Stress deletion or stress demotion? An acoustic study of stress in Spanish lexical compounds

1.1 Spanish lexical compounds

Spanish compound words can be divided into two major classes: (i) lexical (also known as orthographic or proper) and (ii) syntactic (also known as phrasal or syntagmatic); they are mainly distinguished by the criterion of morpho-phonological integration (RAE 2011; Varela 2012). The words forming a lexical compound are morpho-phonologically integrated, as evinced by the preservation of a single primary stress and by structural adjustments taking place at word edges (e.g., identical vowel simplification in telaraña ‘spider web’, cf. *telaaraña); they are normally represented as a single unit.^[1] Syntactic compounds generally show, by contrast, no signs of morpho-phonological integration (e.g., cuenta corriente ‘checking account’). It is often posited that they retain the stress of each constituent word, and they are orthographically represented as separate units.^[2] The two categories of compound, according to Varela (2012), may also differ in their way of inflection. Lexical compounds always have their inflectional markers at the end of the whole compound, whereas in syntactic compounds it is their head that always receives the inflectional markers (p. 218).^[3]

Within lexical compounds, two additional distinctions are relevant. The first one involves the notion of head, which is usually defined on syntactic and semantic grounds. Syntactically, the head determines the lexical category of the whole compound, and semantically, it functions as a hypernym. Depending on whether the head is internal or external, compounds may be endocentric or exocentric (Kornfeld 2009, p. 437; Rainer and Varela 1992, p. 122; Val Álvaro 1999, p. 4766; Varela 2012, p. 218; among others). In the former, one of the embedded words functions as the head (e.g., media noche ‘half night’ → medianoche ‘midnight’), whereas in the latter none of the embedded words plays that role (e.g., cara rota ‘face broken’ → cararrota ‘cheeky person’). The other relevant distinction is based on the relationship between the constituent words (Scalise and Bisetto 2009; Scalise and Vogel 2010). When the constituent words are of equal status, the compound is coordinate (e.g., agrio y dulce ‘bitter and sweet’ → agridulce ‘bittersweet’). When the constituent words differ in status, the compound may be subordinate, if one word complements the other (e.g., saca corchos ‘pulls corks’ → sacacorchos ‘corkscrew’), or attributive, if one word assigns an attribute to the other (e.g., noche buena ‘good night’ → nochebuena ‘Christmas Eve’). According to Bustos Gisbert (1986), Spanish compounds of different morphosyntactic and semantic types might exhibit different stress patterns (p. 184). To verify this hypothesis, the current study examines lexical compounds that represent all the different types as introduced above.

1.2 Spanish stresses and their phonetic correlates

Unlike English which has two levels of word stress (i.e., primary and secondary), Spanish has only one word stress, i.e., primary. It is noteworthy, however, that in the literature several authors have observed secondary rhythmic stresses falling on alternating syllables from the syllable bearing primary stress (e.g., Harris 1983, 1991; Navarro Tomás 1918; Roca 1986) or on the word-initial syllable (e.g., Bolinger 1962; Prieto and van Santen 1996; Stockwell et al. 1956). This type of stress has been often found associated with a falling pitch accent and an amplitude peak (Prieto and van Santen 1996). Linguists have recently defined this type of stress as rhetorical secondary stress for it being an optional post-lexical phenomenon used for attention-getting or authority-seeking (Hualde and Nadeu 2014). It has been reported in radio talks, news broadcasts, political discourses and lectures, but it is relatively rare in conversational speech styles (Fabregat and de la Mota 2010; Hualde and Nadeu 2014; Hualde 2007, 2009, 2010; Kuder 2020; Quilis 1981; Rao 2011). Moreover, it should be noted that one of the basic parameters of lexical stress is its distinctive function (Hualde 2012, p. 153). For example, the primary stress in the verb pa ra ‘he/she/it stops’ seems to be the only sound property distinguishing this word from the preposition para ‘for’ within a phrase. By contrast, the rhetorical secondary stress never contrasts words, and accordingly, it is not considered lexical stress.

In Spanish, all content words including nouns, adjectives, verbs, and adverbs, as well as some classes of function words invariably receive lexical (or primary) stress on one of their last three syllables, which is known as the Three-Syllable Window. Previous studies have identified four phonetic properties that may serve to cue lexical stress in Spanish: pitch, duration, intensity, and spectral tilt. They do not appear to be equally reliable, though. It has been observed that variations in pitch and duration correlate with variations in stress more systematically than variations in intensity and spectral tilt do.

In Spanish linguistics literature, the view that pitch accent is the most important indicator of lexical stress in Spanish is dominant (Bollinger and Hodapp 1961; Contreras 1963, 1964; Enríquez et al. 1989; Figueras and Santiago 1993; Gili Gaya 1975; Llisterri et al. 2005; Monroy-Casas 1977; Quilis 1971, 1981, 1993). The reason why pitch or f0 is recognised as a reliable stress cue is that in stress-accent languages like Spanish, there is a strong tendency for stressed syllables to be accompanied by pitch accents (e.g., Beckman and Edwards 1994; Beckman and Pierrehumbert 1986; Bolinger 1958; Gussenhoven 2004; Ladd 1996; Pierrehumbert 1980; etc.). In the case of Spanish, seven types of pitch accent have been recognised. Using the ToBI labelling scheme, they are transcribed as H*, L*, H + L*, L* + H, L + H*, L + ¡H* and L + <H* (Aguilar et al. 2009; Hualde and Prieto 2015).^[4] Among them, the rising pitch accents are the most common types associated with lexical stress in Spanish words in declarative non-phrase-final position (Hualde 2003; Llisterri et al. 2005; etc.). Moreover, rising and high pitch accents are deemed perceptually more prominent than falling and low pitch accents. According to Prieto et al. (2005), rising and high pitch accents tend to be perceived as high tones, whereas falling and low pitch accents are perceived as low tones (p. 375).

Despite the interweaving of stress and pitch accent, one must bear in mind that stress and pitch accent embody different types of prominence. Stress signals prominence at word level, whereas pitch accent signals prominence at phrase level. It is thus possible for a syllable to be stressed or unstressed depending on the pattern of lexical prominence of its word. Additionally, it is possible for a stressed syllable to be accented or unaccented depending on the intonational pattern of the utterance where it is found (e.g., Beckman and Edwards 1994; Ladd 1996; Pierrehumbert 1980). Convergence of both types of prominence on the same syllable raises the problem of how to tease them apart. To ensure that the effects of stress are not confused with those of pitch accent, this study scrutinises the linguistic structures under investigation under two different conditions: in normal declarative sentences where pitch accents are expected to cue stress, and in post-focal position where no pitch accent is expected and any prosodic prominence on a syllable can be considered a reliable indication of stress (for deaccenting in post-focal position, see de la Mota 1995, 1997; Face 2001).

Duration has proven to be a reliable stress cue in numerous languages (Beckman and Edwards 1994, for English; van Heuven and Sluijter 1996, Sluijter et al. 1997, for Dutch; Dogil and Williams 1999, for German; Ortega-Llebaria 2006, Ortega-Llebaria and Prieto 2008, 2010, for Spanish; etc.). Stressed syllables and their vowels tend to be significantly longer than unstressed ones. This tendency holds strong in Spanish; in fact, studies such as Canellada and Kuhlman-Madson (1987), Garrido et al. (1995) and Díaz-Campos (2000) argue that duration is the main cue to Spanish stress. One must be cautious, however, because stress is not the only factor affecting duration. Segmental structure, position within words and phrases, intonation, and speech rate do so too (e.g., Clark et al. 2007, p. 333; Ortega-Llebaria 2006, p. 111; Yang 1998, p. 42). To prevent interference by those additional factors, the present study takes the following precautions: the syllables to be compared share similar segmental structure and within- word position, the words containing them are kept away from phrase-final position, the intonational context surrounding the words under comparison is held constant, and relative rather than raw duration values are employed.

As for intensity, there was a time when it was regarded as the main stress cue in Spanish. Back in the early 1900s, Navarro Tomás (1918) claimed that intensity is the strongest cue to stress in this language, and to underscore its centrality, he dubbed this type of prominence acento de intensidad (‘intensity stress’). This view has been endorsed by recent acoustic studies such as González et al. (2006) and Urrutia Cardenas (2007). Their statistical analyses revealed that intensity is the most important indicator of stress at both word and phrase levels. Not all recent studies agree, though. Ortega-Llebaria and Prieto (2008) found that while stressed and unstressed vowels differ in overall intensity within declarative sentences (i.e., accented contexts), that is not so within parenthetical sentences (i.e., unaccented context). In addition, Ortega-Llebaria and Prieto (2010) established that although stressed vowels have greater overall intensity than unstressed ones, the difference is virtually imperceptible.

Spectral tilt, the intensity difference between the lower and higher frequencies of the spectrum, has also been linked to stress; however, its role is controversial. Evidence that it serves to cue stress is provided in Ortega-Llebaria and Prieto (2008). They found that stressed syllables, regardless of context, tend to have greater intensity in the higher frequencies, and consequently, their spectral tilt tends to be flatter than that of their unstressed counterparts. In direct conflict with this are the findings of another study by the same authors. Ortega-Llebaria and Prieto (2010) concluded that while spectral tilt does correlate with vowel quality in Catalan, it does not correlate with stress in Spanish.

These controversial findings, as well as those presented above on intensity, might be attributed to the failure of those studies to control mouth-to-microphone distance during recording. It should be noted that intensity levels decrease significantly when sound source moves further away from microphone. Therefore, for intensity and spectral tilt being reliable amplitude measurements of vowels and syllables, the distance between the speaker’s mouth and the microphone must be kept constant. To this end, the current study used a head-mounted microphone for recording.

1.3 Research questions

The current study aims to provide solid acoustic evidence concerning the stress pattern of Spanish lexical compounds by conducting a comparative analysis of all known stress correlates across different types of linguistic structure. Three specific questions will be addressed. First, is the initial constituent of lexical compounds pronounced with stress? Second, if so, which level of stress is it and what are its phonetic cues? And last, are there any cross-compound-type variations in the prosodic behaviour of lexical compounds?

2.1 Materials

Eighteen lexical compounds representing different syntactic and semantic types were carefully selected from those used in the previous studies. Each of them was matched with a non-compound word or a phrase that has a similar segmental structure. In view of the fact that formant structure, duration and intensity are properties that vary from one segment to another, the only case in which it would be warranted to attribute differences in acoustic values to suprasegmental phenomena is when they arise in the same segmental context. Therefore, it was imperative to ensure that the compound words and their counterparts have the same phonological strings, at least in the syllables where stress cues are expected to arise. An example of such word group is given in (2) below, where the syllables to be examined and compared (i.e., the target syllables) are underlined and the primary-stressed syllables are indicated in bold.

Additionally, in the selection of words, mid to high-frequency words were preferred over low-frequency ones, utilising Mark Davies’ Corpus del Español: Web/Dialects (2016) for assistance. For example, sacapuntas, with 548 counts, was favoured over sacamuelas (‘tooth-puller’) with 51 counts. The purpose of adopting this criterion was to make it easier for participants to produce natural speech in the elicitation tasks. The complete list of the selected items is attached in Appendix A (37 items in total).

For each selected item, two tasks, each involving a scripted question and a picture, were designed (sample tasks are given in Appendix B). The first task was designed to elicit an answer using the target item in a normal declarative sentence, while the second was intended to elicit it in post-focal position, i.e., an unaccented context. Moreover, in order to avoid the interference of boundary tones, phrase accents and final-lengthening effects, the task design ensured that the target items were located away from phrase-final or prepausal position.

Tasks that incorporate pictures with partially scripted conversation were selected as our elicitation method for two reasons. Firstly, non-verbal materials such as pictures are more promising for the elicitation of spontaneous, natural and meaningful speech (e.g., Himmelmann and Ladd 2008, p. 266). Secondly, partially scripted conversation enables researchers to control which linguistic structures will be produced (Caldecott and Koch 2014, p. 225). A task of this kind is typically performed as follows: while presented with a picture, the speaker is asked to answer a scripted question based on what s/he sees on that picture. Since questions and pictures are designed ahead of time, researchers can select the lexical items to be elicited but are still able to attain natural and spontaneous productions with the help of non-linguistic prompts and conversational style.

2.2 Participants

Thirty native Spanish speakers (three male and three female from each of the following countries: Spain, Mexico, Colombia, Chile and Argentina) participated in the experiment. They were either visiting or residing in Auckland and all of them were maintaining consistent use of Spanish on a daily basis. Of these participants, 21 were university students, while the remaining 9 also possessed a high-level education background. None of them had speaking or hearing problems. Moreover, none of them had training in phonetics or linguistics; that is, all were naive speakers. Their participation in this study was approved by the Human Participants Ethics Committee of the University of Auckland (June 2017; reference number 018498).

2.3 Data collection

Data were collected in a sound-treated recording studio at the City Campus of the University of Auckland. The participants’ responses were audio-recorded using a lightweight head-mounted microphone (Senneheister HSP2), which was connected to a recorder through an in-line preamplifier. Given that head-mounted microphones allow to maintain a consistent mouth-to-microphone distance throughout the recording, they are often considered the best option for measuring intensity.

Prior to performing the tasks, the participants were asked to provide their personal information (e.g., occupation and educational background). This informal talk served to help the participants feel more comfortable with using the microphone. Prior to the formal recording, the participants were provided with instructions and a sample task. They were encouraged to practice with the sample task as many times as they wanted. They were also assured that production errors would not matter, for they could repeat if they felt they made a mistake. When unnatural or unclear pronunciations were detected, the researcher requested the participants to repeat.

2.4 Data processing

The recordings were processed using Praat (Boersma and Weenink 2019). Target items were firstly extracted from the recordings and saved as separate sound files, i.e., 2,220 tokens (37 items * 30 speakers * 2 conditions). Then, 334 tokens (which constitute 15 % of the data) were removed as they were found affected by recording or production problems such as unexpected noises, unnatural pronunciations or pauses and failure to produce deaccented words in post-focal position. The removed items included 179 lexical compounds, 121 non-compound words and 34 phrases. After removal of all the disqualified items, there were a total of 1,886 tokens.

For each of the 1,886 tokens, boundaries of syllables and syllabic nuclei were manually labelled on separate tiers in Praat textgrids. This was done manually on the basis of visual inspection of waveforms and spectrograms (display frequency range: 0–8,000 Hz, bandwidth: 260 Hz), as well as with reference to auditory evaluation. Even though manual labelling is in principle the most precise and reliable method, labelling consistency needs to be monitored. To reduce inter and intra-item labelling discrepancies, explicit segmentation criteria were set out following Machač and Skarnitzl’s (2009) Principles of Phonetic Segmentation and Hualde’s (2005) introduction to acoustic characterisation of Spanish sounds. The criteria are as follows. Vowel boundaries were placed with reference to the appearance of full formant structure in the spectrogram. When two vowels were in contact, the midpoint of the formant-transition phase in the spectrogram was taken as the separation point. As for liquids and nasals, a simpler wave shape, lower amplitude and drop in the formant structure were attended in labelling their boundaries. The segmentation criteria for fricatives were fairly straightforward: the presence of high-frequency formant columns in the spectrogram and ‘spiky’ or ‘hairy’ waveform. Absence of formant structure provides a reliable labelling indication for both voiceless stops /p t k/ and voiced stops /b d g/. However, the segmentation of the approximant allophones of /b d g/ was problematic. In cases where changes in the formant structure and the waveform did not allow unambiguous boundary placement, the boundaries were placed at the temporal midpoints of the transition phases. Moreover, to increase the intra-item consistency, the manual segmentation was performed on an item-by-item basis, rather than a speaker-by-speaker basis.

2.5 Acoustic measurements

The acoustic measurements were also made using Praat (Boersma and Weenink 2019). The procedure involved running several scripts to automatically extract information on f0, pitch slope, duration, intensity and spectral tilt from each token.

2.6 Statistical analysis

The acoustic measures that were entered into statistical analyses are summarised in Table 1 below. All the statistical analyses were performed using the statistical software R (R Core Team 2017).

To test for significant differences in the above acoustic measures among different linguistic structure types (i.e., lexical compound, non-compound word and phrase), linear mixed model (LMM) analyses were carried out using the “lmer” function from the lmerTest package, which extends the lme4 package to provide p-values for the fixed effects through Satterthwaite’s or Kenward–Roger’s degrees of freedom approximations (Kuznetsova et al. 2017). LMM analyses were performed using each of the above acoustic measures as response, with linguistic structure type as predictor, the effects of linguistic structure type within each word group as random slopes, and participants as random intercepts.^[6] Among those LMM analyses, the analyses on f0 and pitch slope were conducted on accented data only since they are irrelevant in unaccented data. For duration, intensity and spectral tilt measures, LMM analyses were carried on the full set of data, and then repeated on the unaccented data only. It should also be noted that in these analyses, the diagnostic qq-plots (using “qqnorm” and “qqline” functions) for pitch slope measures show strong evidence of outliers; outliers were excluded from the corresponding analyses with the help of the interquartile range method (IQR). Moreover, the duration measures were log-transformed due to their skewed distribution; log transformations significantly improved the normality of the duration distributions.

It is also imperative to acknowledge that conducting multiple statistical tests on the same dataset can lead to the multiple comparison problem, consequently elevating the familywise error rate (FWER), which represents the likelihood of committing a Type I error. In response to this concern, we adopted the Bonferroni-adjusted significance level of 0.005, as determined by the formula presented in (5), to effectively control the FWER.

Lastly, to assess potential variation among types of lexical compounds, LMM analyses were conducted on the acoustic measures of lexical compounds, with the types of compounds (i.e., endocentric vs. exocentric, attributive vs. coordinate vs. subordinate) serving as predictors, and items and participants as random intercepts. The Bonferroni-adjusted significance level of 0.005 was also applied in these analyses to identify statistically significant differences.

3.1 Acoustic differences among lexical compounds, non-compound words and phrases

3.1.2 Results on pitch slope

The results on pitch slope surrounding V1 are shown in Table 3 below. Phrases exhibit the highest pitch slope value, which is approximately 21 ST/s higher than that of lexical compounds and non-compound words. This is also evident in Figure 1 below, where the pitch slope values are mostly positive in phrases but generally negative in lexical compounds and non-compound words. Furthermore, if we recall the way in which pitch slope was calculated (Section 2.5.2), here the positive pitch slope values in phrases suggest that V1 of phrases are generally surrounded by rising pitch contours, whereas the negative pitch slope values in lexical compounds and non-compound words are a natural consequence of the frequent falling pitch contours surrounding their initial syllable.

Table 3:

Results of LMM on pitch slope.

Measures	Factor	b	SE	df	t	p
Pitch slope	Lexical compounds (intercept)	−23.55	3.30	13	−7.13	<0.001
	Non-compound words	0.56	2.84	10	0.20	=0.85
	Phrases	21.10	4.48	21	4.71	<0.001
	Non-compound words (intercept)	−22.98	3.16	12	−7.26	<0.001
	Lexical compounds	−0.56	2.84	10	−0.20	=0.85
	Phrases	20.54	4.50	16	4.57	<0.001

Figure 1:

Pitch slope values in lexical compounds (LC), non-compound words (NC) and phrases (Pr).

3.2 Results on cross-compound-type variations

Significant differences across different compound types have only been found in the relative H1-A2 values with a significance threshold set at 0.05, as seen in Table 10. Coordinate compounds have lower relative H1-A2 values than attributive and subordinate compounds; however, with the adoption of a Bonferroni-adjusted significance level at 0.005, these differences are no longer considered statistically significant.

4.1 Main findings

4.2 Theoretical implications

The findings have important theoretical implications. To begin with, the findings support the use of stress as a diagnostic for compoundhood (Lieber and Štekauer, 2009, p. 8). What makes this possible is the presence of stress on each constituent of lexical compounds. This prosodic property differentiates lexical compounds from non-compound words, which in Spanish, are limited to a single word stress. A prosodic diagnostic will be useful when dealing with borderline cases, such as telenovela ‘soap opera’, autoescuela ‘driving school’, fotocopia ‘photocopy’, and the like. In such words, the morphological status of the initial constituent is not clear. If that constituent is regarded as a root, the overall structure would be a compound word, but if that constituent is regarded as a derivational affix, the overall structure would be a non-compound word (Gifre 1984, p. 223; Rainer and Varela 1992, p. 121). Future experimental studies will be able to settle this issue by applying the prosodic diagnostic.

We also gain insight into the prosodification of Spanish compounds. The finding that Spanish lexical compounds are produced with a secondary stress on the initial constituent and a primary stress on the final constituent rules out some options. They cannot be represented as phonological phrases (PPhs) because that would require retention of primary stress on each constituent word. Their representation as prosodic word (PW) is not viable either, as that would overlook the fact that this language assigns only one stress per PW. An alternative to consider is the representation of Spanish lexical compounds as instances of prosodic word recursion (e.g., Itô and Mester 2007, 2009, 2012; Selkirk 2011). On this view, the compound would be a recursive PW, within which there would be two embedded PWs. These units would be prosodically asymmetrical. The second would function as the head, in virtue of which it would retain its primary stress. The first embedded PW would, by contrast, be a subordinate to the head, and thus, would inevitably have its stress demoted from primary to secondary. In a phonological model where PW recursion is not allowed, the option would be to represent the compound as an intermediate constituent between the PW and the PPh, e.g., a prosodic word group (Vigário 2003, 2010), or a composite group (Vogel 2009, 2010, 2020). The PWs arising from the compound constituents would then be embedded within this new prosodic category. This analysis would also involve prosodic subordination: the second embedded PW would be the head, in virtue of which it would retain its primary stress, whereas the first embedded PW would be a subordinate to the head, and thus, its stress would be demoted from primary to secondary.

4.3 Directions for future research

Although the current study offers fresh insight into the prosody of Spanish lexical compounds, it leaves a number of gaps to be filled in the future. Firstly, while having identified acoustic differences in the prosodic behaviour among non-compound words, lexical compounds and phrases, this study defers to subsequent research the task of determining the perceptual salience of these differences. This inquiry into perceptual salience is crucial for understanding not only how these prosodic variations are processed cognitively by listeners but also how phonetic cues contribute to the distinction of different linguistic structure types.

Secondly, recall that, for the purpose of obtaining natural speech, infrequent items were left out during word selection. As a result, the question of whether lexical compounds with relatively low usage frequencies exhibit the same stress pattern or not cannot be answered here. Lang (1992) suggests that the frequency of use is a useful criterion that helps to distinguish compounds from other types of linguistic structure (p. 94). This may imply that less frequently used compounds could present variations in stress pattern not observed in this study, highlighting the need for further investigation into how frequency impacts the prosodic behaviour of compounds.

Lastly, despite the implementation of measures to render the elicited speech as naturalistic as possible, it remains categorised as laboratory speech, which diverges from spontaneous speech in several phonetic dimensions (Face 2003). Therefore, subsequent research is necessary to examine compound stress patterns within spontaneous speech to determine the extent to which these findings are replicated in more naturalistic linguistic contexts.