Accounting for effects of monitored speech: vowel nasalization, nasal consonant weakening and /s/ production in Spanish

2.1 Variation in /s/ production: region and speech style

Weakening of /s/ is undoubtedly the most studied phonological process in Spanish linguistics. Succinctly put, /s/ exhibits tendency towards debuccalization and deletion in the coda environment, both word-internally and across word boundaries, especially when preceding a consonant. The examples in (1) provide a simple sketch of the variation. One crucial distinction is the presence (pe[s]te or pe[h]te) versus absence of frication (pe[ø]te, or other attested variants with no frication such as gemination of the post-fricative consonant, e.g., pe[t̪t̪]e, or lengthening of the preceding vowel, e.g., p[e:]te). Variation in /s/ production, however, is complex, influenced by phonological, regional, social, and stylistic factors. This section does not intend to provide an exhaustive account of these patterns, but rather to zoom in on the relevant facts. For this reason, I concentrate on two aspects of the variation in word-final /s/ production: regional (Argentine vs. Dominican Spanish), and stylistic.

Both Argentine and Dominican Spanish are s-weakening dialects, albeit to different degrees. In the Dominican Republic, weakening of /s/, whether as aspiration or deletion, is the norm, regardless of educational attainment (Alba 1982, 1990; Henríquez Ureña 1940; Jiménez Sabater 1975; López Morales 1990; Núñez-Cedeño 1980; Terrell 1986). López Morales (1990: 130) reports deletion rates ranging from 98 % among illiterate speakers to 66 % among college-educated participants. He does not report aspiration rates, which he considers it as instances of retention of /s/. Alba (1990: 60), however, reports 90 and 25 % deletion rates for the same populations, respectively. In Argentina, the scope of weakening has been more limited, with low rates reported, at least among educated speakers. Terrell (1978: 44) reports aspiration at a 40 % average rate and deletion at 14 %. Thus, a key difference between these dialects is the rate of /s/ deletion, which becomes even more disparate in educated speech. In fact, when compared to Dominican Spanish, Argentine Spanish remains a somewhat conservative dialect in this respect (Hualde 2014: 159).

Regional variation aside, /s/ production is also susceptible to the effects of (broadly defined) speech style. Following a Labovian sociolinguistic model, weakening is typically more prevalent when speakers pay little attention to speech, as is the case with informal speech in sociolinguistic interviews (Labov 1972). When attention to speech is encouraged (e.g., when style becomes more formal), as is the case with a word-list task, ‘standard’ or prestige variants (i.e., in many speech communities, mostly retained [s], or possibly [h]; crucially not [ø])^[1] increase in frequency of use. Of relevance to the present study is Alba (2011), which investigated speech patterns among television announcers and reporters in the Dominican Republic. He reported that even though weakened variants – [ø] and [h] – are prevalent across all social groups, newscasters in Dominican television produce fully retained variants of /s/. The use of cameras, studio lights, and a wide and public audience elicit a more formal and careful speech style from these speakers. To this end, Alba (2011) analyzed ten TV news segments of 10 min each to understand the patterns of variation. A key insight from this study was that people on television (including interviewees, reporters, and news anchors) retained /s/ at a rate of 63 %, five times the rate that is observed in spontaneous learned speech (or habla culta, the targeted audience for news reporting). In fact, when only news reporters were considered (that is, no interviewees), there was no variation, with retained /s/ being the only variant observed (i.e., no [h] nor deletion). Additionally, a qualitative analysis of a few tokens of what some could label as an “exaggerated” /s/ showed an average duration of 163 ms, compared to 93 ms for spontaneous learned speech. Regarding newscaster /s/ in other /s/-weakening dialects (including Argentina), Alba (2011) noted that the most frequent variant was typically [h], and not full [s]. This is confirmed by Colantoni and Hualde (2013: 29) who, writing about the phonetics and phonology of Argentine Spanish, make the observation that Argentine newscasters default to aspiration, unlike what may be found in other areas, such as Southern Spain and the Hispanic Caribbean. Thus, the repertoire of newscaster speech in some [s]-weakening dialects includes lenited variants such as [h]. In the Dominican Republic, this does not appear to be the case. The findings in Alba (2011) suggest that one way in which attention to speech surfaces among Dominicans may be by means of a fully retained /s/, even though it may not be the expected realization elsewhere in Dominican society.

2.2 Regional variation in nasal consonant weakening and anticipatory vowel nasalization

The nasal system in Spanish includes three consonant phonemes that contrast in place of articulation. As shown in (2), bilabial, alveolar and palatal nasal consonants freely occur in the syllable onset.

In the coda, only one nasal is permitted, /n/, which typically exhibits contextual and regional variation. As is common cross-linguistically, Spanish coda /n/ assimilates to the place of articulation of the consonant that follows (whether word-internally or across a word-boundary), such that six allophones are typically described in this domain, as shown in (3).

Additionally, coda /n/ exhibits considerable regional variation, as shown in (4). In broad terms, dialects of Spanish belong to one of two groups with regard to this variable (Canfield 1981; Lipski 1994; Resnick 1975). A first group, which I label ‘[n]-dialects’, exhibit a tendency towards alveolar variants of /n/ in word-final position, as shown in examples (i) in (4). The exception is the case preceding a velar consonant, as in 4(c)i (following the assimilation process explained above). There is a second group, ‘[ŋ]-dialects’, which opt for the velar variant in all contexts, as seen in examples (ii) in (4), even when not preceding a velar consonant, as is the case with 4(a)ii, 4(b)ii and 4(d)ii (Bakovic 2001; Cedergren and Sankoff 1975; D’Introno and Sosa 1988; Haché de Yunén 1981; Hyman 1956; López Morales 1980; Muñiz-Cachón 2003; Salvador 1987; Terrell 1975; Trigo Ferre 1988).

Velarization of /n/ (examples [ii]) is traditionally viewed as part of a process of nasal consonant attenuation, that by way of consonant deletion, leads to nasalization of the pre-nasal vowel. As such, [ŋ]-dialects exhibit greater anticipatory vowel nasalization (though it is not clear whether “greater” means longer time-course or larger velar port opening) compared to [n]-dialects. Given these facts, most research on word-final nasal consonant weakening and anticipatory vowel nasalization has understandably focused on [ŋ]-dialects in the Caribbean region. Whether from a formal or a sociolinguistic perspective, empirical (albeit impressionistic) research has confirmed several trends. Firstly, nasal consonants are prone to loosening of the oral constriction, particularly in the word-final domain. Backing of /n/ (i.e., [ŋ]), on the other hand, is the most common variant. Anticipatory vowel nasalization surfaces with nasal consonant deletion. Finally, given the high frequency of nasal consonant deletion for some [ŋ]-dialects, it is generally accepted that nasalization in the pre-nasal vowel may have conventionalized and putatively phonologized. That is, nasalization has become a feature of the vowel and no longer a marker of nasal consonant weakening.

In the last two decades, instrumental work has also enriched the research landscape by providing detailed phonetic characterizations of variation patterns, as well as adding comparisons to [n]-dialects. For example, electropalatographic studies have confirmed more back articulations of the nasal consonants in [ŋ]-dialects, in addition to lesser degrees of constriction, when compared to [n]-dialects (Colantoni and Kochetov 2012; Ramsammy 2013). Using a nasograph,^[2] Lederer (2000, cited in 2003) probed regional differences by comparing the time-course of nasalization in an [n]-dialect (Peninsular Spanish) and an [ŋ]-dialect (Cuban Spanish). Following Solé (1992), who showed that the onset of nasalization of phonetic/coarticulatory (Peninsular Spanish) and phonologized (American English) nasalization are different, Lederer found that the patterns of anticipatory nasalization in the [ŋ]-dialect were more similar to those of English (which exhibits fully nasalized pre-nasal vowel allophones), than to those of the [n]-dialect.

The crucial link between nasal weakening and anticipatory nasalization is the potential for realignment of oral and nasal gestures, a well-documented historical and synchronic pattern (e.g., Beddor 2009; Hajek 1997; Zhang 2000): as weakening of the nasal consonant increases in magnitude, so does the temporal extent of nasalization in the pre-nasal vowel. However, investigating this relationship among dialects of Spanish, whether [n]- or [ŋ]-, has not been possible so far due to the nature of electropalatography and nasography. Electropalatography is unable to capture velum lowering, while nasography is unable to characterize the degree of constriction. Recent work employing nasometry has permitted such an analysis, to which we turn next.

Building on the cross-dialectal findings in the extant literature, Bongiovanni (2021a, 2021b) examined nasal consonant weakening and anticipatory vowel nasalization in Santo Domingo Spanish (Dominican Republic, an [ŋ]-dialect) and Buenos Aires Spanish (Argentina, an [n]-dialect). Results in Bongiovanni (2021a) showed earlier onset of nasalization for the Santo Domingo data compared to Buenos Aires Spanish. Bongiovanni (2021b), on the other hand, correlated the nasalization findings with weakening of the word-final nasal consonant. A crucial – and unexpected – finding was that earlier onset of nasalization obtained with little – and arguably without – weakening of the nasal. Specifically, findings revealed similar degrees of nasal consonant weakening (measured in terms of duration and consonant-to-vowel ratio) across dialect groups, unlike what would have been predicted based on previous research. These findings suggested that differences in nasalization cannot solely be attributed to weakening of the nasal consonant. Thus, this instrumental work provided experimental support to the hypothesis of phonological differences across dialects: for the Santo Domingo data, nasalization may no longer be a marker of nasal consonant weakening, but a feature of the pre-nasal vowel allophone.

The lack of nasal consonant weakening, particularly in the Santo Domingo corpus, requires further explanation. Bongiovanni (2021b) hypothesized that the findings were the result of heightened attention to speech, due to a combination of factors, including the use of specialized equipment, a reading task, well-known differences between dialect groups regarding the evaluation of their own variety (Büdenbender 2010; Bugel 2012; Llull and Pinardi 2014; Toribio 2000), and an out-of-community researcher for the Dominican group. There were also more reading errors in the Dominican than in the Argentine group. Additionally, some Dominican participants exhibited dialect “inappropriate” variants, such as the use of the interdental fricative, [θ], for the graphemes <c, z> (a total of 13 tokens, or 0.1 % of the corpus, produced by four speakers: three women and one man), which suggests a perceived “supra-regional” norm.^[3] Far from (negatively) evaluating participants’ performance, I take these pieces as anecdotal evidence of participants attempting a ‘more formal’ speech style or, at the very least, heightened attention to speech.

4.1 Participants and equipment

This study compares Buenos Aires (Argentina) and Santo Domingo (Dominican Republic) Spanish. Data come from 54 native speakers from these metropolises (Santo Domingo, n = 28, aged 18–27 years old, 18 women and 10 men; Buenos Aires, n = 26, aged 19–28 years old, 18 women and 8 men). Except for two who did not answer, all Dominicans lived in Santo Domingo. Four participants reported having also lived in Jarabocoa (n = 1, until the age of 10), Monseñor Nouel (n = 1, until the age of 16) and Monte Plata (n = 2, both until early adolescence). All Argentine participants, with the exception of two who did not provide this information, reported living in Buenos Aires (or the greater metropolitan area) for their entire lives.

All participants were university students at the time of data collection. While working with young, educated participants raises questions about the representativeness of the sample and, thus, the generalizability of the results to other populations, this choice was made to ensure comparability across dialect groups: even as Santo Domingo and Buenos Aires differ in territorial and population size, each serves as the political, economic, and cultural center of its respective country.

Recordings were made with a Glottal Enterprises Nasometer, a split set of microphones divided by a plate that rests above the speaker’s lip. This tool allows for stereo recordings, with separate tracks for oral and nasal cavities respectively, and is ideal for analyses of nasalization (see spectrogram in the next subsection). Given its lightweight, portable nature and lack of calibration requirements, this device is particularly well-suited for field data collection, as employed in this study. Recordings were done in Praat (Boersma and Weenink 2016), in stereo, and sampled at 44.1 kHz.

4.2 Materials

Target stimuli were real words, chosen according to a number of linguistic variables: height of the pre-nasal vowel (/a e o/), stress (stressed, unstressed), syllable type (CVC, CVN, NVN, NV) and whether or not the nasal consonant was in an assimilatory context (pre-consonantal vs. pre-vocalic/pre-pausal). Note that these variables will not be analyzed here, as they were the object of study in Bongiovanni (2021a, 2021b). The focus of the present paper is the comparison of CVC and CVN. In order to keep the task from feeling too repetitive, target words were embedded in noun and verb phrases (e.g., el capitán tacaño ‘the stingy captain’) rather than reusing the same carrier sentence (see Appendix for a full list of elicitation materials).

Elicitation materials were displayed in Microsoft PowerPoint on the same laptop used in recording (a MacbookPro with a 13″ screen). Slides were randomized and set up to advance automatically in order to standardize speech rate across repetitions (to the extent possible).

4.3 Annotation and instrumental analysis

A total of 7,639 tokens were extracted, of which 3,812 were CVN and 3,827 were CVC. Two hundred and forty-two (242) were eliminated from analysis primarily due to misreading and disfluencies, thus 7,397 were analyzed in total. Praat was used for segmentation and analysis. Given that recordings were in stereo, with each channel recording material from different cavities in the vocal tract, the annotation and segmentation of target sounds utilized cues in each channel, as well as the spectrogram. Figure 1 presents spectrogram and waveform samples illustrating the target segments. The onset of the vowel was marked at the beginning of periodicity in the waveform and consistent formant structure in the spectrogram. With regard to the nasal consonant, a sudden absence of audio in the oral channel (Channel 1) and presence of intensity in the nasal channel (Channel 2) were taken as indicators of onset of the nasal consonant. To segment /s/, the converse criteria were employed (i.e., presence of intensity in the oral channel and a flat waveform in the nasal channel), as well as an aperiodic waveform typical of fricative sounds. See below procedure for when no oral constriction was identified (that is, deleted tokens), whether /s/ or /n/.

Figure 1:

Waveform and spectrogram of the words sartén ‘skillet’ (top) and veintidos ‘twenty-two’ (bottom), produced by a Santo Domingo speaker.

Weakening of the coda consonant, oral and nasal, was analyzed in terms of several acoustic cues. Following more recent research on Spanish fricatives (e.g., Brogan and Bolyanatz 2018; Erker 2010, 2012; File-Muriel and Brown 2011), acoustic cues were preferred over impressionist coding as they allow the researcher to capture production of /s/ without over imposing categories. Weakening of /s/ was examined in terms of three cues: duration, center of gravity (COG) and zero-crossings. Center of gravity characterizes the frequency at which sonic energy is maximally concentrated; it is the mean frequency of the spectra over a duration of time. Lowering of COG is interpreted as a weakening tendency (File-Muriel and Brown 2011: 228). In a similar vein, the rate of zero-crossings characterizes acoustic differences between productions of /s/. Compared to [h], [s] tends to have energy concentrated in a high frequency range, resulting in a higher zero-crossing density (Ruch and Harrington 2014). For this reason, the number of zero-crossings was divided by the duration of the frication interval. Extraction of sibilant acoustic parameters were automated with a Praat script (Elvira-García 2014). For deleted observations, duration and rate of zero-crossings were set at zero. A zero value for COG is not interpretable, and for this reason, deleted tokens were removed from the COG analysis.

With regard to /n/, for simplicity, I only examine weakening in terms of temporal reduction (i.e., duration).^[4] When no oral constriction was identified, duration was established at zero. In the case of nasalization, the nasal track was used as a proxy for velum lowering: when the velum is lowered, the acoustic signal increases in the nasal track. Thus, energy measurements (in pascals) were taken at 20 equidistant time points during the course of the pre-nasal vowel and another 20 during the nasal consonant, automated with Praat scripts written by the author for the purpose of this research. Energy measurements were then smoothed using a 5-point window. For each individual token, the difference between the highest and the lowest energy reading in the vowel-nasal sequence was calculated. A 15 % of this difference was established as noise band. Onset of nasalization was operationalized as the first time point at which the nasal energy surpassed the 15 % noise band (Bongiovanni 2021a, 2021b; Delvaux et al. 2008; Solé 1992). Given that consonantal duration and vowel nasalization values vary as a function of vowel type, stress, and environment (Bongiovanni 2021b; Clumeck 1976; Krakow 1999), prior to running any statistical analyses, averages of duration and onset of nasalization were calculated per speaker, vowel type, environment, and stress conditions. That is, averages were calculated within speaker and within conditions.

All visualizations in this article were done in R (R Core Team 2019), using the tidyverse suite of packages (Wickham et al. 2019) and the lmer() function (Bates et al. 2015). Statistical modelling was also done in R (details are provided in the next section). Prior to statistical analysis, duration and COG measurements were normalized using the scale() function (Winter 2019).

5.1 Acoustic characterization of /s/

Given the documented regional and stylistic variation in /s/ production, I first present the acoustic characterization of /s/ production in both dialects. Prior to examining the three acoustic correlates (duration, COG, and rate of zero-crossings), the data was explored for instances of deletion. The Buenos Aires data showed 37 observations of elided /s/, which represent a 2 % application rate. The Santo Domingo data, on the other hand, presented 102 observations of deleted /s/, a frequency of occurrence of 5.6 %. The rate of deletion is in keeping with the previous literature documenting that Santo Domingo Spanish would exhibit a higher frequency of deletion. Nevertheless, the percentage of use of deletion was well below the numbers reported in the literature for either dialect (c.f. Alba 1990; Terrell 1978).

Regarding the acoustic parameters of /s/ production, Figures 2 through 4 illustrate the distribution of the data, per dialect group, for duration, COG and rate of zero-crossings, respectively.

Figure 2:

Boxplot and violin plot of (raw) duration of /s/ in ms, per dialect group. Diamonds represent the mean.

Figure 3:

Boxplot and violin plot of (raw) COG measurements, per dialect group. Diamonds represent the mean.

Figure 4:

Density plot of zero-crossing rate, per dialect group. Diamonds represent the mean.

Starting with duration, the dialects exhibited great acoustic overlap in the temporal dimension (Figure 2). Contrary to what the previous literature would predict, Santo Domingo Spanish exhibited considerably longer productions of /s/ (Santo Domingo, M = 87.3, SD = 74.4; Buenos Aires, M = 79.6, SD = 95.4). Figure 3 showed that the distribution of COG across dialect groups was skewed in opposite directions, with higher values for Santo Domingo compared to Buenos Aires (Santo Domingo, M = 4,701, SD = 1,151; Buenos Aires, M = 3,411, SD = 1,413). Zero-crossing rates in Figure 4, on the other hand, showed much a wider distribution (i.e., more variability) for the Santo Domingo data in comparison to Buenos Aires, which presented data tightly concentrated towards the lower end of the range (Santo Domingo, M = 42, SD = 27.8; Buenos Aires, M = 24.5, SD = 20.4). Welch two-sample t-tests, one per acoustic parameter, showed that group differences were statistically significant (duration: t (3,490.2) = −2.7, p < 0.01; COG: t (3,455.5) = −29.8, p < 0.001; zero-crossing rates: t (3,232.3) = −24.3, p ≤ 0.001).

An important observation is that the three acoustic parameters represent different acoustic manifestations of /s/ weakening. Therefore, we would expect to see evidence of collinearity. To this end, a Pearson’s correlation test between the acoustic measurements was run. Duration of /s/ and COG were weakly correlated, r (3,675) = 0.17, p < 0.001. Duration of /s/ and rate of zero-crossings were moderately correlated, r (3675) = 0.32, p < 0.001. COG and rate of zero-crossings were strongly correlated, r (3536) = 0.6, p < 0.001.

In closing this section, there are two main take-aways from these data. First, findings suggest that productions of /s/ were significantly more lenited for Buenos Aires than for Santo Domingo (despite the latter dialect exhibiting more instances of deletion). This finding offers some evidence that Santo Domingo speakers’ attention to speech was heightened compared to that of Buenos Aires. To fully diagnose this scenario a comparison with these participants’ speech in other communicative situations (such as sociolinguistic interviews) would be necessary. The lack of such comparison does not hinder the analysis. For the Santo Domingo data, these findings are in stark contrast with the previous literature of “rampant” weakening. Second, the fact that COG and rate of zero-crossings showed the same patterns as observed in the temporal domain indicates that results cannot solely be due to speech rate differences.

5.2 Vowel nasalization, nasal consonant weakening and /s/ production

Prior to presenting the analyses, two reminders are in order. First, averages were calculated per speaker and per phonological context (vowel type, following segment, stress). That is, each speakers’ productions of the word-final consonant in, for example, C-án#t (/a/, stressed and in pre-coronal position) were averaged together. Each point in the scatterplot represented values for the exact same phonological context. Thus, average values of duration of /n/ in C-án#t were compared to duration of /s/ in the exact same context (i.e., C-ás#t) for that particular speaker. In other words, the measurements indexed by each point were taken in equivalent contexts. Second, for deleted consonants, duration was set at zero.

Due to the present of collinearity between acoustic parameterse of /s/, I will focus on duration of /s/ in relation to nasal consonant weakening and onset of nasalization. The correlation between duration of /s/, COG and ratio of zero-crossings renders the information redundant, potentially complicating the interpretation of results. Additionally, the choice to focus on duration in the analysis aims to maintain a temporal comparison with /n/ and nasalization. By homing in on the duration of /s/, I aim to streamline the analysis and avoid the challenges associated with disentangling the unique contributions of each parameter in the context of the study.

Figure 5 presents scatterplots of average duration of /s/, as a function of the average durations of the nasal consonant (top) and onset of nasalization (bottom). Starting with the relationship between duration of /n/ and duration of /s/, visual inspection of the data showed fewer tokens towards the left of the scatterplot (the side that would reflect shorter sibilants) than expected, considering that both dialects are /s/-weakening. Another readily visible trend is the positive relationship between weakening of /s/ and weakening of /n/. In other words, longer productions of /s/, arguably signaling more speech monitoring, co-occurred with longer nasal consonants.

Figure 5:

Duration of /n/ (top) and onset of nasalization (bottom) as a function of duration of /s/.

To test whether the relationship between duration of /n/ and duration of /s/ was statistically significant, a mixed effects linear regression model was run, with duration of the nasal consonant as the dependent variable, duration of /s/ and dialect group as the fixed effects, duration of /s/ by dialect group as interaction term, with a by-speaker random intercept. Buenos Aires Spanish was the reference level. The results are summarized in Table 1. Results indicated a fixed effect for duration of /s/ in both dialects. Dialect group and the interaction term with duration of /s/ were not statistically significant. These results indicated that longer durations of /s/ co-occurred with longer durations of /s/. As a reminder, given that other acoustic parameters of /s/ were collinear with duration, this is a result that cannot be exclusively attributed to speech rate differences.

Moving on to the relationship between onset of nasalization and duration of /s/, visual inspection of the data suggests a positive relationship for Buenos Aires, but not for Santo Domingo. That is, for Buenos Aires Spanish, longer word-final /s/ co-occurred with earlier onset of nasalization in pre-nasal vowels. For Santo Domingo Spanish, on the other hand, there appeared to be no relationship, whether positive (both measurements move in the same direction) or negative (as one decreases, the other increases). A second mixed-effects linear regression model was run. The onset of nasalization was entered into the model as dependent variable; fixed effects included duration of /s/, dialect group, and the interaction term between duration of /s/ and dialect group. Speaker was entered as random effect. Buenos Aires Spanish was the reference level here as well. Table 2 summarizes the results.

Findings indicated that mean duration of /s/ and dialect group were significant main effects, as was the interaction term between the two. The pattern of earlier onset of nasalization co-occurred with longer durations of /s/ for Buenos Aires Spanish, but the same effect was not true for Santo Domingo Spanish.