1.1. Consonant lenition in Spanish and the Canary Islands dialect

In Spanish, the typical manifestation of intervocalic consonant lenition is the approximantization of /b d g/. In most Spanish dialects, the degree of openness of the resultant sound differs depending on such factors as stress, place of articulation, quality of the flanking vowels, speech rate, etc. (see Hualde, 2005 for an overview).¹ It is also possible to see /b d g/ delete completely in some cases.

Some dialects of Spanish have also developed a further change – weakening of the /p t k/ series. Canary Islands Spanish is perhaps the most notorious in this respect, with voiceless stops undergoing voicing and (variably) approximantization post-vocalically (Oftedal, 1985; Trujillo, 1980).

Furthermore, while voiceless stop lenition is restricted to post-vocalic environments, /b d g/ also weaken after other sounds. The common generalization in the literature is that /b d g/ are phonetically realized as stops after pauses, nasals and after the lateral /l/ (in the case of /d/), and as approximants elsewhere (Navarro Tomás, 1918; Lozano, 1979; Lipski, 1996; Hualde, 2005). Moreover, the degree of approximantization of the resultant sound depends on the preceding segment: the greater its constriction, the more constricted the target consonant (Machuca, 1997; Martínez-Celdrán & Regueira, 2008). Preceding fricatives, especially a preceding /s/, both when retained and when weakened to [h], typically trigger a less approximantized pronunciation of the underlying voiced stops (Eddington, 2011).

However, there have been reports on varieties of Spanish in which stop allophones of /b d g/ rather than approximants are preferred after consonants. Such a pattern can be seen in the Spanish spoken in Costa Rica, Nicaragua, El Salvador and Panama (Fernandez, 1982; Amastae, 1989; Lipski, 1996; Quesada, 1996; Carrasco et al., 2012, see Quesada, 2010 for a review), in Judeo-Spanish (Hualde, 2013) and in the Colombian highlands (Canfield, 1962).

It is not entirely clear which of the weakening trajectories is followed by Canary Islands Spanish. Previous literature suggests that the dialect may be closer to the abovementioned Latin American varieties. Focusing on the behavior of /b d g/ after /s/, several researchers observed that the sibilant is usually deleted in these contexts and a stop rather than an approximant is produced following this deletion in the Spanish of Gran Canaria (Trujillo, 1980; Almeida, 1982; Dorta & Herrera, 1993; Herrera, 1997; see also Hualde, 2005). Additionally, a compensatory durational effect takes place: The resultant stop is longer and even geminated according to some sources, marking the plural in the absence of the /s/ (Trujillo, 1981; Almeida, 1982). Phonologically, it has usually been interpreted as tense, with the notation [b: d: g:] used to refer to it.²

As for the comparison with other contexts, a study by Herrera (1997) looks at differences in VOT and duration of /b d g/ after deleted /s/ compared to /b d g/ after vowels and nasals. It is unclear, however, how /b d g/ behave after pauses and other consonants, especially aspirated /s/. [h] is reported in Almeida (1982) as a minority pronunciation of the /s/ that is followed by lax [b d g] unspecified for continuancy, as opposed to tense non-continuants [b: d: g:] observed after deleted /s/. These may be interpreted as approximants, which would mean that some approximantization takes place after fricatives in the dialect. There appears to be variation, however, as Almeida also lists [h] followed by tense [b: d: g:] as minority variants. Furthermore, a recent study of spontaneous speech from Gran Canaria shows some variability in voiced stop productions in the context of a preceding consonant (Broś et al., 2021). Here, preceding sound deletion diminishes the probability of weakening. Approximantization still occurs, but it is much less frequent (a drop from 95% to 30% on average). The study does not provide a comparison with pronounced preceding consonants, however, so we are again unable to say whether the same decrease in weakening rates would occur e.g., after an aspirated /s/.

To summarize, stop lenition in Canary Islands Spanish concerns both voiced and voiceless stops. Typically, /p t k/ are either voiced or voiced and approximantized, while /b d g/ are approximantized or deleted in post-vocalic position (Oftedal, 1985; Almeida & Díaz, 1988). Additionally, in post-vocalic contexts derived by preceding sound deletion, weakening is largely diminished or even blocked. There is also some indication that /b d g/ do not weaken after a preceding aspirated /s/, which would place Canarian Spanish next to Nicaraguan, Costa Rican and other varieties that diverge from the mainland Spain pattern (cf. Hualde, 2005). Table 1 presents a summary of these generalizations, listing the most frequent as well as minority realizations in each context.³

Given the data and the current state of knowledge concerning the occurrence of stop weakening in the Canarian dialect, it is worth delving into the role of the left-hand environment. More specifically, this study will investigate the effect of a preceding consonant, pronounced or not, on the constriction of the target stop.

1.2. The roles of lower and higher-level prosody in lenition

As mentioned in Section 1.1., stop lenition has been shown to be influenced by various factors other than the segmental environment. In this paper, we want to take a closer look at the prosodic effects, as not all their aspects have been explored to date. Previous research reported, for instance, that more lenition occurs in /b d g/ when the sound is preceded by a stressed syllable (Colantoni & Marinescu, 2010; Cole et al., 1999; Ortega, 2004). At the same time, other studies report sensitivity to the stress of the following vowel. According to Hualde et al. (2011), there is a significant increase in the voicing of /p t k/ in Iberian Spanish in unstressed syllables. Carrasco et al. (2012) found a similar effect of stress on /b d g/ in Costa Rican Spanish. Eddington (2011), on the other hand, showed that the following unstressed vowel is associated with more lenition while the following stress diminishes the rate of approximantization in /b g/ in several varieties of Spanish. Finally, studies on Canarian Spanish show that the weakening of both /p t k/ and /b d g/ is promoted by the following unstressed vowel (Broś & Lipowska, 2019; Broś et al., 2021).

Although previous studies focused on some word-level factors other than lexical stress, such as the presence of a word boundary (Cole et al., 1999; Eddington, 2011; Katz, 2016), to the best of our knowledge there has been no study so far that focuses on prosodic effects beyond the word domain.⁴ Since lenition affects segments inside words and across word boundaries in Spanish and has been shown to be sensitive to stress, it is worth investigating higher-level prosody and phrasing effects.⁵

Such investigations should be informed by the accentual characteristics of Spanish. To provide some background, in Spanish, the main stress usually falls to the rightmost part of the sentence. Thus, the nuclear stress position is typically the last content word of the intonational phrase (Beckman et al., 2002; Hualde & Prieto, 2015). This is in line with the so-called Nuclear Stress Rule (NSR) by which nuclear stress is assigned to the rightmost lexical category of a sentence (Zubizarreta, 1988). Moreover, the Focus/Prominence Rule (FPR) requires that the focused constituent be the most prominent one. In a broad focus statement, where none of the words are under special emphasis, all lexical items are equally highlighted (Ladd, 1980) and stress is typically marked with an F0 rise in prenuclear position, with each peak being slightly lower than the previous one in the sentence (a pattern referred to as downstepping), until final lowering by which the pitch returns to the baseline level at the end of a neutral declarative sentence. This general rule has been confirmed for many varieties of Spanish (Face, 2000, 2001; Face & Prieto, 2007; Vanrell & Fernández-Soriano, 2018).⁶

Furthermore, it has been shown that Spanish should be classified as a stress accent language (Beckman, 1986), which uses duration in addition to pitch to signal accents. Thus, the nuclear position is metrically the strongest and its phrasally determined prominence is often associated with final lengthening (Ortega, 2006; Ortega & Prieto, 2007). Consequently, to explore the effects of phrasal/sentence stress on the lenition patterns of stops, we decided to design a study with sentences in which stops in stressed syllables would be placed either in prominent (focus) or non-prominent words. Syntactically, the right edge of an IP should fall on the object of an SVO sentence, and it is the object of the sentence that is the expected location of nuclear stress in Spanish. Thus, we determined the prominence of a given word based on the syntax of the sentence and compared sentence-final lexical items with those placed before the verb, toward the beginning of the utterance. The assumption of the study is that since stressed syllables inhibit lenition and promote more stop-like pronunciations, focus position should have an even greater effect on consonantal constriction.

1.3. Stop lenition – acoustic correlates

Stop lenition in Spanish and other Romance languages has been explored to date using a series of acoustic correlates. Some of them are qualitative or categorical, e.g., presence/absence of a burst in stops (Dalcher, 2008), presence/absence of formants on the spectrogram that helps identify approximants as opposed to stops or fricatives (Recasens, 2016; Katz & Pitzanti, 2019) or voicing (Herrera, 1997; Dalcher, 2008; Romero et al., 2007). Continuous measurements, on the other hand, typically include sound and/or constriction duration (Dalcher, 2008; Hualde et al., 2011; Lavoie, 2001; Parrell, 2011) and acoustic intensity-based calculations (Carrasco et al., 2012; Hualde et al., 2010; Hualde & Nadeu, 2011, Hualde et al., 2012). The latter is perhaps the most widely tested and accepted method of measuring the degree of constriction and hence weakening of consonants. Moreover, acoustic intensity has been shown to correlate quite well with lip aperture measured during articulation (Parrell, 2010). Thus, although intensity is not an articulatory measurement, it can, in principle, be used to determine consonant aperture. We therefore decided to use it as a baseline criterion for detecting different degrees of lenition in the present study.

While different ways of using intensity have been applied to date (intensity ratio, intensity difference and maximum velocity), intensity difference is by far the most popular as a marker of relative constriction degree. We will therefore focus on this parameter. In this case, two alternative measurements can be considered. Since stop lenition usually happens in intervocalic position, both a comparison with the preceding and the following vowel can be used. The intensity difference is obtained by subtracting the minimum intensity of the consonant from the maximum intensity of the vowel. In our study, we follow Martínez-Celdrán and Regueira (2008), Figueroa and Evans (2015) and Broś et al. (2021) in calculating the intensity difference with respect to the preceding vowel. We also consider the alternative measurement using the following vowel, as done by Hualde et al. (2011), Carrasco et al. (2012) and others. However, since we are interested in the effect of stress and accent (see Section 1.2.), we want to avoid confounding lenition with changes in the intensity of the vowel caused by its prosodic status. Thus, we believe that taking intensity from the preceding vowel is more suitable. We return to this issue briefly in the Discussion.

1.4. Ways of measuring aperture

While the obvious way of measuring aperture directly is electromagnetic articulography (EMA), its major disadvantage is that it is both time-consuming and resource intensive. For this reason, articulography studies are usually limited to just a few participants. Moreover, while all places of articulation can be studied via EMA, it has been shown that measuring constriction degree with this method can be quite complicated for velars and coronals. More specifically, Romero (1995: 39–40) reports that in the case of velars, no direct estimate of constriction degree can be determined for anatomical reasons, including the way soft palate is used during the production of velar sounds. Thus, constriction degree must be estimated indirectly by comparing the spatial positions of the articulators during the production of different velar sounds. In the case of coronals, using the very tip of the tongue instead of the blade may make it more difficult to obtain satisfactory measurements. Nevertheless, once these obstacles are overcome, consistent patterns of constriction degree and location can be found across subjects. To avoid problems with determining the constriction degree, Parrell (2011) decided to focus on labials only in his study on Spanish stop weakening.

To provide a measurement of constriction degree using EMA, the Euclidean distance between the sensors placed on the articulators, e.g., the lower and upper lip in the case of labials, is calculated. In his comparison of /p/ and /b/ productions in Spanish across different contexts, Parrell (2011) demonstrated that the articulatory data point to a difference in target constrictions between the voiceless and the voiced stop. The latter has a more open (i.e., less constricted) target, although it is still beyond the point of closure. The articulatory parameters show that combined with a decreased duration, this can lead to undershoot, resulting in approximantization in intervocalic contexts. At the same time, more constricted targets for /p/ do not lead to approximantization with undershoot (in the studied sample). Thus, articulation aligns with impressionistic generalizations and acoustic measurements of differences between /p/ and /b/, at least in Peninsular Spanish and similar varieties. By extension, any method that will allow for inferring the relative distance between the articulators (here, lips) should be equally suitable for investigating consonant lenition.⁷ Motion capture is a promising technique in this respect.

Recently, there has been substantial interest in linking phonetic and phonological research with physical gestures. Since it has been shown that certain face and body movements serve as an enhancement mechanism for marking salient prosodic positions (McNeill, 2008; Wagner et al., 2014), many researchers have taken on a multimodal approach to studying speech. For instance, Beskow et al. (2006) found that in Swedish, words with a focal accent are supplemented by greater facial movement than words in other positions in the sentence. Similarly, Sangari (2002) and Esteve Gibert and Prieto (2013) showed that the timing of head nods is aligned with prosodic boundaries and the metrical structure of the word in focus. Such studies used motion capture to track facial expressions, with sensors placed on the face and body, and optical tracking. Although they focused on stress and sentence focus, they did not look at segmental phenomena accompanying speech. The motion capture method, however, is in principle suitable for analyzing at least some aspects of consonantal phonology as it allows for tracking facial expressions such as vertical and horizontal lip movements. In this vein, Żygis et al. (2017) focused on orofacial gestures produced in questions and statements in German whispered speech using motion capture. According to their results, there is a significant difference in lip aperture between the two sentence types. The same applied to lip aperture measurements focused on particular vowels. Thus, reliable measurement of lip opening from video data should make it possible to study other more nuanced questions, such as the relative constriction of labial consonants. No study of this kind has been done to date, which makes the present paper the first to provide a motion capture account of the weakening process.

While the motion capture method using optical tracking with sensors is quite successful, it also requires a special lab and equipment. Fortunately, some fieldwork-friendly alternatives to such endeavors are now available. Importantly, Holbrook et al. (2019) demonstrated the viability of optically tracking lip movements by using inexpensive off-the-shelf video recording equipment and algorithms. Krause et al. (2020) extended this approach by basing their measurements on intelligent face-tracking. Recordings can be comfortably made in the lab or in the field using an internet camera. With a specially-designed algorithm based on the OpenFace face-tracking utility (Baltrušaitis et al., 2018), the method allows for extracting lip aperture and lip area measurements from video material, provided that the speaker was seated without substantial movement facing the camera. Most importantly, Krause et al.’s study shows that internet camera-based lip tracking is robust and reliable enough to distinguish speech productions at the segmental level. Furthermore, the method allows not only for detecting lip closure vs. lip opening, but also for distinguishing different types or degrees of lip closure by looking at relative lip compression, measuring on the outer and inner lip contour. This is important from the perspective of the present study as we are especially interested in various degrees of lip closure and slight lip aperture corresponding to voiceless and voiced stops, and approximants, respectively.

1.5. The present study – goals and hypotheses

The principal goal of this study is to investigate the prosodic and segmental phonological effects of stop weakening in Canary Islands Spanish using a novel method. A motion capture study of lip movements gives us a link between the acoustics and articulation, and it allows us to determine whether lip movements tracked from video data can predict different degrees of consonant weakening (from stop voicing to strong approximantization) and whether such data can be used in future research alongside acoustic measurements. Thus, we hope that our study will constitute an important methodological contribution to laboratory phonology.

At the same time, we are interested in two important questions related to the weakening of underlying stops in Spanish. First, previous studies show that the deletion of a preceding consonant makes the stop behave differently than in underlyingly intervocalic position, which suggests that there is a phonological blocking effect that can be expressed as containment, i.e., non-deletion or incomplete deletion (non-pronunciation) of the apparently deleted consonant (Prince & Smolensky, 1993/2004; van Oostendorp, 2006). If the root node of the consonant is still there phonologically, the lack of weakening is explained. Such an interpretation, however, requires empirical support. An articulatory study can help us elucidate this question. A consonantal gesture can still be present even though the sound is not audible, e.g., due to gestural masking (or blending). As argued by Browman and Goldstein (1990), two gestures from two different tiers may sometimes mask or ‘cover’ each other, leading to apparent deletion. Some examples of this come from English. For instance, in sequences of words such as must be, the /t/ seems to be deleted in fast speech but palatography evidence shows that an alveolar closure gesture is there. In our case, it may be that a preceding consonant is masked by the following stop gesture and, as a result, apparently deleted. It is therefore worthwhile to look for comparative contexts. By looking at data elicited in a reading task, given that /s/ deletion is optional, we can compare contexts in which the stop is underlyingly intervocalic with two other types of outputs: a situation in which a preceding /s/ is retained (but e.g., weakened to [h]) and a situation in which it is deleted completely. In this way, we can look at the differences and similarities between the latter two contexts both in terms of acoustics and articulation to disentangle the effect of the preceding /s/ from the effect of a deleted segment. More specifically, we would like to know whether the /b/ in la vaca (the cow) /la#baka/, las vacas (the cows) /las#bakas/ with /s/ retention and las vacas /las#bakas/ with /s/ deletion are three separate surface categories (i.e., the post-deletion /b/ is somewhere in between the VbV and the VsbV) or whether approximantization is blocked in the latter two contexts in a similar way. If blocking takes place after a retained /s/, we should take it as evidence against post-consonantal weakening in the dialect. If blocking takes place after a deleted /s/ as well, there is evidence for containment, as mentioned above.

The second important question is the alignment of facial gestures with prominence in speech. As mentioned in Section 1.4., orofacial gestures produced during speech mark salient positions important for intonation and expressive language marking. This includes the position of focus. Knowing from previous studies that stop lenition in Spanish is sensitive to stress, we want to see whether this effect is magnified in a syllable that is stressed and in focus, i.e., accented, and whether lip movement data will show such effects. At the same time, it will be interesting to see whether focus effects are marked both acoustically and articulatorily or only articulatorily. To the best of our knowledge, no such study has been conducted to date.

Given the three main goals of our study, our hypotheses are as follows.⁸

H1A. Target sounds will have a greater intensity difference (i.e., less lenition) in deletion contexts (VsCV and V(s)CV) compared to underlyingly intervocalic position (VCV). In deletion contexts, there will be a blocking effect when the consonant is retained or deleted.⁹ In both cases, a greater intensity difference is expected. This can be summarized as an ascending lenition trend: V(s)CV, VsCV < VCV, i.e., V(s)CV, VsCV show less lenition than VCV.

H1B. Target sounds will have a smaller lip aperture (i.e., less lenition) in deletion contexts (VsCV and V(s)CV) compared to underlyingly intervocalic position (VCV). In deletion contexts, there will be a blocking effect when the consonant is retained or deleted. In both cases, a smaller lip aperture is expected. This can be summarized as an ascending lenition trend: V(s)CV, VsCV < VCV, as in H1A above.

H2A. Target sounds will have a greater intensity difference (i.e., less lenition) in stressed position (S) compared to unstressed (US), and an even greater intensity difference in stressed syllables in focus (SF). This can be summarized as an ascending lenition trend: SF < S < US in which SF shows less lenition than S, which in turn shows less lenition than US. The intensity difference gets smaller with more lenition.

H2B. Target sounds will have a smaller lip aperture (i.e., less lenition) in stressed position (S) compared to unstressed (US), and an even smaller lip aperture in stressed syllables in focus (SF). This can be summarized as an ascending lenition trend: SF < S < US, as in H2A above. Here, lip aperture increases with lenition.

H3. There is a correlation between the acoustic measurements (i.e., relative intensity difference) and articulatory measurements (lip aperture parameters) in predicting stop lenition. The greater the lip aperture, the smaller the intensity difference.

In all cases, we assume that there will be a difference between /p/ and /b/ in that there will be less lenition in the voiceless stop compared to the voiced one, as measured by both intensity difference and lip aperture.

3.1. Acoustic data analysis

To estimate the degree of lenition depending on the consonant and condition, we first looked at the deletion context, which was an independent variable with four levels: no deletion context (i.e., underlying VCV, annotated as ‘no context’), deletion context (V(s)CV) in which deletion occurred (‘deletion’), deletion context (VsCV) in which the /s/ was weakened to [h] or [ɦ] (‘weakening’), and deletion context in which deletion failed to occur and no weakening happened instead, i.e., /s/ was produced as [s] (‘no weakening’). We fit a linear mixed effect model with consonant (p/b), /s/-deletion context and their interaction as predictors, and intensity difference A as a dependent variable. The participant’s ID and target word were clustering variables. After the adjustment of the initial maximal model structure aimed at obtaining a non-singular model that converged, the model fit was the following:

intensity difference A ~ consonant * context + (1+consonant |participant) + (1|word)

The model was run on a total of 8,184 observations, of which 7,077 represented a no deletion context (i.e., underlying VCV); 406 had non-deleted preceding segments (VsCV), 127 of which were retained as /s/ and 279 weakened to [h]; and a further 701 had deleted preceding segments (V(s)CV). We performed F-testing of fixed effects, obtaining denominator degrees of freedom via Satterthwaite approximation. The results show main effects of consonant, F(1, 37.97) = 99.5, p < 0.001, and context, F(3, 277.6) = 253.17, p < 0.001, and a significant interaction between the two (F(3, 277.49) = 3.719, p < 0.05).

We then estimated marginal means from the model to assess the simple effects of context and to take a closer look at the interaction, testing via t-test, and adjusting for familywise error via Tukey’s method. The simple effects are as follows: The intensity difference of the segment is significantly smaller in no deletion contexts (underlying VCV) compared to both the deletion contexts in which the preceding sound was retained (VsCV, t = –25.99, df = 197.1, p < 0.001 when retained as [s], and t = –21.34, df = 259.6, p < 0.001 when weakened to [h]) and deleted (V(s)CV, t = –21.99, df = 30.1, p < 0.001). At the same time, the intensity difference calculated for the segment in deletion contexts with preceding consonant retention (VsCV) was significantly greater than when the preceding consonant was deleted (V(s)CV, t = 11.40, df = 8151.4, p < 0.001 in the case of the [s], and t = 6.27, df = 8136.5, p < 0.001 in the case of [h]). The difference between [s] and [h] outputs also resulted significant (t = 3.17, df = 8118.6, p = 0.008).

As for the interaction, all contrasts were significant except for the intensity difference of the /p/ in weakened /s/ vs consonants retained as [s] (t = 0.379, df = 8117.8, p = 0.99). Table 3 presents the marginal means for the interaction term and the interaction effects are shown graphically on Figure 4.

Figure 4

Effects plot, based on the estimated marginal means of the mixed-effects model, of the interaction between consonant and /s/-deletion context in predicting intensity difference A. (Error bars: 95% CI.)

To explore the effects of stress and focus on the relative intensity of the consonant, we excluded the deletion contexts from the database, which gave us a total of 7,077 observations, quite evenly distributed between the three levels of the condition variable (S = 2,170, US = 2,532, SF = 2,375). We then fitted a linear mixed effects model with consonant, condition and their interaction as predictors, and participant ID and target word as clustering variables. The resultant model that allowed for convergence was in the form

intensity difference A ~ consonant * condition + (1+consonant+condition|participant) + (1|word)

The results show that the effect of consonant is significant (F(1, 15.61) = 137.78, p < 0.001). There was also a significant effect of condition (F(2, 23.25) = 24.80, p < 0.001) and a significant interaction between consonant and condition (F(2, 134.31) = 10.708, p < 0.001).

As in the deletion model, we estimated marginal means from the model. In this case, we wanted to explore the simple effects of condition and the interaction effect in more detail. The simple effects for condition show that target sounds in S position have a smaller intensity difference compared to the SF position, but the effect is not significant (t = –1.95, df = 14.5, p = 0.16). At the same time, they have a significantly greater intensity difference compared to US (t = 5.18, df = 34.2, p < 0.001). Also, consonants in SF position have a significantly greater intensity difference compared to US (t = 6.66, df = 28.1, p < 0.001). This means that there is less lenition in SF than in US contexts.

The pairwise comparisons for the interaction demonstrate that underlying /p/ has a greater intensity difference in SF position compared to both /p/ and /b/ in all other positions, and these effects are significant in all but one case: The comparison with /p/ in S position yielded insignificant results (t = –2.739, df = 18.7, p = 0.113). As for underlying /p/ in S position, it shows a greater intensity difference compared to /p/ in US position (t = 6.040, df = 58.9, p < 0.001), and compared to /b/ in S position (t = –11.573, df = 17.5, p < 0.001), SF position (t = 9.587, df = 16.1, p < 0.001), and US position (t = 11.19, df = 16.2, p < 0.001). Finally, /b/ has a significantly lower value of intensity difference in US position compared to /b/ in SF position (t = 3.171, df = 46.0, p < 0.05), but not compared to /b/ in S position (t = 2.561, df = 48.8, p = 0.126). There is also no significant difference between /b/ in S and SF positions (t = –0.945, df = 17.7, p = 0.929). Table 4 presents the emmeans for all consonants and condition levels, and Figure 5 presents the interaction term graphically.

Figure 5

Effects plot, based on the estimated marginal means of the mixed-effects model, of the interaction between consonant and condition (S, SF, US) in predicting intensity difference A. (Error bars: 95% CI.)

3.2. Correlating the acoustics with the articulatory measurements

We next evaluated which of our measured articulatory variables most reliably covaried with acoustic intensity difference A. Intensity difference is an established means of assessing lenition. However, to our knowledge perhaps only two relevant articulatory studies have been performed to date.¹⁵ Parrell (2011) used minimum lip aperture (dubbed “maximum constriction”) to assess approximantization of Spanish labial consonants in different phonetic contexts. As for a direct comparison of the articulatory and acoustic metrics, Parrell (2010) presented a small-scale study on two subjects, showing a negative correlation between constriction degree and intensity difference. Similarly, Hualde, Shosted and Scarpace (2011) showed a good correlation between acoustic intensity and articulatory data based on electropalatography in intervocalic contexts for the coronal stop in Spanish.¹⁶ The results of these studies suggest that a similar relationship might be revealed in our data.¹⁷

A priori, there appear theoretically well-motivated reasons to consider at least three of our measured variables: minimum lip aperture, relative aperture, and maximum closure speed. All three variables might be expected to predict intensity difference on the assumption that, for bilabial consonants, quieter/more consonant-like acoustic performance arises when the lips are brought closer together. On its face, relative aperture would seem most similar to (relative) intensity difference. However, while a canonical bilabial closure requires the lips to touch, most of the vowels in our inventory should be tolerant of a wide range of vertical apertures. This might reduce the explanatory power of relative aperture and/or maximum closure speed since both variables depend on the specific magnitude of maximum aperture.¹⁸

We made our determination empirically, using mixed-effects multiple regression. Because intensity difference, relative aperture, maximum closure speed, and minimum aperture all showed substantial positive skewness, we first transformed all four variables. Minimum aperture and relative aperture were converted to their binary logarithms. Relative intensity and maximum closure speed were converted to their square roots, as this yielded somewhat better distributions. Where necessary, small constants were added to the distributions to eliminate negative values prior to transformation. For the remainder of this paper, references to these four variables should be taken as references to their transformed versions.

In the model, acoustic intensity difference was entered as the outcome variable, while minimum lip aperture, relative aperture, and maximum closure speed were entered as numeric predictors. Consonant (/b/ vs. /p/) was also entered as a categorical control variable. Participant ID and target word were entered as clustering variables. The random effects structure was determined via backwards selection from the maximal model. Backwards selection prioritized first de-correlating random intercepts and slopes and then eliminating random slopes – higher-order terms first, until a convergent, non-singular model resulted. The final model had the form

intensityDifference ~ consonant + minimumAperture + relativeAperture + maximumClosureSpeed + (consonant + minimumAperture + relativeAperture + maximumClosureSpeed || participant) + (minimumAperture + relativeAperture || word)

where the double-bar notation indicates that slope-intercept correlations have not been estimated for a given cluster. Despite correlations arising between multiple fixed effects, the model was not markedly collinear (all VIFs < 1.5). This model was run on a total of 8,184 observations.

T-tests of model slopes (using Satterthwaite-approximated degrees of freedom) revealed that of the three predictors of interest, only minimum lip aperture reliably differed from 0, estimate: –0.118, SE = 0.022, t(28.62) = –5.36, p < .001. Unsurprisingly, the control variable of consonant was also reliable, estimate (using /p/ as the reference value): 1.248, SE = 0.159, t(58.92) = 7.86, p < .001. An effects plot showing how minimum lip aperture predicts intensity difference appears in Figure 6.

Figure 6

Effects plot, based on the estimated marginal means of the mixed-effects model, of how minimum lip aperture predicts intensity difference A. (Minimum lip aperture has been converted to its binary logarithm, and intensity difference A to its square root. Error band: 95% confidence.)

3.3. Motion capture data analysis

We next examined the effects of the experimental manipulations on articulation. Following the findings described in the prior section, we used minimum lip aperture (binary log transformed) as the dependent variable of interest.

As we had for the acoustic results, we first tested the effects of the /s/-deletion context, treating it as a four-level predictor: trials with no deletion context, trials with deletion context but no recorded deletion and no weakening, trials with deletion context but no recorded deletion and weakening, and trials with deletion context and recorded deletion (i.e., the no context, no weakening, weakening and deletion levels, respectively). We fit a linear mixed-effects model with consonant, deletion context, and their interaction as predictor variables, and participant ID and target word as clustering variables. After model simplification, the final fitted model was of the form

minimumLipAperture ~ consonant + deletion_context + consonant* deletion_context + (consonant + deletion_context + consonant* deletion_context|| participant) + (1 | word)

As was the case for the acoustic analysis, this model was run on 8,184 observations (7,077 VCV, 127 V[s]CV, 279 V[h]CV, and 701 V(s)CV). Because the deletion context predictor had four levels, we performed F-testing of fixed effects, obtaining denominator degrees of freedom via Satterthwaite approximation. The main effect of deletion context was reliable, F(3, 35.96) = 14.09, p < .001. The interaction was also reliable, suggesting the effect of deletion was moderated by consonant, F(3, 230.10) = 11.30, p < .001.

We estimated marginal means from the model to explore the interaction. We tested the resulting simple effects via t-test, using Satterthwaite-approximated degrees of freedom. Familywise error was contained using Tukey’s method. For /p/-initial target words, none of the pairwise comparisons were reliable. For /b/-initial words, the effects were more nuanced. When /s/ was fully deleted, the resulting minimum apertures did not statistically differ from those cases in which /s/ was fully retained or weakened to [h]. Trials in the no context level had larger minimum apertures than trials where /s/ was weakened, estimate: 0.883, SE = 0.113, t(362.00) = 7.78, p < .001, and trials where /s/ was fully deleted, estimate: 0.531, SE = 0.123, t(21.9) = 4.32, p < .01. Trials in the no context level may also have resulted in larger minimum lip apertures than trials where /s/ was fully retained, but the result was on the margin of reliability, estimate: 0.317, SE = 0.110, t(12.90) = 2.87, p = 0.56. Finally, trials in which /s/ was fully retained resulted in larger minimum lip apertures than those in which /s/ was weakened, estimate: 0.566, SE = 0.138, t(30.20) = 4.10, p < .01. Table 5 presents the marginal means for the interaction term. An effects plot of the interaction appears in Figure 7. Note that the trends in this plot, as well as the fact that a reliable contrast arises between retained and weakened /s/ but not between retained and deleted /s/, suggesting that lip apertures might be quite small on weakened trials. We will revisit this point in the Discussion.

Figure 7

Effects plot, based on the estimated marginal means of the mixed-effects model, of the interaction between consonant and /s/-deletion context in predicting minimum lip aperture. (Minimum lip aperture has been converted to its binary logarithm. Error bars: 95% CI.)

Next, we removed trials in the deletion context from the dataset and examined the effects of the remaining conditions (7,077 total observations, S = 2,170, US = 2,532, SF = 2,375). We treated condition as a three-level independent variable: trials where target word was unstressed, trials where target word was stressed, and trials where target word was both stressed and focused (i.e., the U, S, and SF conditions, respectively). We fit a linear mixed-effects model with consonant, condition, and their interaction as independent variables, and participant ID and target word as clustering variables. After model simplification, the final fitted model was of the form

minimumLipAperture ~ consonant + condition + consonant*condition + (consonant + condition | participant) + (condition | word)

Because the condition variable had three levels, we again performed F-testing of fixed effects. The interaction was not reliable. However, the main effect of condition was reliable, F(2, 32.73) = 4.92, p = .014. The main effect of consonant was also reliable, F(1, 23.17) = 7.54, p = .011.

We estimated marginal means from the model to assess the simple effects of condition, testing the simple effects via t-test, and adjusting for familywise error via Tukey’s method. SF trials resulted in reliably smaller minimum lip apertures than S trials, estimate: –0.084, SE = 0.032, t(21.7) = –2.60, p = 0.042. SF trials may also have resulted in smaller minimum lip apertures than US trials, though the effect was on the margin of reliability, estimate: –0.134, SE = 0.056, t(58.00) = –2.39, p = 0.052. US and S trials did not differ from each other (p = 0.661). Table 6 presents the marginal means for the main effect of condition. An effects plot of the main effect of condition appears in Figure 8.

Figure 8

Effects plot, based on the estimated marginal means of the mixed-effects model, of the main effect of condition on minimum lip aperture. (Minimum lip aperture has been converted to its binary logarithm. Error bars: 95% CI.)

3.4. Bilabial closure timing in /s/ deletion contexts

Though not the primary focus of this study, the articulation of the bilabial closure may add relevant evidence to the question of whether the phonetic /s/ ([h]) is fully deleted, is truncated, or is masked by coproduction with /p b/. For example, ‘early closure’ might appear to be suggested by Figure 2, in which the closure phase for full deletion trials appears to cover a much larger part of the early timeframe, compared to other contexts (at least for /b/ words). However, one should not be too hasty to draw conclusions from Figure 2, because it is plotted over proportional time. Judged in terms of the overall duration of the utterance, the bilabial closure will always be delayed for cases of /s/ retention, relative to cases of /s/ deletion, regardless of the underlying deletion mechanism. This is because in cases of /s/ retention, the implementation of the /s/ has been ‘slotted’ into the time between the first vowel and the implementation of /p b/ closure. Therefore, more specific follow-up was called for. Note that, since we have no direct measure of either tongue-tip or glottal constrictions, any evidence gleaned from this exercise is necessarily incomplete.

We generated a new Python script that detected kinematic landmarks of bilabial closure in the vertical lip aperture trajectory data. These landmarks were based on those estimated via the findgest algorithm in Mview (Tiede, 2005), although the specific implementation was new to our script. Briefly, the script searched for the moments of peak closure velocity and peak opening (release) velocity following the closure. During the closure phase, gestural onset was logged as the first moment at which the closure velocity rose above 20% of peak, and target achievement was logged as the first moment at which closure velocity fell beneath 20% of peak.

Gestural onset was frequently logged as 0 ms, creating a sizeable floor in the distribution, and making it an undesirable dependent variable. We therefore fit a linear mixed-effects model with target achievement time (binary-log-transformed) as the dependent variable, and consonant, deletion context, and their interaction as independent variables. The main effect of deletion context was reliable, F(3, 16.61) = 5.39, p < .01. After Tukey correction, two reliable pairwise contrasts were observed: between the no context and weakening levels, and between the no context and deletion levels (see Figure 9).

Figure 9

Effects plot, based on the estimated marginal means of the mixed-effects model, of the main effect of deletion context on target achievement timing for the bilabial closure. (Error bars: 95% CI.)

Since we modeled target achievement time, rather than gesture onset time, one might wonder if the delays in the weakening and deletion contexts (relative to no context) reflect additional time needed to form a tighter closure (as opposed to the approximantized closures observed for /b/ in the no context level). To address this concern, we estimated a follow-up model testing the linear association of minimum lip aperture with target achievement time (using consonant type as a control factor). However, none of the estimated effects were statistically reliable (all ps > .2).

We also fit a linear mixed-effects model testing how consonant and deletion context predicted acoustic duration of the overall phonetic sequence. There was a reliable main effect of deletion context, F(3, 31.20) = 38.32, p < .001. All pairwise comparisons were reliable after Tukey correction (see Figure 10).

Figure 10

Effects plot, based on the estimated marginal means of the mixed-effects model, of the main effect of deletion context on phonetic sequence duration. (Error bars: 95% CI.)

We suggest that these differences in sequence duration likely arise due to one or more of the following issues. First, there may be differences in overall speech rate. Perhaps /s/ retention is more likely at lower rates of speech. Second, it makes sense for the acoustic duration of the /s/ to be “missing” from V(s)CV tokens. It is worth noting that the average 60-ms difference between the no weakening and deletion levels is roughly equivalent to the average acoustic duration of [s] in no weakening cases. Intriguingly, the durations of deletion and no context cases themselves differed, suggesting that the timing of V(s)CV and ‘true’ VCV tokens were not identical. This was further borne out by our finding that the bilabial closure target was achieved later for deletion cases than for no context cases, which might suggest that some inaudible trace of the /s/ remains phonetically present in deletion trials. If true, the present data cannot adjudicate whether the underlying gestures are truncated or ‘masked’ by coproduced /p b/ gestures (or both).

4.1. Phonological effects – Hypotheses 1A and 1B

Hypotheses 1A and 1B stated that the target sound will have a greater intensity difference and a smaller lip aperture (i.e., less lenition) in deletion contexts compared to underlying VCV and that it will be the same in V(s)CV contexts, where the preceding consonant is deleted, and VsCV, where it is retained.

The results of the study show that there are significant differences in stop weakening depending on the phonological environment, as expected based on previous literature. Thus, stops in VCV sequences are significantly more weakened than stops in either V(s)CV or VsCV sequences. This was confirmed by acoustic and articulatory measurements.

As for the difference between the deletion contexts, there is a discrepancy between the acoustic and the motion capture analysis. Intensity difference shows slightly smaller values after the deletion of a preceding consonant (V(s)CV) compared to sequences with a retained /s/. This is against our expectations: We assumed that there should be no difference because weakening is not expected in either case. Importantly, the results tell us that stops do not weaken at all after the fricative [s] (according to previous studies, an intensity difference of 29 dB on average is equivalent to a full, usually voiceless stop pronunciation). Furthermore, it seems that the non-weakening environment is maintained both when the /s/ is weakened to a glottal fricative (around 27 dB on average) and after consonant deletion (23 dB).¹⁹ These results point to a ‘qualitative’ gradient effect: The greatest lenition blocking²⁰ happens after a full-fledged alveolar fricative and dissipates in line with the degree of lenition of the underlying /s/. When there is only partial weakening by loss of oral constriction, there is less of a blocking effect and even less so with full consonant deletion. Nonetheless, the resultant sounds are nowhere near lenited consonants of the approximant type in the case of /b/ (around 5–10 dB) or voiced stop in the case of /p/ (around 15–18 dB). Post-consonantally, the outputs of /p/ do not fall below 29 dB of average intensity difference and the outputs of /b/ are not lower than 16 dB in this respect.

The discrepancy between the behaviors of /b/ and /p/ is also worth noting. The values of intensity difference are the same in voiceless stops regardless of whether the underlying /s/ was retained unchanged or realized as a glottal fricative. However, we can see a difference between there being some remnant of the /s/ and full deletion in both /p/ and /b/. The latter suggests some kind of a transition stage between underlying and derived VCV sequences. Further studies are needed to explore stop weakening after other consonants, e.g., sonorants, although it is worth noting that our data from this and other studies contain some instances of stops preceded by /l/ in which no lenition ensues, e.g., el perro (the dog), salvajes (wild). It is therefore quite plausible that both voiced and voiceless stop lenition in Canarian Spanish is post-vocalic only. All in all, we can say that Hypothesis 1A was partially confirmed.

As for the motion capture data, they show a difference in lip aperture between the VCV and deletion contexts, and no reliable difference between VsCV with a full [s] or glottal [h] and V(s)CV (with /s/ deletion). It is also worth noting that the discovered effects were only reliable in the case of /b/, suggesting a different behavior of voiced vs voiceless stops.²¹ Moreover, there is also a surprising difference between sequences with /s/ weakening to [h] and /s/ deletion on the one hand, and sequences with a retained [s]. The trend suggests that /b/ preceded by a weakened /s/ may be the most constricted, which is contrary to the findings based on acoustic intensity. The stop preceded by an [s] is different from /b/ in underlying VCV sequences but seems to show an intermediate effect, which might be compatible with the hypothesis that there is some weakening after consonants but not as much as in post-vocalic contexts, in line with earlier acoustic studies of some dialects of Spanish. Unfortunately, we have no previous studies to compare our results to and hence we cannot say whether the minimum lip aperture values shown in the marginal means table roughly correspond to a weakened (perhaps approximantized) variant of /b/ or simply a weaker stop pronunciation.

All in all, Hypothesis 1B was also partially confirmed. If we combine this result with the lip aperture trajectories from Figure 2, we can see that the level of aperture is indeed the same regardless of whether or not the preceding /s/ was deleted. A slight discrepancy between the trajectories can be observed in the part of the sequence before the target stop, where we see a slower drop in lip opening when /s/ is retained due to the presence of some form of a consonantal gesture (be it [s] or [h]). Interestingly, in /b/, the trajectories of V[h]CV and V[s]CV are aligned and differ from V(s)CV, whereas in /p/ there is alignment between V(s)CV and V[h]CV, with V[s]CV showing a different temporal path. These discrepancies should, however, be treated with caution given that time normalization results in distortions that may make visual comparisons unreliable. Another difference in the trajectories consists in the initial aperture of the lips depending on the context: The starting (vowel) level is lower when the vowel is flanked by a coda consonant, regardless of whether the consonant is pronounced or not.

All in all, these results lead to two important generalizations. First, there is no stop weakening after /s/ in the studied variety of Spanish, which is in line with some earlier experimental studies on different Spanish dialects (see Eddington 2011 for a review) and, indirectly, with some of the earlier reports on Gran Canarian Spanish (Almeida, 1982). This applies both to preceding alveolar fricatives and their weakened allophones. There is also a blocking effect of a preceding deleted /s/, which confirms previous findings, starting from reports of strengthened or tense voiced stops following /s/ deletion in Canarian Spanish (Trujillo, 1980, Dorta & Herrera, 1993) and followed by reports of unweakened voiced (and voiceless) stops following consonantal deletions in Gran Canarian (Broś et al., 2021) and some other Spanish dialects (e.g., Honduran, Amastae, 1989 or Costa Rican, Carrasco et al., 2012).

Second, the data are not entirely clear as to the phonological difference between contexts with /s/ deletion, /s/ weakening and /s/ retention. While acoustic data seem to point to a somewhat gradual weakening pattern and perhaps an intermediate category of stops produced in derived VCV contexts as opposed to underlying VCV on the one hand, and post-/s/ VsCV positions on the other, articulatory results are not fully in line with the acoustics, and /p/ does not necessarily follow the same pattern as /b/. While no differences have been detected for /p/ in motion capture data, perhaps due to the excessive coarseness of the method, the data for /b/ suggest a different pattern, i.e., a weakening/deletion blocking effect as opposed to (perhaps) some degree of weakening (an intermediate variant) after a full [s]. Nonetheless, taken together, the differences between the three deletion contexts are very small compared to the difference between VCV and deletion contexts in general and do not seem to reflect categorical changes, such as the change from a voiceless to a voiced stop or a change from a voiced stop to an approximant. Thus, we may look at them as gradient changes caused by articulatory overlap.

As for the phonological interpretation of the deletion context results, we can say that /p b/ weakening is blocked after consonants, and this blocking effect persists even after the consonant is deleted phonetically. There are several interpretations of this result. First, the apparently deleted segment is still there phonologically, i.e., it is not deleted completely, which makes the preceding vowel phonetically not adjacent to the stop consonant. Thus, the stop consonant behaves as if it was standing after a non-deleted segment and does not weaken. Another interpretation might involve gestural overlap in the phonetics component: the /s/ is not deleted completely and is instead masked by the adjacent stop gesture. As a result, no /s/ is phonetically present on the surface and the stop does not weaken because it stands after a consonant and not a vowel. Given the methods used in the experiment we cannot fully exclude the gestural overlap interpretation in cases of apparent deletion. On the one hand, we did find that the temporal trajectories in VsCV vs V(s)CV sequences differ by an average of 60ms in sequence duration, which roughly corresponds to the duration of the [s] in the dataset. This would suggest that the acoustic deletion of /s/ is complete, i.e., that no /s/ gesture occurs at all in these sequences. On the other hand, the durations of VCV and V(s)CV differ slightly and the timing of reaching the target for bilabial closure is delayed in the latter case. This might suggest that some remnant of the /s/ gesture is there nonetheless (see Section 3.4.). Thus, we leave the question of gestural overlap open for the time being.

Another question concerns gestural coordination in the sequence. In line with our data, the gestures in V(s)CV sequences are not coordinated in the same way as gestures in underlying VCV sequences. If full deletion ensues, we might expect gestural reorganization, which would lead to lenition in derived intervocalic contexts, yet this does not happen. Similar effects were reported by Shaw and Kawahara (2023), for instance, who analyzed gestural coordination in CVC sequences compared to CC sequences in which the intervening vowel was deleted in Japanese. They found a blocking effect in those consonantal contexts in which vowel weakening is less frequent, with no gestural reorganization and the preservation of the temporal structure of the sequence despite deletion. One of the interpretations put forward by Shaw and Kawahara (2023) is that less consistency in an optional process of vowel deletion leads to the preservation of a weakly activated gesture that inhibits the coordination between the now-adjacent consonants. Although our case looks at consonant deletion and vowel-to-consonant gestural coordination, there is a possible parallel. Since /s/ weakening and deletion is variable and optional across speakers, we might assume that the structural coordination between the vowel and /p b/ is maintained as with an intervening segment even in the phonetic absence of the /s/. Some support for this can be found in the delayed timing of reaching the bilabial closure in V(s)CV compared to VCV reported in Section 3.4.

The above interpretation might find its phonological equivalent in gradient symbolic representations (Smolensky & Goldrick, 2016; Zimmermann, 2019) which show gradient activation in phonology. Note that there seems to be a gradual change in the intensity difference of /b/ depending on the featural representation of the /s/. While a full alveolar [s] triggers a fully voiced occlusion, weakening to [h] (weaker activation) changes the [b] in the direction of a weakened stop, and /s/ deletion (weaker still) leads to a pronunciation that is even closer to an approximant. However, these gradual changes do not lead to approximantization and hence do not involve a featural change in phonological terms. Besides, articulatory data do not confirm these effects; neither were they found for /p/. Thus, we are more inclined to determine that, in the absence of more convincing evidence, there seems to be no phonological difference between contexts with /s/ deletion and /s/ retention. The deleted consonant behaves as if it was still there, blocking the process of weakening in the same way as a retained consonant. As a result, a derived intervocalic environment leads to a different phonetic representation than an underlying one. The difference between the two environments is not a phonetic but a phonological (structural) one.

Phonological theory has dealt with the above problem on numerous occasions, based on the assumption that structural relations are not necessarily equivalent to surface phonological or phonetic relations. It is therefore possible to encounter some types of covert structure in the form of a root node that is left phonologically unparsed (Prince & Smolensky, 1993/2004) or that is projected but not necessarily pronounced (Goldrick, 1998). The usual situation, in which no deletion ensues, requires reciprocity between the projection and pronunciation lines connecting the underlying representation with the surface representation, i.e., full bidirectionality. Unpronounced projections are not visible nor perceptible on the surface because they cannot be interpreted phonetically. A schematic representation of our three types of sequences taking into account covert structure is presented in Figure 11 below.

Figure 11

A schematic representation of the three types of sequences analyzed in the study. Projection lines are represented by upward arrows, pronunciation lines by downward arrows; segments correspond to lexical representations. Note that the pronunciation of the /s/ as [s] leads to the same structural representation as in b).

The above representations can be used as a basis for a formal analysis, which can be easily operationalized in one of the leading phonological frameworks. Looking at the opaque blocking effect in Gran Canarian Spanish, Broś (2016), for instance, uses a simplified version of Colored Containment (van Oostendorp, 2006) and shows a successful analysis of the non-weakening of /p t k/ in deletion contexts under Optimality Theory. Although the assumptions made in Broś’ paper were purely phonological, the present study confirms their correctness in capturing the blocking effect via acoustic and articulatory analysis. The preceding consonant may be unpronounced but still blocks stop lenition and hence must be present in the phonological output structure (the root node and its projection line are still there).

An interesting implication of a containment-based analysis is that it leaves the question of whether a given segment was deleted (no articulatory movement was made/planned) or masked by another gesture open. Note that the phonetic component interprets the output of phonological computation. If a given segment is there structurally (it contains a projection line) but is not pronounced, as in Figure 11c, this is treated differently than both an absent segment (Figure 11a) and a projected and pronounced segment (Figure 11b) and may be realized phonetically as gestural overlap. Otherwise, the lack of the pronunciation line might be interpreted as a lack of an /s/ gesture (target) and hence a phonetic VCV sequence. The unpronounced output of phonology might give rise to different, perhaps gradient phonetic interpretations, which is supported by the slight acoustic differences between sequences with retained and deleted /s/ in our data. Such an in-between option was also proposed by van Oostendorp (2008) in his analysis of incomplete neutralization in stop devoicing observed in many languages of the world. Thus, the containment approach does not allow for a formal distinction between the two types of phonetic interpretations of the data, which is an important issue raised by this journal’s associate editor. Perhaps, ultimately, all deletion is due to gestural overlap, which is something that should be studied further in Spanish and other languages.²²

4.2. Prosodic effects – Hypotheses 2A and 2B

Hypotheses 2A and 2B stated that target sounds should be characterized by a greater intensity difference and a smaller lip aperture (i.e., less lenition) in stressed vs. unstressed positions, and these tendencies should be even stronger in focus vs. no focus, resulting in a US > S > SF lenition trend.

In this context, the second part of our analysis revealed a key discrepancy between acoustic and articulatory measurements. The acoustic analysis showed that there is a difference between stops in stressed vs. unstressed positions in terms of degree of weakening. At the same time, there seems to be no difference between the two types of stressed syllables: S and SF. Thus, Hypothesis 2A is only partially supported. The motion capture data, on the other hand, demonstrate a significant difference between the focus position and the other two contexts, but no differences between S and US, contrary to our expectations. Hypothesis 2B was partially supported. All in all, the results suggest that lip movements show higher-level effects, such as sentence prosody (here: stress in focus position), which is compatible with previous studies on orofacial gestures produced during speech, while acoustic measurements are more sensitive to word-level processes and lexical as opposed to sentence stress.

Another explanation of the results might be sought in experiment design. While lexical stress was easy to elicit, focus was determined syntactically, on the expectation that speakers would put main emphasis on the rightmost word/phrase in the sentence, in accordance with the rules of nuclear stress placement and prominence in broad focus in Spanish (cf. Section 1.2.). Given the nature of the study, all the sentences were constructed as simple or complex statements, and we did not manipulate sentence structure or try to elicit narrow focus in a particular position e.g., by asking speakers to answer questions about particular parts of the sentence. We also did not control the number of words or follow the same syntactic structure in all sentences to provide a stimulus set, ensuring a more natural variability in /p b/ production. Consequently, it is possible that the speakers did not use neutral focus marking at least in some cases, which may have affected our results. The prominence of the consonant in S and SF conditions may have blended to some extent. Since we could confirm the effect of focus in the motion capture data, however, it seems that our expectations concerning intonation and prominence marking were borne out. If we look at the second vowel of the sequence depending on the condition, the data are in line with the expectation that the nuclear stress should be characterized by greater vowel duration in Spanish (see Figure 12).

Figure 12

The duration of the second vowel depending on the condition. Note that this vowel is the longest in SF, and the shortest in US.

At the same time, it must be noted that the intensity of the second vowel is the lowest in SF. Although this parameter is not a sentence prominence marker in Spanish and plays a greater role at the level of lexical stress distinctions instead (see e.g., Ortega-Llebaria & Prieto, 2007), the low intensity of the SF vowel may be an indication of an overlap between stress and accent with boundary effects. Domain finality, such as the end of a sentence or utterance, has been associated with various types of prosodic weakness and articulatory reduction, including accent retraction, pitch lowering, etc. (Cheng & Kisseberth, 1979; Hyman, 1977; Beckman & Pierrehumbert, 1986), which may have some impact on the intensity-based measurements centered on lenition.²³ All in all, the acoustic data show sensitivity to intensity differences in stressed vs. unstressed syllables, in line with previous literature, and no discrimination of SF vs. S based on this parameter, also as expected (since accent is cued by duration and pitch rather than intensity in Spanish). Articulatory data, on the other hand, show sensitivity to the focus position. Thus, the cues relevant to stress and accent in Spanish are apparently reflected not only in the vowels, but also in onset consonants produced as a part of the studied syllables. Finally, it is worth mentioning that the discrepancy between the acoustics and articulation in capturing stress and accent effects in stop lenition in Canarian Spanish finds a parallel in at least one previous study. Shaw et al. (2020) found that acoustic and articulatory data are correlated in velar lenition in Iwaidja, as in this study, and that they seem to provide overlapping but distinct channels of information about lenition.

4.3. The reliability of lip measurements vis à vis acoustics – Hypothesis 3

Finally, Hypothesis 3 assumed that there is a correlation between the acoustic measurements and articulatory measurements in predicting stop lenition: the greater the lip aperture, the smaller the intensity difference. Our study showed that motion capture data obtained via lip tracking from outputs recorded with an internet camera can be successfully used to predict stop lenition. There is a significant negative correlation between one of the three tested articulatory parameters, i.e., minimum lip aperture, and its acoustic equivalent, i.e., relative intensity difference. Thus, hypothesis 3 has found support in our data.

It follows from the above that lip measurements obtained from video recordings are fine-grained enough to accurately predict the degree of weakening in stops in intervocalic position. Furthermore, they not only constitute a viable, easy-to-implement alternative method of studying lenition, they also provide additional information on those aspects of speech which have not been captured by the acoustics, such as nuclear stress/focus effects (see Section 4.2.). As a further extension, they can be used to explore other prosodic characteristics of segmental phenomena and therefore be recommended to linguists interested in linking segmental phonology with intonation research.

Note that intensity, measured in dB, has been demonstrated to reliably show even the slightest differences between sounds undergoing lenition. For instance, Broś et al. (2021) showed that underlying voiced stops realized as [b d g] are significantly different from underlying voiceless stops that underwent voicing by lenition, the difference between them being in the order of a few dB of relative intensity. In our data, we did not label surface realizations as either lenited or not, voiced or voiceless, or approximants. We relied on the underlying specifications only. However, the acoustic data help us categorize these sounds based on the predictions made in previous studies. In Figure 13, we can see the intensity difference calculated for the target sounds depending on the condition. Judging by the mean values, we can conclude that /b/ is produced as an approximant in all VCV contexts (relative intensity below 10 dB) and as a stop in the deletion context (values close to 20 dB). At the same time /p/ is in the range of a voiced or a partially voiced stop in VCV contexts (15–22 dB), and in the range of a tense voiceless stop in deletion contexts.

Figure 13

Intensity difference of the target sound by context and condition.

Against this background, the reader may wonder how much precision and discrimination potential can be found in lip measurements made at a distance. A similar comparison with raw lip measurements as a discriminating factor shows that while differences in lip aperture are much smaller than the relative differences in intensity, the former reliably distinguish different degrees of weakening, which is exemplified by /b/ (see Figure 14). As for /p/, the differences between unweakened /p/ and its partially weakened or voiced counterparts, for example, are perhaps physically too small to capture by lip tracking. This was also pointed out in Section 3.

Figure 14

Minimum lip aperture for the target sound by context and condition.

It is also worth mentioning that intensity measurements are a more general metric that ‘sieves in’ more than just differences in the constriction of the produced sounds. By contrast, lip aperture is by definition a consequence of fewer speech-relevant kinematic degrees of freedom than is acoustic intensity. Looking at this metric can only speak to relative consonant aperture and not e.g., voicing. Thus, any mismatch between the two metrics is possibly driven by the fact that intensity differences result from degrees of freedom other than lips and jaw.

To summarize the comparison between the two methods, given the intra- and inter-category discrepancies in the predictions allowed by the two types of measurements, we can conclude that they are complementary in nature and that each one of them constitutes an added value. For this reason, we encourage researchers to use both acoustic and articulatory data in future studies on consonant lenition.

1. List of sentences used as stimuli

• La bamba es un baile latinoamericano muy conocido.

• En toda América Latina se baila la bamba.

• Mi padre pasó todo el fin de semana bailando la bamba.

• La bajada de agua de nuestra casa se ha estancado otra vez.

• A mi padre le duele la barriga.

• La base científica del Covid es indudable.

• La tienda de deporte más conocida de la isla se llama Base.

• La banda de música empezó el concierto con la bamba.

• La báscula demuestra que he bajado la barriga.

• La bala le ha atravesado la barriga.

• La baba de caracol se usa para producir cosméticos.

• A la niña se le cae la baba.

• El perro tiene hambre, se le cae la baba.

• La vaca de Juan cuesta mucha pasta.

• La vaca de Juan ha dado mucha leche este año.

• A Juan se le escapó la vaca.

• La papilla de avena le gustó mucho a la niña.

• La paciencia de esa mujer me tenía bastante impresionado.

• La paciencia de mi padre siempre ha sido objeto de admiración.

• La patrulla ha estado buscando al ladrón desde hace días.

• La paella valenciana es la más auténtica de todas las paellas.

• La vacuna contra el Covid debería ser obligatoria para todos.

• El jugador del Real Madrid comenzó la jugada por la banda.

• Las bandas de música más conocidas de la historia son The Rolling Stones y Metallica.

• La valla protege la casa de los animales salvajes.

• El ladrón casi se escapa, pero al final se chocó contra la valla.

• La barriga de ese hombre casi no cabe en la puerta.

• La basura acumulada en el Océano Atlántico es una pasada.

• La parte más difícil de ser padre es tener que aprenderlo.

• Vimos las primeras dos películas y ahora tenemos ganas de ver la última parte.

• La pandilla de mi barrio es muy conflictiva.

• La pasta de dientes que compramos no sirve para niños.

• La baraja española tiene cuarenta cartas con cuatro palos.

• La página web del gobierno anunció la erupción del volcán de La Palma.

• Para acceder a más información sobre las clases online tienes que ir a la página web.

• Las papas arrugadas son una comida típica de las Islas Canarias.

• Vivo en Las Palmas, pero soy de Gáldar.

• Las Palmas de Gran Canaria es la capital de la isla.

• Las Vacas Locas es una banda de música de Tenerife.

• Las bases aéreas han parado de operar a causa de la pandemia.

• La palanca de cambios de mi coche dejó de funcionar.

• Con una palabra lo dijo todo.

• Está mal dicho decir “la padre”.

• En esta celebración se suelta al final la paloma de la paz.

• Me temo que la paz mundial es imposible de conseguir.

• Se llama panza de burro cuando está nublado en verano.

• La camisa es muy pequeña y se le ve la panza.

• La batata del potaje no era muy dulce.

• La barrera estaba mal colocada y el portero no veía.

• Estoy a dieta y tengo miedo de subirme en la báscula.

• La paga mensual es más baja de lo que pensaba Paco.

• Tuve un accidente muy grave y por eso pedí una paga.

• Después de cuatro horas en el quirófano los médicos lograron sacarle la bala.

• El gato pardo siempre sabe cómo llenarse la panza.

2. Examples of TextGrid annotations

Figure 15

Annotation of a VCV sequence in the word vaca (cow) (condition: SF).

Figure 16

Annotation of a VsCV sequence in the word baja (low) with /s/ retention (condition: DEL).

Figure 17

Annotation of a V(s)CV sequence in the word vacas (cows) with /s/ deletion (condition: DEL).