2.1. Phonological and phonetic interactions between tones and syllables across languages

Restrictions on contour tones are typically modeled in phonological theory as constraints on associations between tonal targets and moras, an abstract unit of timing associated with segments in the rime (Woo, 1969; Hyman, 1989). Contour tones are composed of two tone targets, typically a high and low, where the high precedes the low in a falling tone and follows the low in a rising tone. In contrast, level tones are assumed to consist of a single tone target.

Languages with syllable-based restrictions on contour tones prohibit the association of more than one tone to a single mora. It thus follows that only syllables containing at least two moras may support contour tones in these languages (Figure 1).

Figure 1

Association between tones and moras in languages with contour tone restrictions.

In all languages, phonemic short vowels are monomoraic, while phonemic long vowels are bimoraic. The primary source of cross-linguistic variation lies in the moraic status of consonants. In languages that allow contour tones on syllables closed by a sonorant but not those closed by an obstruent, e.g., Kiowa, only sonorant codas are moraic. In other languages, e.g., Krongo, no coda consonants are moraic; as a result, only syllables containing a long vowel may support a contour tone. In some languages, syllable-based tone restrictions are manifested primarily as static constraints on the distribution of tones, e.g., Cantonese (Bauer & Benedict, 1997), whereas in other languages, e.g., Krongo and Kiowa, active processes also simplify contour tones in syllables that do not support them.

Given their perceptual underpinnings, it is not surprising that the categorical restrictions on tone contours in many languages are mirrored by phonetic effects designed to mitigate the challenges of implementing tonal excursions on syllables that provide suboptimal backdrops for those excursions. Such effects are observed both for lexical tone and for intonational tones, even in languages purported to allow a full range of tone contrasts regardless of syllable type. Gordon (2001a) and Zhang (2002) present several case studies of languages revealing varied strategies to facilitate the accommodation of tones on syllables that are intrinsically deficient in their ability to support contour tones. These measures include, in addition to phonologically deleting a tone, as in Krongo, (1) rescaling of tonal targets (i.e., lowering high targets and/or raising low targets) to minimize the extent of f0 excursions, (2) lengthening the syllable with which tones are associated in order to allow more time for the realization of tonal excursions, and (3) shifting one or more tone targets to adjacent syllables in order to minimize tonal crowding. These strategies also shed light on the few languages reported to allow contour tones on syllables closed by an obstruent but not on open syllables containing a short vowel.

Rescaling is a common phenomenon. In Luganda (Bantu; Uganda), which allows contour tones on syllables ending in an obstruent coda, a falling tone associated with a syllable closed by a voiceless obstruent is realized with less of a drop in f0 than a falling tone on other syllable types carrying contours (Dutcher & Paster, 2008). Hausa (Afro-Asiatic; Niger, Nigeria), another language tolerating contour tones on obstruent-closed syllables, employs lengthening of vowels bearing contour tones in order to mitigate tonal crowding in syllables closed by an obstruent (Gordon, 2001a).

Temporal shifting of tones is another well-documented strategy for reducing tone crowding, including in intonation systems, where it has been claimed to underlie typological biases in stress location attributed to avoidance of tonal crowding between pitch accents and boundary tones (Hyman, 1977; Gordon, 2001b). It can even lead to categorical tonal processes as observed in Mbui (Niger-Congo; Cameroon) phrasal tonology: For example, underlying /lɔ̀ɔ́ + wa/ is realized as [lɔ̀ɔ̀ wá] ‘look for me’, where the underlyingly toneless object pronoun wa receives its surface tone from the second component of the low-high contour on the root (Hyman & Schuh, 1974).

All of these temporal and scaling strategies have the effect of mitigating tonal crowding. The same strategies for reducing crowding of tones through rescaling, lengthening, temporal shift, and categorical deletion are observed in intonation systems, where they are modeled in autosegmental metrical theories of intonation in terms of associations between moras and intonational tones, including boundary tones, phrasal tones, and post-lexical pitch accents (Pierrehumbert, 1980; Beckman & Pierrehumbert, 1986). The parallel between lexical and intonational tones in their response to tonal crowding is not surprising given that the realization of both is subject to the same physical constraints.

The tonal crowding factors at play in the realization of tone sequences within syllables also exert themselves across syllables by reducing the extent of f0 excursions between adjacent syllables. This effect can be seen in processes involving either partial or complete assimilation of adjacent tones. For example, in FeʔFeʔ-Bamileke (Atlantic-Congo; Cameroon) (Hyman & Schuh, 1974), a low tone raises to a raised low preceding a non-low tone: si¹ pʉɐ¹ ‘without a bag’ vs. si² mɔh³ ‘without a fire’ where higher numbers represent higher tones. It may be noted that the distinction between the raised low and low tone is phonemically contrastive in FeʔFeʔ-Bamileke, cf. zɔk¹ ‘knee’ vs. ʧɐk² ‘pot’. Temporal mitigation of tone crowding is also observed in intonation systems. In Chickasaw (Muskogean; Oklahoma) (Gordon, 2008), the final word in a question contains a H* pitch accent followed by a final L% boundary tone. The pitch accent is confined to a three-syllable window at the right edge of a word where the location of the pitch accent is not predictable from stress as word-final syllables carry primary stress in most cases. The H* occurs on a final syllable only if it contains a long vowel, following the common cross-linguistic pattern whereby a falling contour tone is restricted to the most sonorous syllable type (cf. Krongo), e.g., katahˈtãː ʧi.haː.ˈʃaː^H*L% ‘Who are you angry at?’ Otherwise, if the final syllable does not contain a long vowel, H* retracts to a heavy (CVV or CVC) penultimate syllable, e.g., kaˈtaːt ok.ʃit^H*.ˈta^L% ‘Who is closing it?’, or, if the penult is light (CV), to the antepenult, e.g., kaˈtaːt hon^H*.ko.ˈpa^L% ‘Who is stealing it?’ Crucially, not only is the phonological alignment of the pitch accent sensitive to syllable type, but the f0 peak associated with H* is also timed progressively earlier within the syllable, the closer the syllable carrying the pitch accent is to L%, i.e., earliest in ʧi.haː.ˈʃaː^H*L% and latest in hon^H*.ko.ˈpa^L%, a pattern that reflects the pressure to increase the distance between H* and L% both within and across syllables.

These cross-syllable strategies to mitigate tonal crowding indicate that constraints against tautosyllabic contour tones are merely one instantiation of a more general tendency to minimize f0 excursions, a pressure that itself belongs to the broader class of coarticulatory processes affecting tone. Although not universal (see Hyman & Schuh, 1974, and Hyman, 2014, on dissimilatory tone processes), tonal coarticulation is a typologically pervasive phenomenon operative on phonological tone whether, lexical or intonational.

2.2. Tone and syllable structure in Croatian

This paper explores the role of syllable structure in the realization of tone contrasts in the Split variety of Croatian (Rešetar, 1900; Browne & McCawley, 1965; Magner, 1978; Inkelas & Zec, 1988; Smiljanić, 2004; Godjevac, 2005; Kapović, 2015). Croatian belongs to a larger cluster of other national languages that includes Bosnian, Serbian and Montenegrin and is collectively known by the acronym BCMS. Like Standard Croatian, Split Croatian possesses a contrast between a rising and a falling tone on stressed non-final syllables, where the realization of the complete contour often extends into the post-tonic syllable. It may be again noted that in the Croatian linguistic orthography, tone type is crossed with vowel length to yield the four categories of short falling, short rising, long falling, and long rising: pȁra (SF) ‘steam’, pàra (SR) ‘money’, Lȗka (LF) ‘Luke’, lúka (LR) ‘port’. These categories can be labelled “prosodemes,” and we use this terminology throughout the paper. As the examples in Table 1 show, the falling and rising tones are both possible on all syllable types, including syllables containing a long vowel (CVV), as well as those containing a short vowel with a sonorant coda (CVR), an obstruent coda (CVO), or no coda at all (CV).

Croatian is typically regarded as a case of a pitch-accent language in which tone has a smaller functional load relative to prototypical tone languages. As in other pitch accent languages (and unlike canonical tone languages), there is a single accent per prosodic word in Croatian, and the location of the pitch accent is positionally limited to the stressed syllable in a word, though its phonetic exponents may extend beyond the tonic syllable. The vast majority of languages in Zhang’s (2002) survey are canonical tone languages rather than pitch-accent languages. Typological data on syllable-based tone restrictions in pitch-accent languages, in which tone plays a more limited role in signaling distinctions, is sparser, likely due in part to analytic indeterminacy about the phonological treatment of tone in these languages (see Hyman, 2006, on the taxonomy of pitch accent in relation to tone). Because pitch-accent languages typically observe culminativity—the requirement that there be a single syllable per prosodic word that is more prominent than others—they are often amenable to alternative analyses that resemble those used to account for stress. Particularly relevant to the present paper is the possibility of analyzing pitch accents as a single tone target, typically a high tone, rather than a sequence of multiple tone targets. As such, issues of tonal crowding are potentially less acute in pitch-accent languages. Nevertheless, evidence suggests that the same types of syllable-sensitive constraints on tone are operative in pitch-accent languages as in more prototypical tone languages, perhaps because lexical pitch accents may have two (or more) tonal targets in practice, even if a theoretically more parsimonious account appealing to a single tone target is, in principle, available. In fact, both the Krongo-type and Kiowa-type sonority-driven tone restrictions are observed in linguistic relatives of Croatian that are typically classified as pitch accent languages: Long vowels are thus privileged licensers of tone in some South Slavic varieties, e.g., standard Slovene (Carlton, 1991, pp. 312–322; Greenberg, 2003; Kapović, 2015, pp. 81–85), while both long vowels and syllables closed by a sonorant preferentially support tonal distinctions in the Balto-Slavic languages: Lithuanian (Senn, 1957–66; Kenstowicz, 1972) and Latvian (Derksen, 1996).

As in other pitch accent languages, different analyses of the lexical tone contrasts in Croatian have been proposed, although we are not aware of formal treatments of tone specifically for the Split variety. Previous analyses of Croatian tone and, more broadly, tonal BCMS varieties, differ in terms of both the assumed lexical tone targets and in the temporal docking of these tones. Some of these analytic possibilities are shown schematically in Figure 2 for disyllabic words, drawing on the insights gleaned from other instrumental analyses of BCMS tone (e.g., Purcell, 1973; Lehiste & Ivić, 1986; Smiljanić & Hualde, 2000; Smiljanić, 2004; Godjevac, 2005; Pletikos, 2008; Zsiga & Zec, 2012; Zintchenko Jurlina, 2013, 2019; Langston, 2018).² On a phonological level, falling prosodemes consist of a high tone followed by a low tone and rising prosodemes of a low tone followed by a high tone, though the actual realization of the prosodemes varies depending on context. However, it has to be stressed here that different Neo-Štokavian (BCMS) varieties can have phonetically widely different systems—the basic systems can be the same, i.e., the opposition of short and long “falling” and “rising” tones—but their phonetic makeup can vary widely, i.e., the “rising” tones may not actually be phonetically rising. Thus, results from analysis of one regional variety cannot be generalized, nor should one approach them as the same system, except in very abstract terms, which was often the mistake made in much of the previous literature (Kapović, 2015, pp. 687–688).

In the simplest possible phonological analysis, the tones comprising both the falling and rising prosodemes may be assumed to fall within the stressed syllable. This analysis, which we term the monosyllabic accent analysis (Figure 2a), is reflected in the traditional Western South Slavic linguistic orthography and is also consistent with the strong typological affinity of both lexical tones and postlexical pitch accents to stressed syllables. Following our earlier discussion, closed syllables and those containing a long vowel may be assumed to be bimoraic, though nothing crucial hinges on this decision for purposes of the phonetic analysis; for these syllable types, there is a one-to-one mapping between tones and moras phonologically, though the acoustic realization of tone may differ between the syllable types. On the other hand, open syllables containing a short vowel are monomoraic and thus display a phonological association between two tones and a single mora.

A variant of this approach, which we designate the disyllabic accent analysis (Figure 2b), assumes that the first tone of the bitonal sequence is realized in the stressed syllable and the second trailing tone surfaces in the post-tonic syllable. This approach is consistent with the observation that the second tonal target, in particular for the rising tone, is characteristically realized on the post-tonic syllable in certain prosodic contexts.

Figure 2

Potential phonological analyses of Split Croatian accent.

A third possibility, which we label the monotonal accent analysis (Figure 2c), is that only the high tones are underlyingly specified and that the low tones are supplied on the surface by a default post-lexical rule of low tone insertion (Inkelas & Zec, 1988; Zsiga & Zec, 2012). Under this approach, the underlying contrast between falling and rising tone is in the location of the high tone: early in the first syllable for the falling tone, and either late in the first syllable or in the post-tonic syllable (as depicted in Figure 2c), in the case of the rising tone.

Finally, because the post-tonic syllable is typically associated with a fall in f0 in many contexts, yet another analytic option is to assume that both accents have an L target to the right of the stressed syllable at the boundary of a prosodic domain larger than the syllable. We term this approach the Final L account (Figure 2d), where there are different domains with which the final L could potentially be linked. An association with the right edge of the immediately post-tonic syllable could be construed as a foot-based final L, given that the stressed and post-tonic syllable are amenable to analysis as a metrical foot (see Köhnlein & Cameron, 2024, on foot-based analyses of pitch accent linked to stress). Other possibilities include a linkage with the right edge of the prosodic word or a phrase, as proposed by Zsiga and Zec (2012) in their study of the Belgrade variety of BCMS, although it has to be noted that modern Belgrade dialect is in a very advanced stage of tone loss and the great majority of speakers do not have a pitch-accent system anymore (unlike, e.g., Split or neighboring Novi Sad in Serbia. See Kapović, 2023, p. 247).

It may be noted that the Final L analysis has subvariants. For the falling accent, one could assume that the stressed syllable also has a low target, as in the monosyllabic analysis, or that the only low target occurs after the stressed syllable, as in the disyllabic analysis. For the rising accent, one could likewise assume that the high target is either in the stressed syllable, as in the monosyllabic analysis, or that it is realized in the post-tonic syllable, as in the disyllabic analysis. For expository simplicity, the representations in Figure 2d adopt the disyllabic approaches to both accents in combination with the low target in the post-tonic syllable.

In this paper, we explore which of the analyses in Figure 2 is most compatible with the Split variety of Croatian across different syllable types that diverge in their ability to phonetically support tone excursions. It may be noted that the four analyses depicted in Figure 2 do not represent an exhaustive set of possibilities for treating the accent distinction. Different combinations of timing (monosyllabic vs. disyllabic) and accent composition (monotonal vs. bitonal) are conceivable for either the rising or falling accent, and the two accent types do not necessarily pattern together in either their timing or tonal composition. Other potential analyses will be considered in the context of the discussion of the phonetic results. As a working hypothesis, however, we adopt the monosyllabic analysis assumed in traditional phonological analyses of Croatian in which the falling and rising prosodemes are represented, respectively, as HL and LH sequences aligned with the stressed syllables (as in Figure 2a).

Importantly, regardless of the analysis ultimately adopted, the tonal contrast is only two ways, as there is no level accent to contrast with the falling and rising accents given that the accents are historically linked to stress. In particular, the rising accent arose from words that shifted their stress one syllable to the left but retained the original location of the pitch properties associated with stress. The result was a phonetically rising tone on the newly stressed syllable. In this way, the original contrast in stress was transformed into one of tone, as observed in modern standard (Neo-Štokavian) word pairs like jȁgoda [ˈjâɡoda] ‘strawberry’ with inherited initial stress and falling tone vs. lòpata [ˈlǒpata] ‘shovel’ with original second syllable stress that now also has initial stress and rising tone.³

2.3. Tone and syllable structure in Split Croatian

This paper examines the timing of f0 peaks and troughs associated with the prosodemes of Split Croatian to (1) assess the impact of syllable structure on temporal and scaling aspects of their realization and (2) evaluate the phonetic evidence for the different analyses of accent advanced in Section 2.2. The Split variety was chosen for the purposes of another study,⁴ which considered this variety in light of the historical development of lexical pitch accent from Proto-Slavic.⁵ In the present study, we test the following hypotheses about the realization of tones, which are guided by the typology of tone restrictions dependent on syllable structure and assume the monosyllabic analysis of Croatian tone:

H1: Composition and Alignment of Tone Targets: Tone targets for both the falling and rising prosodeme will be associated with the stressed syllable. Phonological tone targets (f0 peaks and troughs) will be diagnosed through their association with f0 maxima (for high tones) and f0 minima (for low tones) that display less variability in their temporal alignment than f0 landmarks resulting from interpolation between phonological targets.

H2: Timing: Tone targets will be timed differently for syllable types depending on how conducive they are to supporting a tonal contour, i.e., a sequence of two phonological tone targets. We hypothesize that open syllables containing a short vowel (CV) and closed syllables containing a coda obstruent (CVO) are least amenable to contours. By contrast, syllables containing a long vowel (CVV) are most amenable to tonal contours, with syllables closed by a sonorant (CVR) ranking in between CVV and CV/CVO in terms of conduciveness to supporting multiple tone targets. In terms of direction, we see two possibilities: The first possibility is that the degree of temporal separation between High and Low targets will be increased for syllable types less well suited to supporting contour tones. In this situation, the phonetic realization of a contour tone is more likely to extend beyond the syllable with which the tone is phonologically associated for lower sonority syllable types. The second possibility is that the degree of temporal separation between High and Low targets will be decreased for syllable types less well suited to supporting contour tones, as the tonal targets are forced to be realized in a shorter time frame. Either way, one may expect a difference according to syllable type.

H3: Scaling: Tone excursions, i.e., the difference between the maximum and minimum f0 values, will be smaller in syllables least equipped to carry contour tones. Once again, we hypothesize that open syllables containing a short vowel (CV) and open syllables containing a coda obstruent (CVO) are least amenable to contours. By contrast, syllables containing a long vowel (CVV) are most amenable to tonal contours, with syllables closed by a sonorant (CVR) ranking in between CVV and CV/CVO in terms of conduciveness to tonal contours. That is, the difference in f0 level between High and Low targets will be decreased for syllable types less well suited to supporting contour tones. Thus, for the difference between maximum and minimum f0 values, CV/CVO < CVR < CVV.

H4: Timing and Scaling: Tone targets associated with rising tones will be more prone to greater temporal separation and reduced f0 excursions than tone targets associated with falling tones—that is, rising tones will be particularly sensitive to differences in syllable structure because they take longer to realize than falling tones (Sundberg, 1979).

It may be noted that these hypotheses are not mutually exclusive. It is thus possible that a combination of pitch rescaling and temporal shift could be employed to minimize tonal crowding or that the rising and falling prosodemes could behave asymmetrically with respect to their phonological alignment of tones or the strategies they adopt to reduce crowding of tones.

Because Croatian lacks phonemic level tones, there is less pressure to realize the two-way tone contrast as a distinction between a rising and falling tone, since one of the two could easily be rescaled to minimize the tone excursion, effectively recasting the contrast as one between a contour tone and a level tone. In practice and in keeping with H4, one would expect the rising tone to be most prone to leveling out.

Furthermore, because tonal contrasts are limited to the stressed syllable in Croatian, there is additional room on which to maximize the temporal separation of the low and high tone targets that comprise the two contour tones: HL in the case of the falling accent and LH in the case of the rising accent. This differs from more canonical tone languages in which contrasts in lexical tone are associated with adjacent syllables.

In short, the absence of a contrast between contour tones and level tones coupled with the confinement of tone distinctions to the stressed syllable in Croatian provide considerable leeway for increasing temporal separation between tone targets and/or rescaling them to minimize the articulatory and perceptual demands of contour tones.

3.1. Speakers and recordings

Thirteen speakers (including one male) of the Split variety of Croatian were recorded in March 2022 at the Department of Phonetics recording studio at the Faculty of Humanities and Social Sciences, University of Zagreb, under the supervision of a recording technician and author MK. Speakers were recorded using an AKG C414 B-ULS microphone and RME Fireface UFX soundcard. Recordings were saved as mono WAV files using a 44.1 kHz sampling rate and 16-bit bit-depth. Collected data were also employed in another study of acoustic characteristics of tone and stress in Split Croatian.

All speakers were born in Split between 1996 and 2002 (with most born between 1999 and 2002) and arrived in Zagreb for their university studies between 2015 and 2021. It should be noted that although the Split variety is not the standard variety, it is in fact a prestigious and well-known variety in Croatia, associated with popular music and with various celebrities in the country. With more than 160,000 inhabitants, Split is the second largest town in Croatia and the main city in the south of the country and the Dalmatian coast. Thus, the speakers in our recordings were not speaking a stigmatized variety of the language in the university setting and were encouraged to pronounce the words as they normally do. Author MK verified that the speakers spoke the Split variety of Croatian during the recording.

3.2. Stimuli

Speakers read a list of 65 real Croatian words that illustrated the four prosodemes. Four of these words contained a syllabic /r/ in the stressed syllable, and we will not consider these words in the present study due to difficulties involving segmentation of this sound. This leaves us with 61 lexical items in the present study.

Words were read in isolation, with no carrier phrase. The words were either stand-alone lexical items or prosodic words with enclitics (Croatian has an extensive system of enclitics and proclitics). All words contained lexical stress on the first syllable. Words contained at least three syllables and up to six syllables (15 words with three syllables, 23 words with four syllables, 19 words with five syllables and four words with six syllables). The purpose of having longer words was to make sure there were at least two syllables following the initial stressed syllable to allow for any phrasal tones that may occur in these single-word utterances.

In setting up the wordlist, we endeavored to find words in which the lexical pitch accent occurred in six different syllable types: an open syllable containing a phonemic short vowel (CV); an open syllable containing a phonemic long vowel (CVV); a short vowel syllable closed by sonorant (CVR); a long vowel syllable closed by a sonorant (CVVR); a short vowel syllable closed by an obstruent (CVO); and a long vowel syllable closed by an obstruent (CVVO). Syllable structure is notoriously difficult to diagnose in Croatian as typical diagnostics used in other languages are not applicable. In particular, the common strategy of comparing word-initial and word-medial clusters is not informative because a rich array of initial clusters is permitted in Croatian (Kapović, 2023, pp. 216–222), including many that speakers would potentially treat as spanning a syllable boundary intervocalically. We thus adopted the typologically most common strategy for syllabifying intervocalic consonant clusters according to which at least one consonant belongs to a coda and at least one to an onset. In practice, this means that a CC cluster, the only type of cluster found in our data, is split by a syllable boundary. Syllables were distributed as follows in our data (Table 2–see also Appendix A for wordlist):

A word of caution is needed at this point. There is an overall paucity of words containing a target syllable closed by a sonorant. This is a reflection of the lexical situation of the language, particularly for the trisyllabic and longer words that we targeted. As a consequence, results for syllables closed by a sonorant coda must be interpreted with caution.

The wordlist was randomized, and each speaker read four repetitions of the list. (One speaker did not produce the word klȁckali ̮su ̮se [Short Falling] ‘they were see-sawing’, due to the addition of this word to the list after the speaker had been recorded.) We thus have a total of 3172 possible word tokens for analysis (61 words × 4 repetitions × 13 speakers, minus four tokens for the speaker who did not read a particular word).

One final note concerns vowel quantity on the post-tonic syllable. Croatian contrasts long and short vowels in all positions of the word (except before the stressed syllable). In the present study, we were not able to properly control for post-tonic length of the vowel, given the other factors we attempted to control. For this reason, both long and short vowels occur in the syllable following the stressed syllable in our word list. Nevertheless, we have included post-tonic vowel quantity as a factor in our inferential statistical analyses (see below).

3.3. Data analysis

Phonetic transcriptions of the words were imported from a spreadsheet and used for preliminary phonetic segmentation with the Munich AUtomatic Segmentation system (MAUS: Kisler et al., 2017) pipeline function MAUS->PHO2SYL. Manual correction of the phonetic MAUS labelling was conducted using the EMU Speech Database Management System (Winkelmann et al., 2017; Winkelmann et al., 2019), interfaced with the R statistical software package (R Core Team, 2020). The Snack signal processor (Sjölander, 2014) was used for calculating formants within the VoiceSauce software package, which was also used to extract voicing measures including Straight f0 (Shue, 2010; Vicenik et al., 2020). All of these signal measures were extracted at a sample rate of 1000 Hz (i.e., every 1 ms). Plots were generated using the ggplot2 package in R (Wickham, 2016).

3.3.1. Identifying peaks and troughs in the f0 traces

For the purposes of identifying peaks and troughs in the f0 traces, we chose the vowel as the window of analysis (in either the stressed or post-stressed syllable). To maximize our chances of finding a peak and/or trough within this window, we added 10 ms before the vowel, and 10 ms after the vowel. We were unable to use the syllable as the basis for our analysis because many consonants either side of the vowel were (voiceless) obstruents, with either no f0 or unreliable f0 traces.

Dominant peaks and troughs were identified by progressively smoothing the signal until only two roots (zero crossings) of the first derivative of the smoothed signal were present. This strategy eliminated peaks and troughs caused by noise, while selecting the strongest peaks/troughs. We chose to search for two roots as the target number of roots, since we considered it possible to find both a maximum and a minimum in each vowel window, based on our examination of pitch traces in a related study using the same database (from which Figure 3, presented further below, is taken). In cases where a smoothing iteration produced fewer than two roots, the function returned the roots (e.g., three or four) from the previous smoothing iteration. However, if the first smoothing iteration identified two or fewer roots, it returned the root(s) from this first smooth.

Figure 3

The four prosodemes of Split Croatian. This plot is based on 3376 tokens from our database. Please note that “pitch” and “quantity” refer to pitch and quantity of the stressed syllable.

Smoothing was implemented using the smooth.spline() function in baseR, which also allows prediction of the derivative, while the rootsolve package was used to identify the roots (Soetaert, 2009; Soetaert & Herman, 2009). Note that in 17 instances, the function failed to converge even after 20 smoothing iterations (this total of 17 instances includes both the stressed and post-stress vowels).

Examination of data returned by the smoothing function showed that most tokens contained either two or three turning points, and a smaller number of tokens contained only one turning point. Very few tokens contained either zero turning points (i.e., a straight line) or four or more turning points: These tokens were discarded. In addition, of the tokens that contained only one turning point, very few of these contained a pitch minimum as opposed to a pitch maximum. For this reason, tokens with only a single minimum were also discarded. Moreover, examination of the data suggested that where three turning points were found, the last turning point was quite late in the token. Given that we had added 10 ms to the end of the vowel window, we decided to discard the third turning point where three turning points were found and only keep the first two for that particular token.

Finally, we removed any turning points which had an f0 value of less than 100 Hz or greater than 300 Hz, since these were likely to be a result of tracking errors. All of the above filtering procedures left 3000 tokens where the stressed syllable was analyzed, and 2678 tokens where the post-stressed syllable was analyzed from the original 3172 tokens submitted to the smoother. Table 3 shows the number of tokens for each prosodeme in the two syllable positions (stress and post-stress).

Table 3

Number of tokens by prosodeme (falling or rising, and short or long) and syllable position (stress or post-stress).

	Stress		Post-Stress
	Short	Long	Short	Long
Falling	701	523	646	487
Rising	904	872	782	763

Table 4 shows the number of turning points used for statistical analysis after the above filtering procedure. Note that this table combines tokens where only one turning point was found in a syllable, with tokens where two turning points were found in the syllable. As such it is not possible to add any two cells in a meaningful way. The table can simply tell us that for falling pitch accents, it was a little more common to find a maximum in the stressed syllable than a minimum; for rising pitch accents, both maxima and minima were found in the stressed syllable. Also, for rising pitch accents, a maximum was more likely to be found than a minimum in the post-stressed syllable. None of this precludes the possibility that both a maximum and a minimum were found for a syllable (and of course, also the possibility that neither was found).

Table 4

Number of individual minima and maxima found for each pitch accent type, in each syllable position.

	Stress		Post-Stress
	Minimum	Maximum	Minimum	Maximum
Falling	1020	1223	1080	1133
Rising	1720	1776	1302	1543

3.4 Statistics

Linear Mixed Effects (LME) analyses were conducted using the nlme package of R (Pinheiro et al., 2016). LME models allow us to set speaker and word as random effects in the data analysis and are robust against differing numbers of tokens in each cell, as is the case in the present study.

For vowel duration, we used the following command:

lme(Duration~SyllableWeight, data=data, random=~1|speaker/words)

where SyllableWeight was one of Open; (Closed-by-)Sonorant; or (Closed-by-)Obstruent. We conducted separate tests for the Short Falling, Short Rising, Long Falling and Long Rising prosodemes (i.e., four prosodemes); and for the stress and post-stressed syllables (two syllable positions). This gives us eight separate tests (four prosodemes multiplied by two syllable positions). As a result, we set a significance level of p < 0.00625 after Bonferroni correction (0.05 divided by eight tests). These results are presented in Appendix C.

For the maxima and minima (i.e., pitch peaks and troughs), we used the following command:

lme(Measure~Quantity*SyllableWeight, random=(~1|speaker/words)

where Quantity was either Long or Short (length of the stressed vowel); and SyllableWeight was one of Open; (Closed-by-)Sonorant; or (Closed-by-)Obstruent. As can be seen, for studying the pitch minima and maxima, we treated Quantity as an independent variable, since we expected that the dependent variable would be strongly affected by whether the stressed vowel was long or short. We conducted separate tests for the falling and rising pitch accents (i.e., two accents); for the stress and post-stressed syllables (two syllable positions); and for pitch minima and maxima (two turning points). This, once again, gives us eight separate tests (two accents multiplied by two syllable positions multiplied by two turning points). As a result, we once again set a significance level of p < 0.00625 after Bonferroni correction.

The measures examined in these LME models were the f0 values (in Hertz); the relative timing of the turning points (i.e., timing based on normalized time–this is therefore a dimensionless value between zero and one); and the absolute timing of the turning points (i.e., timing without any normalization of time–measured in milliseconds). These results are presented in Appendix D.

Although it would have been desirable to include post-stress vowel length as a factor in these LME models, our models did not converge when we did so. Note also that LME models, which included a random slope for speaker with another variable, failed to converge. Nevertheless, we included the single male speaker together with our female speakers, since examination of the f0 data suggested that the male speaker did not pattern obviously differently from the other speakers and did not necessarily have the lowest f0 of the 13 speakers. Indeed, examination of the speaker random effects intercept values for the various LME models we ran showed that at least one female speaker had a lower f0 than the male speaker. This is consistent with the sociolinguistic observation that female speakers of Croatian tend to have a low f0.

4.1 Duration

Figure 4 shows the vowel duration according to syllable type, and LME results are presented in Appendix C. As one would expect, in the stressed syllable (the left column), the long vowels have greater duration than the short vowels, as shown by solid versus dashed lines. A first, broad inspection of these boxplots suggests that any differences according to syllable type are not extensive. However, a closer inspection of the boxplots in this left column shows that the long vowels are longer when they are in an open syllable (solid brown boxes in the middle). The LME results estimate that long vowels in open syllables are about 22 ms longer (than in obstruent-closed syllables) for the falling accent and about 16 ms longer for the rising accent. No such effect applies in the case of short vowels in open syllables. However, the LME results do suggest that in the case of the short rising prosodeme, the vowel is about 12 ms longer in a sonorant-closed syllable than in an obstruent-closed syllable.

Figure 4

Vowel duration according to syllable type (shown as different colours). Data show the different pitch accents in rows (falling and rising); the different stress vowel quantities according to linetype (long and short); and the different syllable positions in columns (stress and post-stressed syllables). Note that long/short refers to quantity of the stressed syllable.

Turning now to the post-stressed syllable (right column), the reader is reminded that the long/short coding in these panels refers to the quantity of the stressed syllable, not of the post-stressed syllable. In addition, the reader is also reminded that we were unable to control for quantity in the post-tonic syllable (i.e., whether the post-stress vowel is long or short), given the other factors which we also sought to control. With these caveats in mind, it may be noted that the vowel following a short rising syllable (broken line boxes in the bottom right panel) is longer than the vowel following a long rising syllable (solid boxes in the same panel). We suggest that this effect is essentially due to the fall on the post-stressed syllable being a defining feature of the rising prosodeme in this variety of Croatian; that is, a slightly longer vowel duration is needed on the post-stressed syllable in order to fully realize the final fall of the short rising prosodeme. It may also be noted that there are otherwise minimal effects according to syllable type in the post-stressed syllable. LME results estimate that the long rising prosodeme is about 22 ms shorter in the sonorant-closed syllable context than in the obstruent-closed syllable context (solid brown box in the bottom right panel). However, the reader is reminded that there is only one lexical item in our wordlist which has a sonorant coda on a long rising prosodeme–this is the word bólničārka [bǒːlnit͡ʃaːrka] ‘nurse’ which has a short post-tonic vowel /i/. For this reason, we may consider this result non-informative for our purposes.

In summary, the data show minimal effects of syllable type on either stressed or post-stress vowel duration. The exception is for the long prosodemes (falling and rising), where the vowel is longer in an open syllable. We attribute this to the common typological pattern of open syllable vowel lengthening (Maddieson, 1985) but limited to phonemic long vowels which, unlike short vowels, are free to lengthen without potentially jeopardizing the phonemic contrast in vowel length.

4.2 Pitch minima and maxima regardless of syllable type

We now turn to the timing and f0 levels of the accents. In this section we offer informal observations of the results, regardless of syllable type, to make sure that the maxima and minima we are finding are in line with expectations. Figure 5 shows the f0 values of the minima and maxima across the four prosodemes of Split Croatian. Minima are shown on the top row and maxima in the bottom row; the stressed vowel is shown in the left column and the post-stress vowel in the right column.

Figure 5

f0 values (in Hertz) for minima (top row) and maxima (bottom row) of stressed (left) and post-stress (right) vowels. Rising and falling pitch accents are shown by colour. Long and short vowels are shown by line type. Note that long/short refers to quantity of the stressed syllable.

In the bottom left panel, we can see that the maxima in the stressed syllable have a higher f0 for the falling accents (blue boxes) than the rising accents (brown boxes), for both the long and short vowels (denoted by linetype). This is in line with expectations based on the f0 traces shown in Figure 3. At the same time, the maxima in the post-stressed syllable have a higher f0 for the rising accents than for the falling accents (bottom right panel). This is also in line with expectations based on Figure 3.

Figure 6 shows the normalized time values (from zero to one) of the minima and maxima across the four prosodemes. Time was normalized within the get_trackdata() function in the emuR package. For reasons of space, we do not present results for absolute time (in ms) in this paper. However, broadly speaking, they mirror the results for normalized time.

Figure 6

Normalized time (from zero to one) for minima (top row) and maxima (bottom row) of stressed (left) and post-stress (right) vowels. Rising and falling pitch accents are shown by color. Long and short vowels are shown by line type. Note that long/short refers to quantity of the stressed syllable.

In the stressed syllable (left column), maxima for the rising accents (brown boxes–bottom left) occur later than the maxima for the falling accents (blue boxes–also bottom left). The maximum for the falling accent occurs quite early, at about the 20% point of the vowel duration. This is what we would expect for a rising versus falling pitch accent docked onto the stressed syllable.

Conversely, the maxima in the post-stressed syllable (bottom right) occur earlier in the rising accent than in the falling accent–this is likewise what we would expect given that the rising accent has a clear high fall on the post-stressed syllable in the Split variety. The maximum for the rising accent similarly occurs quite early in the post-stressed syllable, also at about the 20% point of vowel duration.

Importantly, the variability in the timing of the maxima differs between the rising/falling accents. The falling accent has less variability in the timing of the maximum in the stressed syllable, while the rising accent has less variability in the maximum in the post-stressed syllable. This reflects the importance of the stressed syllable peak for the falling accent, but of the post-stressed syllable peak for the rising accent.

Considering now the minima of the stressed syllable (top left panel), we see that although there is less of a difference in the means between rising and falling accents, there is a clear pattern of less variability in the timing of the minimum of the rising accent. The minimum of the rising accent seems to be timed for about the halfway point of the stressed vowel, whereas the timing of the minimum of the falling accent is much more variable. As for the post-stressed syllable (top right panel), we see that the minimum for the rising accent tends to occur later than the minimum for the falling accent. This is in line with the fact that the f0 maximum for the rising tone is not reached until the post-stress syllable, which forces the immediately following minimum to be realized relatively late in the same syllable.

The f0 values in Figure 5 and the timing results in Figure 6 provide insight into the analysis of phonological targets for the two accents. For the falling accent, the f0 maximum is higher in the stressed syllable than the post-tonic syllable and is less variable in its timing, occurring consistently early, a pattern that is consistent with an early H target in the stressed syllable. Furthermore, the f0 minimum is lower and less variable in the post-tonic syllable than in the stressed syllable for the falling accent, suggesting the possibility of an L target in the post-tonic syllable. These results are consistent with both the Disyllabic (Figure 2b) and the Final L (Figure 2d) analyses and are inconsistent with both the Monosyllabic (Figure 2a) and the Monotonal (Figure 2c) analyses.

The rising accent is more probative in distinguishing between the Disyllabic and the Final L analysis. The f0 minimum, though only slightly lower than the f0 maximum, is consistent in its alignment, occurring shortly before the midpoint of the stressed vowel, suggesting the presence of an L target in the first syllable in contrast to the prediction of the Monotonal account. Furthermore, the presence of an f0 maximum on the post-tonic syllable suggests an H target early in the post-tonic syllable. There is also a clear downward trend in pitch for the rising tone following the H target in the post-tonic syllable. Overall, f0 results for the rising tone are most compatible with the Final L analysis positing an LHL sequence, with the first L tone docking on the stressed syllable, the H on the post-tonic syllabic, and the second L either at the end of the post-tonic syllable or to its right.

One remaining question concerns the domain with which the final L is associated: the foot, word, or phrase? Our data do not allow a definitive analysis, though they suggest an asymmetry between the falling and rising tone. For the falling tone, the relatively early realization of the f0 minimum (at around the halfway point) in the post-stressed syllable is consistent with an L target in the post-stressed syllable. In contrast, the f0 minimum for the rising accent is reached quite late in the post-stressed syllable. One possibility is a foot-based approach (Köhnlein & Cameron, 2024), in which the first two syllables form a trochaic foot characterized by an L in the weak (post-tonic) syllable final preceded by an H for the falling accent and an LH sequence for the rising accent. For both accents, the tone(s) preceding the weak-syllable L are plausibly phonologically aligned with the stressed syllable, but, in the case of the rising accent, tonal crowding attributed to the initial L forces the trailing H to be realized early in the post-tonic syllable.

The asymmetry between the falling and rising accents accords with f0 data in other studies (e.g., Pletikos, 2007; Zsiga & Zec, 2012), showing a pitch trough in the post-stressed syllable followed by a low plateau through the end of the word in words with a falling accent but a continuous fall in pitch through the end of a word with a rising accent. Clearer insight into the source of the final low tone after the rising accent must await further acoustic study of syllables to the right of the immediately post-tonic one in words and phrases of differing lengths.

4.3. Pitch minima and maxima according to syllable type

We now turn to the timing and f0 of the accents according to syllable type. We present results separately for the falling and rising pitch accents. LME results for the data in this section are presented in Appendix D.

4.3.1. Falling

Figure 7 shows the f0 maxima and minima for the falling pitch accents according to syllable type. (The reader is reminded that, in all cases, open~obstruent~sonorant refers to the structure of the stressed syllable, not of the post-stressed syllable.)

Figure 7

Minimum (top row) and maximum (bottom row) f0 values of stressed (left) and post-stress (right) vowels with a falling pitch accent. Different syllable types are shown by color. Long and short vowels are shown by line type. Note that long/short refers to quantity of the stressed syllable.

Looking at the top two panels, the f0 minimum falls on the post-tonic syllable across different types of stressed syllables in keeping with the overall result in Figure 5. Likewise, the bottom two panels confirm a consistent f0 peak within the stressed syllable. Focusing on the top left panel (minima of the stressed syllable), we can see that the f0 minimum is higher for short vowels than for long vowels (by about 11 Hz, according to the LME estimate). At the same time, the f0 maximum (bottom left panel) is lower for short vowels than for long vowels (by about 9 Hz, according to the LME estimate). One can therefore conclude that the f0 fall is greater in long vowels than in short vowels, by about 20 Hz. Given that the f0 values for our speakers tend around 200 Hz, this represents a difference of about 10%. Thus, there is a re-scaling of f0 values in the stressed syllable of the falling accent based on vowel length in support of Hypothesis 3.

However, there seems to be minimal effect of the presence or type of coda consonant on f0 minima and maxima in the stressed and post-stressed syllables. The exception is an effect of open syllables on the f0 maximum in the stressed syllable–however, this only applies for long vowels, which have a lower f0 maximum in open syllables compared to their counterparts in closed syllables. This pitch scaling effect is not predicted since all long vowels are phonologically well suited to accommodating targets in a contour tone. In fact, because long vowels in open syllables are longer than those in closed syllables (Figure 4), they should be especially well suited to accommodating contour tones without any need of rescaling to minimize the pitch excursion between high and low targets. The LME results also estimate a lower f0 minimum for short vowels in the post-stressed syllable by about 5 Hz, a pattern that is inconsistent with predictions given that short vowels offer less time for pitch excursion than long vowels.

Figure 8 shows the relative timing of maxima and minima for the falling pitch accents according to syllable type. The f0 values for the phonological high target in the stressed syllable and the low target in the post-tonic syllable are less variable than their counterparts in the other syllable across all syllable types. Furthermore, it may be noted that there is a clear effect of long/short vowel on the timing of the maximum in the stressed syllable (bottom left panel). According to the LME estimate, the f0 maximum occurs 0.12 later in the short vowel than in the long vowel (where relative time is measured between zero and one). It is also important to note the very low variability in this panel compared to the other panels. At the same time, the variability in timing of the f0 maximum in the stressed syllable is greater for short vowels than for long vowels. This suggests a greater stability in timing for long vowels within an overall greater stability in timing for this particular pitch event compared to other pitch events.

Figure 8

Timing of minimum (top row) and maximum (bottom row) for stressed (left) and post-stress (right) vowels with a falling pitch accent. Different syllable types are shown by color. Long and short vowels are shown by line type. Note that long/short refers to quantity of the stressed syllable.

In terms of syllable type, perhaps the most obvious effects can be seen for the minimum in post-stressed syllables (top right panel)–in this case, the minimum is timed later following an open syllable (brown boxes in the middle) by about 0.11 (normalized time) according to the LME estimate. In addition, the LME estimates that the maximum in the stressed syllable (bottom left panel) is also timed later in open syllables–by about 0.07–though this is not so obvious when looking at the plot. In general, one could tentatively assume that pitch events are timed later in open syllables compared to closed syllables. This difference is not dependent on vowel length as would be predicted by Hypothesis 3, according to which one might expect the High pitch target for the falling tone to be realized earlier in an open stressed syllable with a short vowel in order to allow sufficient time for the pitch fall.

In summary, for falling accents, we see that pitch is re-scaled in the stressed syllable, with a lesser fall in short vowels than in long vowels (pitch floor raised, and ceiling lowered). At the same time, the pitch peak is timed later in short vowels than in long vowels, a result that perhaps runs counter to Hypothesis 3, which would predict an earlier realization of the High target in order to allow adequate time for the fall in a short vowel. Any effects of presence/absence of coda consonant or type of coda seem to be limited to a slightly later timing of the pitch minimum in post-stressed syllables following an open stressed syllable. This difference is mainly attributed to stressed long vowels and is also not predicted by Hypothesis 2, according to which an earlier realization of the post-stress minimum would be expected following short vowels in open syllables in order to increase temporal separation from the high target in the stressed syllables.

4.3.2. Rising

Figure 9 shows the f0 maxima and minima for the rising pitch accents according to syllable type (parallel to Figure 7 for the falling pitch accents).

Figure 9

Minimum (top row) and maximum (bottom row) f0 values of stressed (left) and post-stress (right) vowels with a rising pitch accent. Different syllable types are shown by color. Long and short vowels are shown by line type. Note that long/short refers to quantity of the stressed syllable.

A visual examination of these plots shows no difference according to syllable type or vowel length in the stressed syllable (left column). Most apparent is the tendency for the f0 maximum to be considerably higher in the post-tonic syllable than in the stressed syllable in the open syllable condition, whereas the maximum value is similar across the two syllables in the other conditions. This result suggests that the overall alignment of the H tone of the rising accent in the post-tonic syllable observed earlier in Figure 6 is largely driven by cases in which the rising accent is associated with an open stressed syllable. The parallel behavior of short and long vowels in open syllables is not predicted by Hypothesis 2 since the two syllable types differ considerably in their sonority. Minor characteristics of open syllables observed in the LME results are a slightly lower minimum and slightly lower maximum in open syllables and a slightly higher maximum for short vowels. However, these differences are all around 4 Hz, in the context of an overall data estimate of about 180 Hz. We could therefore conclude that any such differences are unlikely to be perceived. The LME results also show no significant differences for either syllable type or vowel length in the case of the post-stress minima (top right panel); however, they do suggest a higher f0 maximum in the post-stressed syllable following short vowels (by about 8 Hz) and following an open syllable (by about 12 Hz). These effects on the maximum may be seen in the plot (bottom right panel), but the relative f0 differences are once again relatively small, and therefore not likely to be perceived.

Figure 10 shows the relative timing of maxima and minima for the rising pitch accents according to syllable type (parallel to Figure 8 for the falling pitch accents).

Figure 10

Timing of minimum (top row) and maximum (bottom row) for stressed (left) and post-stress (right) vowels with a rising pitch accent. Different syllable types are shown by color. Long and short vowels are shown by line type. Note that long/short refers to quantity of the stressed syllable.

Considering first the stressed syllable, there are no significant differences in timing of the pitch minima (top left panel) according to either vowel length or syllable type. As regards the timing of the pitch maxima (bottom left panel), the LME results suggest that the maximum occurs earlier in short vowels than in long vowels (by about 0.14 normalized time). However, there is considerable variability in timing, so no conclusions about timing of pitch targets for the falling accent should be drawn. There are also significant LME results involving the sonorant-closed syllables, including an interaction with vowel length–however, the reader is reminded that there is only one lexical item for the long rising pitch accent in a sonorant-closed syllable (the word for ‘nurse’), and as such it would be unwise to consider this a robust result (it is the solid orange box with the noticeably lower mean).

Turning now to the post-stressed syllable, we see a noticeably later timing of the post-stress minimum following an open syllable (top right panel). The LME results estimate that it occurs 0.08 later (on a scale from zero to one). Similarly, the LME results estimate that the post-stress maximum (bottom right panel) occurs 0.12 earlier following an open syllable (there is once again an interaction involving the Sonorant here, which we disregard for the reasons given above). It therefore seems that the final fall is temporally extended following an open syllable, with the peak starting earlier and the trough occurring later in the post-stressed syllable. One may also observe at this point the overall lesser variability in the timing of the post-stress maximum following open syllables, which together with the higher f0 maximum in the post-tonic syllable relative to a stressed open syllable in Figure 8, provides support for the H component of the rising tone being aligned with the post-tonic syllable. However, because a similar pattern is observed for both short and long vowels in open syllables, it does not provide support for a link between the sonority of a syllable and the temporal realization of tone targets as predicted by Hypothesis 2. The reduced variability in timing of the maximum f0 in the post-tonic syllable does, however, present an interesting contrast with the f0 values in Figure 8.

In summary, we see no differences of syllable type or of vowel length on the f0 values themselves. However, the timing of the pitch peaks is earlier in short vowels than in long vowels in the stressed syllable. In the post-stressed syllable, we see an effect of the preceding open syllable, whereby the post-tonic fall is temporally extended: starting earlier and finishing later. It is not clear why there may be such an effect. It is also important to note that the minimum in the stressed syllable is timed for around halfway through the vowel regardless of any suprasegmental factors, and the maximum is timed for early in the post-stress vowel to an extent also regardless of any suprasegmental factors. The relatively lesser variability in the timing of these two pitch targets is once again evident in Figure 10.