absence of contextual effects on prosodic realizations–but remains largely inconclusive. The fact that participants are aware of an ambiguity, the task setting (reading-out loud versus interactive setting with real versus imagined interlocutors), the type of ambiguity (e.g., attachment ambiguities versus pragmatic ambiguities), the length of the to-be-produced utterance, the type of interlocutor (e.g., child versus adult), and other contextual factors, such as absence/presence of noise, seem to influence if and how individuals use prosody to disambiguate syntactic structures. We are going to address the question of situational in/dependence by comparing prosodic realizations in varying situations within individuals. In addition, we will explore differences between speakers as they will give us further insights into the nature of the prosody-syntax link.

1.4. Aims and hypotheses

In this study we systematically explore inter- and intra-individual variability in the production of prosodic boundaries to get insights into the prosody-syntax relation and the function of prosody. According to situationally dependent models of prosodic phrasing (Allbritton et al., 1996; Snedeker & Trueswell, 2003; cf. audience design hypothesis, Bell, 1984) we would predict that, if speakers use prosody to disambiguate different syntactic structures at all (e.g., because they are aware of the ambiguity and/or because an interlocutor is present in the communicative situation), they vary considerably in their prosodic productions between interlocutors with different needs. Contrary, according to situationally independent models of prosodic phrasing (Kraljic & Brennan, 2005; Schafer et al., 2000; Speer et al., 2011), we would predict that speakers use prosody to disambiguate different syntactic structures in any event, because they are doing it ‘automatically’ during speech planning stages. The prosodic realizations should hence be rather clear between conditions without and with internal grouping (example [2] versus [3]), and consistent across different interlocutors–although some variability between speakers is also expected (Speer et al., 2011, p. 88ff.).

We argue that the issues of within-speaker situational in/dependence and of between-speaker in/variability are related to the underlying nature of the prosody-syntax link: If there is a fixed relationship (and dependency) between prosody and syntax, we would predict that speakers ‘automatically’ produce prosodic boundaries in a rather fixed or stable manner to disambiguate the syntactic structure, irrespective of the situation they are confronted with (i.e., situationally independent). If, at the same time, the relationship between syntax and phonology is relational or gradient (e.g., Wagner, 2005), we would additionally predict some variability between speakers with respect to the phonetic correlates they employ to disambiguate the syntactic structure.

Our study thus explores the effect of the type of the interlocutor and presence/absence of noise on variability between and within speakers’ prosodic boundary realizations in a controlled, semi-interactive setting. Specifically, the speakers are asked to utter coordinates with versus without internal grouping (such as [3] and [2]). The five different contextual settings will henceforth be referred to as contexts. The contexts involve four different female interlocutors: a young adult (YOUNG), a child (CHILD), an elderly adult (ELDERLY), and a young non-native adult speaker of German (NON-NATIVE) and a noisy environment (the young adult with white background noise, NOISE).

Speakers are completely aware of the intended syntactic grouping of the coordinates and are asked to utter the name sequences in such a way that the different virtual interlocutors can resolve them. We will focus on the prosodic cues f0-range, final lengthening, and pause at/after the first and the second name, as these are known to be modulated to indicate prosodic boundaries. The results will be discussed referring to the Proximity/Similarity model of syntax-prosody mapping (Kentner & Féry, 2013). We will describe the interplay and combined use of the prosodic cues of prosodic boundaries and how these are affected by inter- and intra-individual variability as these will contribute to our understanding of the prosody-syntax relation and the functional role of disambiguating prosody.

Our research questions are as follows:

Regarding Q1, based on the literature outlined in 1.3, we expect speakers to mark the difference between coordinates with (3) and without (2) internal grouping in line with the Proximity/Similarity model by Kentner and Féry (2013). More specifically, we expect a prosodic boundary realized by an increase of final lengthening and an increased f0-range at the right edge of the group (i.e., on the second name), as well as the insertion of a pause after the grouping in (3) compared to (2). On the first name, we expect a decrease in final lengthening and a smaller f0-range in (3) compared to (2).

With respect to Q2, we confront speakers with five different contexts (YOUNG [baseline], CHILD, ELDERLY, NON-NATIVE, and NOISE) to disentangle the question of situational in/dependence of prosodic variability. If speakers vary their productions between contexts, we expect those variations to be in line with the literature mentioned in 1.2: We expect speakers to mark the difference between conditions with and without internal grouping in a more pronounced way in the non-baseline contexts. If the findings of prosodic cue modifications for contextual situations in different sentence types are transferable to coordinates and if the modifications found for English and French speakers also hold for German, we expect an increase in segmental and pause durations as well as increased f0-ranges for CHILD and ELDERLY. Due to inconsistent findings in previous studies regarding a non-native interlocutor, we explore this context at a rather exploratory level and refrain from a specific hypothesis. Regarding the presence of noise (NOISE), the literature predicts an increase in f0 and segment durations. Note that the NOISE condition is the only condition, which directly affects the speaker, because the virtual interlocutor and the speaker are confronted with the noise.

With Q3 and Q4, we expect to further disentangle the nature of the prosody-syntax link. Regarding Q3, between-speaker in/variability in the combined use, that is, in the interplay of the prosodic cues, will inform us about the type of link between syntax and phonology (i.e., fixed and categorical or relative and allowing for some flexibility). Regarding Q4, within-speaker in/variability will give further insights into situational in/dependence of prosodic cues on the individual level. These two research questions will be addressed in an exploratory manner.

Overall, marked differences between contexts would speak in favour of situationally dependent models and their absence for models of situational independence (given speakers would prosodically disambiguate the conditions with and without grouping). If speakers show inter-individual variability in how they employ and combine prosodic cues at the surface to mark prosodic boundaries, this would speak in favour of models of relative boundary strength (e.g., Wagner, 2005).

3.1. Statistical analyses of prosodic disambiguation and general context-dependent variability of individual prosodic cues on Name1

3.2. Statistical Analyses of prosodic disambiguation and general context-dependent variability of individual prosodic cues on Name2

3.2.1. Rise on Name2

For rise, the estimates for the fixed-effects were extracted from the maximal model (Table 7). We found a main effect of condition and a main effect of the contexts CHILD and ELDERLY. Regarding prosodic disambiguation, speakers, overall, produced a larger f0-range in the brack than in the nobrack condition. Regarding contexts, when addressing the child as well as the elderly person, speakers increased the f0-range of the rise compared to addressing the young adult (cf. Figure 4 and left panel of Figure 5). For a subset of 13 female speakers this increased f0-range can be seen in Figure 4, where the green and blue (CHILD and ELDERLY context, respectively) dashed and solid lines start below the black lines (YOUNG context) at the beginning of Name2 and rise to a level above the black lines towards the f0-peak of Name2. Note, Figure 4 cannot be compared directly to the results of the statistical model, since values in Hertz are plotted in the figure, while the model is calculated on semitones and the figure contains only data of a subset of the speakers. A similar, though statistically non-significant tendency is observable for the other two contexts, addressing the non-native speaker and in noise. The model revealed no statistically significant interactions between contexts and condition.

Table 7

Estimates of the model for rise on Name2 (i.e., f0-range on Name2). Statistically significant effects are marked in bold (p < .05).

Predictor	Estimate	SE	t-value	p-value
condition_brack	2.86	0.27	10.77	<.001
CHILD vs. YOUNG	1.10	0.31	3.56	.003
ELDERLY vs. YOUNG	0.94	0.36	2.59	.021
NON-NATIVE vs. YOUNG	0.59	0.31	1.89	.078
NOISE vs. YOUNG	0.53	0.30	1.76	.097
condition_brack: CHILD vs. YOUNG	0.03	0.20	0.16	.874
condition_brack: ELDERLY vs. YOUNG	0.09	0.22	0.40	.694
condition_brack: NON-NATIVE vs. YOUNG	0.04	0.24	0.18	.856
condition_brack: NOISE vs. YOUNG	–0.09	0.26	–0.33	.746

Figure 4

Time-normalized f0-contours (in Hz) of coordinates in brack (solid lines) and nobrack (dashed lines) conditions produced in five contexts (cf. colours) by a subset of 13 female speakers.

3.2.2. Final lengthening on Name2

For final lengthening, the estimates for the fixed-effects were extracted from a model that included principal components but not the correlation parameters in the random-effects structure (Table 8). Regarding prosodic disambiguation, the data show a main effect of condition, indicating that speakers marked the brack condition with increased final lengthening compared to the nobrack condition. Regarding general context variability, there was an interaction between condition and the ELDERLY context with a negative estimate and an interaction between condition and the NOISE context with a positive estimate (cf. mid panel Figure 5). This indicates that speakers increased the final lengthening when addressing the elderly as opposed to the young interlocutor in the nobrack condition. In the noise context, however, they increased final lengthening in the brack condition, but not in the nobrack condition.

Table 8

Estimates of the model for final lengthening on Name2. Statistically significant effects are marked in bold (p < .05).

Predictor	Estimate	SE	t-value	p-value
condition_brack	4.96	0.58	8.59	<.001
CHILD vs. YOUNG	–0.39	0.58	–0.67	.511
ELDERLY vs. YOUNG	1.34	0.62	2.17	.039
NON-NATIVE vs. YOUNG	0.61	0.57	1.07	.293
NOISE vs. YOUNG	1.45	0.71	2.04	.056
condition_brack: CHILD vs. YOUNG	–0.56	0.51	–1.09	.274
condition_brack: ELDERLY vs. YOUNG	–1.24	0.51	–2.43	.015
condition_brack: NON-NATIVE vs. YOUNG	–0.65	0.51	–1.27	.205
condition_brack: NOISE vs. YOUNG	1.43	0.52	2.75	.006

Figure 5

Mean values and 95% confidence intervals for rise (left panel), final lengthening (mid panel), and pause (right panel) on Name2 for each context and condition (green = condition brack, grey = condition nobrack).

3.3. Exploratory analyses of inter- and intra-speaker variability of cue combinations on Name2

The cue combinations for each speaker (cf., y-axis) and context (cf., x-axis) are plotted in Figure 6. For each speaker and context, the cell is divided into three rows, with the distinction type of rise given in the uppermost row of the cell, final lengthening in the middle row, and pause in the bottom row. The three types of distinction are represented by shading: full colour for clear distinction, light shade for partial distinction, and the lightest shade for no distinction.

Figure 6

Speaker variability of cue combinations across contexts on Name2, showing the patterns of cue combinations (shades of green) used by individual speakers (y-axis) in the contexts YOUNG, CHILD, ELDERLY, NON-NATIVE, and NOISE (x-axis). The shades of green indicate the type of distinction: full = clear distinction (C), light = partial distinction (P), lightest = no distinction (N). For each speaker, the three rows indicate the different cues (R: rise, L: final lengthening, P: pause). The small numbers in italics indicate the mean ratios of the hit-rates for condition in the perception check (i.e., ratio of correct identifications of the intended grouping to all rates; numbers to the left: average per speaker; lowest line: average per context; above cells: average per speaker per context). For example, speaker 16 clearly distinguishes between the brack and nobrack condition, using all three cues in the YOUNG, CHILD, and NON-NATIVE context, but in the context NOISE the speaker uses final lengthening only partially and pause not at all to distinguish between the two conditions. In the YOUNG context, 100% of the rates in the perception check were congruent rates, while in the NOISE context 94% were congruent rates.

Regarding inter-speaker variability, we focused on whether there are different patterns of cue combinations between speakers within the young context only, represented in the left-most column of the plot in Figure 6. In general, five different patterns are observable: CCC_RLP (i.e., all three cues in full green), CNC_RLP, CPC_RLP, CCP_RLP, and PPC_RLP, however they differ in number of occurrences. Seven speakers out of 15 (2, 3, 4, 7, 13, 15, and 16) produced the pattern CCC_RLP, indicating that all three cues were clearly used to distinguish between the two conditions. A further four speakers (1, 6, 9, and 11) produced the pattern CNC_RLP, which means that they clearly distinguished between brack and nobrack using rise and pause, but not using final lengthening. The other three patterns were produced by either two or one speakers. Overall, in four of the five patterns brack and nobrack were clearly distinguished by at least two of the cues. While both rise and pause were used clearly distinctively by 14 out of 15 speakers, only eight speakers used final lengthening in a clearly distinctive way. Notably, the pattern with no distinction in all three cues was never observed in the young context.

Regarding intra-speaker variability, we focused on whether speakers vary the patterns of cue combinations when addressing different interlocutors or speaking in noise. For that purpose, we examined the patterns of cue combinations within speaker across contexts (i.e., the three rows in each cell for each speaker across columns in Figure 6). For speaker 2, the pattern is identical across all five contexts, thus showing stability in the use of the prosodic cues for distinguishing between the brack and nobrack condition across different contexts. Most other speakers show two or three different patterns across contexts (cf., speakers 6, 7, 8, 9, 11, 14 and 3, 4, 5, 10, 13, 15, 16, respectively). Overall, there is more variability across contexts in the use of final lengthening than in rise and pause, visualized by more varying shading in the middle row of the speaker-specific cells. There is one speaker who in one context shows no distinction between brack and nobrack in any of the three cues (cf., NNN_RLP speaker 1, context non-native); for all other speakers and contexts at least one cue is clearly distinctive. In addition, we plotted the mean ratio of the hit-rates for condition in the perception check in Figure 6 for the respective speakers and contexts. This allows us to get an impression of the relation of the produced types of distinction in the three prosodic cues to how well the prosodic boundaries (i.e., conditions) have been perceived by naïve listeners.

4.1. Prosodic disambiguation of coordinates (research question Q1)

Our findings replicate previous studies, showing that the internal grouping of coordinates in German is marked by a prosodic boundary consisting of three prosodic cues from the tonal and durational domain: f0-range, final lengthening, and pause. As expected, speakers used prosodic cues on Name1 as well as on Name2 to clearly distinguish between the two conditions. A perception check with naïve listeners showed that the distinction between the conditions was perceptually recoverable: 96% of the productions were correctly recognized as the intended grouping.

The results of the production study are in line with the Proximity and Anti-Proximity principles that form part of the Proximity/Similarity model introduced by Kentner and Féry (2013) and along with this, they are in line with the literature (Taglicht, 1998; Wagner, 2005, 2010; Watson & Gibson, 2004). Thus, our hypothesis (Q1) was confirmed: In the condition with internal grouping compared to the condition without grouping, we found a statistically significant decrease in final lengthening and f0-range on Name1 along with an increase in final lengthening and f0-range on Name2 as well as the insertion of a pause after Name2. In terms of Proximity, durational and tonal cues of Name1 were decreased, indicating that the neighbouring element to the right (i.e., Name2) forms part of the same group. In terms of Anti-Proximity, the prosodic boundary after Name2 was strengthened, indicating that the neighbouring element to the right (i.e., Name3) does not form a group with Name2. This finding also underlines the assumption that prosodic phrasing is not a local phenomenon with changes of prosodic cues occurring only at the prosodic boundary (cf. in our case Name2) but rather depends on globally distributed prosodic changes (cf. in our case Name1 and Name2) (e.g., Clifton et al., 2002; Frazier et al., 2006; Wagner, 2005, 2010).

We further found that speakers use the pause cue in a slightly different way than f0-range and final lengthening in marking the difference between conditions. Following Name2, a pause was mostly absent in the condition without internal grouping, while it was present in the condition with grouping. The pause, thus, appears rather as a categorical than a continuous variable. Since we were interested in differences in pause duration between contexts, however, we kept pause as a continuous variable for our analyses.

Overall, the syntactic structure (with or without internal grouping) was clearly disambiguated by means of prosody. This can be interpreted as evidence in favour of a close link between syntax and prosody.

4.2. General context-dependent prosodic variability (research question Q2)

The current study is, to the best of our knowledge, the first to systematically investigate prosodic variability in production of coordinates across speakers between various contexts to explore the situational in/dependence of disambiguating prosody and to find out whether the principles of Proximity/Anti-Proximity also hold across situations.

At the group level, we found some variability driven by the different contexts. Nevertheless, variability was rather small and not as distinct as expected. In the following, the contexts CHILD, ELDERLY, NON-NATIVE, and NOISE will be discussed individually in comparison to the baseline context (YOUNG).

In the context CHILD, when addressing the child as opposed to the young adult, speakers changed their productions in the tonal domain: They increased the f0-range on Name2 independent of condition. This can be interpreted as an adaptation to the interlocutor, but without affecting the ease of disambiguation between conditions. The increased f0-range when addressing a child is partly in line with semi-spontaneous speech data from English speakers (Biersack et al., 2005), who additionally showed lengthened vowels. These differences might be due to differences in age of the interlocutor. For a child addressee of the same age as in our study, DePaulo and Coleman (1986) reported longer pauses; a finding that was not evident in our data. A possible explanation for the absence of statistically significant effects in the durational prosodic cues (i.e., final lengthening and pause) in our study might be related to differences in speech style as well as in language-specific factors. Our data were highly restricted with respect to the wording, whereas the data of DePaulo and Coleman (1986) consisted of spontaneous speech and the data of Biersack et al. (2005) of semi-spontaneous speech, both in English.

In the context ELDERLY, when addressing the elderly adult compared to the young adult, speakers modified their speech in the tonal as well as in the durational domain. On Name2, speakers produced an overall larger f0-range in the ELDERLY context along with a longer pause (in the condition with internal grouping). In contrast, final lengthening on Name2 was not used to make the conditions more distinct in the ELDERLY context: Unexpectedly, speakers increased the lengthening in the condition without grouping compared to coordinates addressed to the young adult. Yet, with the increased pause duration, the smaller difference in final lengthening between the conditions was probably levelled out. The findings of increased pause durations and increased f0-ranges, thus, partly confirm our hypotheses and are comparable to observations on other structures in English and German (Kemper et al., 1995; Thimm et al., 1998). Those studies found slower speech due to prolonged vowels and more pauses as well as increased variation in intonation, among other speech adaptations. Regarding the increased number of pauses in the reported studies, again, it needs to be mentioned that the respective data stem from spontaneous speech which probably allows for more pause insertion than the relatively restricted stimuli used in our study. Nevertheless, we suggest that the increased pause durations in our data can be interpreted as comparable speech adaptations. In previous research on speech directed at elderly persons, Kemper, Ferrell, Harden, Finter-Urczyk, & Billington (1998) distinguished two sets of parameters that speakers modify in order to adapt to the needs of their elder interlocutor: semantic and discourse information on the one hand, and fluency, prosody, and grammatical complexity on the other. Kemper et al. (1998, p. 53) discuss that the latter set of parameters does not “appear to benefit” perception, but to the contrary, decreases self-esteem on the side of the interlocutor. This type of speech is referred to as patronizing communication (Kemper et al., 1998; Ryan, Hummert, & Boich, 1995; Thimm et al., 1998; Torrey, Fussell, & Kiesler, 2005) and includes the changes in prosodic cues found in our data.

In the context NON-NATIVE, in response to the non-native interlocutor, the data show no clear effects. This contrasts with reports in the literature, in which non-native speakers were addressed with increased f0-ranges and a more emphatic style compared to native speakers (Smith, 2007).

Finally, in the context NOISE, the interlocutor was the same young adult as in the baseline context. For adaptation to the noise, speakers increased final lengthening on Name2 in the condition with grouping while at the same time, they decreased the relative duration of the following pause. The increase in final lengthening is in line with our hypotheses and findings in the literature, although we would have expected an additional increase in the f0-range. A possible explanation for the unexpected decrease in pause duration is that a silent pause is a less effective cue in a noisy environment than in a quiet one. Instead of a silent pause, speakers lengthened the final segment to mark the boundary. Furthermore, speakers might have tried to fill the noise with their own voice, in order to distract themselves from the noise. Varadarajan and Hansen (2006) interpreted this result as “a sense of urgency on the part of the speaker […] due to persistent exposure of the environmental noise” (Varadarajan & Hansen, 2006, p. 938).

With respect to our research question Q2, we can conclude that we found only some small differences in the three prosodic boundary cues produced in coordinates elicited in different contexts. In addition, the small differences between the contexts could hardly be discriminated on the perceptual side as shown by the weak performance in the perception check regarding the assignment of productions to the differential contexts: Listeners were not able to reliably identify to whom the utterance was addressed.

With regard to the question of situational in/dependence of prosodic disambiguation, the finding of clear production of a prosodic boundary to disambiguate the conditions with/without grouping (Q1) together with the only small contextual adaptations (Q2) in our data, speaks in favour of situational independence. In the context of our study, the prosodic distinction between coordinates with or without internal grouping might have been considered to be more ‘relevant’ than a prosodic adaptation to possibly different needs of the interlocutors.

In the following we will discuss two limitations of our study, before turning to research questions Q3 and Q4:

First, another explanation for the fact that the context effects in our production data were smaller than expected might be based on the somewhat artificial design of the study: The interlocutors were auditorily present before the recording of each stimulus, however, there was no feedback of their perceptual performance. A request for repetition or a misunderstanding may have triggered further accommodations in the speech addressed to the interlocutors. As mentioned above, accommodation to possible needs of an interlocutor can also be interpreted as patronizing by the interlocutor, as Kemper et al. (1998) reported for the speech used by young adults when addressing elderly adults. In our study, speakers either may have perceived no need to adapt any further to their interlocutors or they might have been sensible and avoided an over-exaggerated speech style since no feedback was given. This can apply especially for the elderly adult and the young non-native speaker, as they are both adults. Future studies, nevertheless, might want to include feedback of the interlocutors in order to increase the necessity of speakers to adapt to their interlocutors and to make the interaction more natural.

Second, we focused on three particular prosodic boundary cues and, therefore, cannot disregard the possibility that speakers may have produced additional prosodic cues to adapt to their interlocutors. This could, for instance, apply to the NOISE context: The context NOISE was best identified in the perception check (17 out of the 21 productions that were correctly identified by all listeners had been produced in the NOISE context and a total of 65 productions in NOISE was correctly identified by at least 75% of all listeners). This suggests that speakers used additional (prosodic) cues to adapt to the noise. Other studies looking at speech in noise reported, for instance, increased intensity in the presence of noise, as well as spectral changes (e.g., Davis et al., 2006; Junqua, 1996; Landgraf et al., 2017; Lu & Cooke, 2008; Summers et al., 1988). This could be seen as further evidence that disambiguating prosody is not primarily produced for the interlocutor but automatically produced ‘for’ the speaker during planning and articulation: When speakers are confronted with noise, this might affect the cognitive resources used for the planning and articulation and hence get reflected in their prosodic output. Future studies are needed to test this hypothesis.

In the final two sections, we discuss the results of the exploratory analysis regarding which cue combinations are used by individual speakers to mark the prosodic boundary in the grouped name sequences.

4.3. Inter-speaker variability of cue combinations (research question Q3)

With regard to inter-speaker variability of prosodic cues and cue combinations (i.e., the interplay of prosodic cues) in the YOUNG context, the data show that the majority of the speakers (14 out of 15) employed at least two cues distinctively to mark the prosodic boundary in the condition with grouping on Name2. Furthermore, for 13 speakers these two clearly distinctive cues were rise and pause. To put it simply: The vast majority of speakers clearly used pause and rise on Name2 to distinguish between conditions. In comparison to rise and pause, final lengthening was used more variably in the YOUNG context: Some speakers produced it clearly distinctively, others with partial or no distinction. A post-hoc exploratory visual inspection of the data points that were excluded after the perception check further showed that the ‘clear distinction’-pattern in either of the three prosodic cues was beneficial for perception: Often the perception of the non-intended condition went along with one of the three prosodic cues falling within the range of the values of the perceived condition. In other words, if for instance a grouped item was perceived as having no internal grouping, the value of one of the three prosodic cues was more similar to other items without grouping of that speaker than to grouped items.

Overall, most speakers combined at least two cues to clearly disambiguate the conditions, but still, there is some variability between speakers. This speaks in favour of a close relation of syntax and prosody that nonetheless allows for some flexibility in how prosodic boundaries are phonetically realized at the surface (Wagner, 2005, p. 155). Despite this variability between speakers at the phonetic level, the boundaries are easily and reliably detected by the listeners, as shown by the perception check.

4.4. Intra-speaker variability of cue combinations (research question Q4)

This discussion concerns the question whether individual speakers mark the boundaries on Name2 differently in the five contexts. Mirroring the group analysis (see 4.2.), almost half of the speakers (7 out of 15) were stable across contexts also with regard to the relation between cues, as they used one or two patterns only. A closer look at these speakers revealed that the patterns they used mostly contained alternations in one cue only and were, consequently, quite similar to each other. Again, final lengthening emerges as the cue used least distinctively of the three cues investigated, while rise and pause in most cases clearly distinguish between the two conditions–also across contexts. In conclusion, in terms of cue patterns used, the differences between contexts were quite small and individual speakers rather stuck to their individual ‘prosodic strategy’ of marking the boundaries in the condition with grouping independent of their interlocutor.

Overall, we can summarize that individual speakers showed a limited set of cue patterns with only slight shifts in cue distribution between contexts. Hence, also the analysis of individual speakers in varying contexts is in favour of a relatively limited range of variability or rather stable intra-individual ‘prosodic strategies’ to disambiguate coordinates with versus without internal grouping. This adds to the notion of situational independence of disambiguating prosody that is produced automatically by the speakers in a rather invariant manner.