The status of /f/ in Mayan-accented Spanish

2.1 /f/ lenition

The historical literature on the lenition/aspiration of /f/ is well-documented. According to most scholars, the variation between [f], [h] and even ∅ has been present in Spanish since before the early modern period and has gone through periods of allophonic loss and reintroduction (Brown and Raymond 2012; Lapesa 2008; Penny 2000, 2002; among others). This variability is reflected in the graphemes f and h, the latter not being pronounced, in etymologically related words: fumar [fuˈmaɾ] ‘to smoke’, versus humo [ˈumo] ‘smoke’ (Van Buren 2017: 163). According to Penny (2000: 45–46), word-initial /f/ in words of Latin origin that had become [h] were starting to be dropped altogether by the 16th century and this process was spread to the Americas. Of particular interest to contemporary cases of lenition, /f/ before /uV/ diphthongs was realized as a voiceless labiovelar fricative [ʍ] and, unlike initial [h], this allophone was maintained. Although it is not entirely understood why, /f/ was reintroduced into the phonemic inventory of Spanish and [f] became the canonical variant towards the end of the Middle Ages (Penny 2000: 72–73). Nevertheless, this allophonic variation in cases where /f/ preceded /uV/ still exists today, where non-canonical realizations of /f/ are associated with rural and lower socioeconomic-class speakers (Penny 2000: 163). The exact phonetic realization of non-canonical /f/ before /uV/ in contemporary Spanish varies; as [ʍ] may have become an aspirated, or debuccalized, voiceless glottal fricative [h], a voiceless velar fricative [x], a voiceless labiovelar fricative [ʍ], or a voiceless bilabial fricative [ɸ] in different dialects (Penny 2000; Pensado 1993; Renaud 2014; among others). Nonetheless, an exact definition may be futile as distinctions between several of these variants are lost in this specific phonological context. For example, the difference between [h] and [ɸ] may be negligible as [h] generally assimilates in rounding before the rounded semivowel of the following /uV/ diphthongs and thus indistinguishable from [ɸ] (Ladefoged and Disner 2012). As such, we use the generally accepted variant [h] when discussing cases of /f/ lenition in this paper.

Studies on contemporary Spanish have demonstrated that rates of /f/ lenition are conditioned by both linguistic and extra-linguistic factors. Phonetically, /f/ lenition is most common preceding /uV/ diphthongs. In fact, in some studies it has only been found in this context (Van Buren 2017). Experimental work has demonstrated that some phonetic factors that promote [f]/[h] variation in Spanish include the ambiguity of these two fricatives before rounded vowels, such as /u/, both in terms of production and perception (Foulkes 1997; Greenlee 1992; Renaud 2014). Additionally, several studies have found that the preceding sound may also condition /f/ lenition. Specifically, it has been found to be more common after non-high vowels (Brown and Alba 2017; Brown and Raymond 2012; Mazzaro 2015) and after the voiced alveolar lateral /l/ (Mazzaro 2015). In some studies, stress appears to facilitate lenition as well, as lenition is more common in unstressed syllables (Brown and Alba 2017; Brown and Raymond 2012; Mazzaro 2015); however, stress was not a significant factor in Renaud (2014). Studies that have analyzed the effects of word position on /f/ lenition have produced conflicting results. For example, Mazzaro (2015) finds that word-initial position was a conditioning factor for lenition whereas Renaud’s (2014) analysis found no effect of word position on the lenition of /f/.

Concerning extra-linguistic factors, the sociolinguistic work on /f/ lenition in Spanish has produced some expected results that pattern after general sociolinguistic trends (Labov 2001): females and younger speakers have been shown to prefer the prescriptive “standard” [f] (Mazzaro 2015; Resnick 1975; Van Buren 2017). The most common variable associated with /f/ lenition in the literature is that of socioeconomic status, often analyzed via the proxy of education, as many authors have associated /f/ lenition with lower socioeconomic classes (Agüero Chaves 1962; Alvar 2013; Canfield 1981; Lipski 2008; Moya 1981; Parodi 2001; Penny 2000; Quesada Pacheco 2013; Resnick 1975; Zamora Vicente 1996). This correlation has been verified via quantitative studies such as Calvo Shadid (1996), Mazzaro (2015), Renaud (2014), and Van Buren (2017); Brown and Alba’s (2017) analysis demonstrated a trend in this direction that approached significance. As some of these authors explain, higher levels of education often go hand in hand with higher levels of literacy; as speakers’ knowledge of Spanish orthography increases, they are better able to connect the graphemes f and h with specific words, particularly those that are more frequent in everyday speech, and produce far less cases of non-prescriptive or “non-standard” realizations of /f/ (Mazzaro 2011: 164; Renaud 2014: 267).

Many of these studies investigate speakers that are bilingual and in language contact situations; however, few of them have analyzed bilingualism or language contact as a variable. Van Buren (2017) presents one of these few studies of the lenition/aspiration of /f/ in an investigation of Spanish monolingual and Spanish-English bilingual Mexican migrant workers in the Pacific Northwest of the United States. Although some findings of /f/ lenition in this context parallel results in the abovementioned studies, Van Buren’s analysis reveals that education is not a significant variable among this population. However, bilinguals produce /f/ lenition at a lower rate than Spanish monolinguals. She hypothesizes that within this context, bilingualism is a sign of upward mobility and could be functioning as a proxy for socioeconomic class instead of education. Additional results from this study of bilinguals and monolinguals that contrast the previous studies include the heighted use of [h] among younger speakers, and a higher rate of [h] following the high vowel /i/.

In Guatemala, the context is not only multilingual, but also one in which the Spanish /f/ is found in a phonemic conflict site with Mayan languages. Although there are few studies on /f/ lenition in Guatemalan Spanish, one of the first is that of Lentzner (1893: 42), who briefly mentions that it is quite common among bilinguals. Predmore’s (1945: 278) description of Guatemalan Spanish states that /f/ is lenited among bilinguals, those with less education and in more colloquial speech, and that there is also an allophonic variant [ɸf]; this specific variant, however, has not been corroborated by any other researcher. Among speakers in southwestern Guatemala, Alvar (1980: 272) finds most cases of [h] in word-initial position and lenition is common before all vowels, not just /uV/ diphthongs. Speakers in Alvar’s study also associated /f/ lenition with lower socioeconomic classes and the Maya population. Utgård (2010: 58) states that [f] is by far the most common realization of /f/, but that lenited forms are more common before non-front vowels, especially before /uV/ diphthongs. Utgård also concludes that it is generally the older speakers that produce more lenition. Finally, French and García Matzar (2023) specifically list /f/ lenition among Kaqchikel-Spanish bilinguals as the direct result of a phonemic conflict site and only report it when preceding /uV/ diphthongs. In sum, the existing literature on /f/ lenition in Guatemalan Spanish offers some linguistic factors that may differ from other varieties of Spanish and mentions of extra-linguistic factors coincide with Spanish in general. In some cases, contact with Mayan languages is presented as a factor. However, these studies were largely based on observations and not quantitative statistical analyses, and none present any acoustic data.

2.2 /f/ fortition

Unlike /f/ lenition, the fortition of /f/ cannot be traced to any diachronic variation in Spanish, but to language contact. Across different varieties of Spanish, /f/ fortition has only been reported in specific areas of contact with languages that do not possess the phoneme. For example, Lipski (1987: 215) and Quilis (1996: 241) note that in the Philippines, /f/ → [p] occurs among bilinguals, especially those that speak Chavacano creole, Rodríguez Cadena (2008: 155) reports that in Colombian Spanish in contact with Wayuunaiki, “semi Spanish speakers” replace /f/ with [p^h], as in “[kon p ^h amilia] ‘with family’”, and Chappell (2021: 189–190) states that it occurs in Nicaraguan Spanish in contact with Miskitu, as in [p]echa for [f]echa ‘date’. Outside of Guatemala, Flores Farfán (2000: 147) reports that some Spanish-Yucatec Maya bilinguals in Southern Mexico, especially those whose L1 is Yucatec Maya, produce /f/ fortition: “[p]ernando for [f]ernando”. Further examples of this process can be seen in the language in contact with Spanish. For example, in Spanish loan words in Mayan languages, /f/ is primarily replaced with /p/: Spanish /ˈfɾuta/ → K’iche’ /ˈprut/ ‘fruit’; Spanish /kaˈfe/ → K’iche’ /kaˈpe/ ‘coffee’ (Baird 2023b: 5).

Even though there is ample contact with Mayan languages, there has been some debate among scholars as to whether the fortition of /f/ to [p] occurs in Guatemalan Spanish. In one of the first accounts of Guatemalan Spanish, Lentzner (1893: 42) states that when Maya individuals “try to speak Spanish as an L2,” they replace the [f] with [p] and gives the examples such as p amilia for f amilia ‘family’ and p os p or for f ós f oros ‘matches’.^[2] However, Alvar (1980: 272) refutes this claim, or at the least states that if did occur a century before when Lentzner first reported it, that it no longer occurs. On the other hand, García Tesoro (2008: 105–106) states that while /f/ fortition still occurs in Guatemalan Spanish, “we have only observed it among bilingual speakers with low levels of education.” Aleza Izquierdo (2010: 86) states that “[i]n zones with Mayan influence … the consonant /f/, which does not exist in Mayan languages, is pronounced as the bilabial occlusive [p] among bilinguals with low levels of education.” In the same year, Utgård (2010: 59) reports that “what Lentzner described in 1893 [fortition] has not been confirmed among the participants of this study.”

These mentions of /f/ fortition in Guatemalan Spanish in the literature are brief, as they are not the primary focus on these works. In fact, no one offers any linguistic factors that may condition /f/ fortition, nor any acoustic data. However, those that corroborate /f/ fortition bring up the same extra-linguistic factors; it primarily occurs among bilinguals with lower levels of education. In a pair of perception studies, Baird (2023a, 2023b) used the matched guise technique to analyze the sociolinguistic attitudes of /f/ fortition among native speakers of Guatemalan Spanish. The results corroborate the brief sociolinguistic comments of these previous authors as they demonstrate that /f/ fortition is highly stigmatized among Guatemalans and indexes a lower socioeconomic status and indigeneity.

2.3 The present study

Although the existing research on the lenition and fortition of /f/ in Guatemalan Spanish is sparse, two specific characteristics may be inferred. First, lenition appears to be more widespread than fortition, as the former is not disputed in the literature. Second, both lenition and fortition are more likely to occur among Spanish speakers that are bilingual in a Mayan language and have lower levels of education. However, as previously mentioned, the previous accounts lack acoustic data and quantitative analyses. As such, this study was guided by the following two research questions: First, what is the overall frequency of lenition ([h]) and fortition ([p]) realizations for /f/ present in the speech of simultaneous bilingual K’iche’-Spanish speakers? Second, what are the linguistic and extra-linguistic factors that condition the processes of lenition and fortition in the speech of simultaneous bilingual K’iche’-Spanish speakers?

3.1 Participants

In the summer of 2019, 40 K’iche’-Spanish simultaneous bilinguals (23 females, 17 males) participated in sociolinguistic interviews. All participants were from the municipality of Nahualá in the western highlands of Guatemala and ranged in age from 18 to 67 (M: 31.3, SD: 13.0). The participants varied in educational attainment in which 6 had completed less than a secondary education, 21 had completed secondary education, and 13 had at least some college education. Each participant completed the Bilingual Language Profile (henceforth BLP) (Birdsong et al. 2012) to examine language dominance. For reference, BLP scores range from −218 to +218 with a score of 0 being an idealized balanced-bilingual. In the current study, larger negative scores indicate more K’iche’ dominance, and larger positive scores indicates more Spanish dominance. Scores ranged from −83.9 (K’iche’-dominant) to 25.4 (Spanish-dominant) (M: −28.8, SD: 33.5). Of note is that 33 of the 40 participants were K’iche’-dominant according to the BLP scores. This is in line with the larger community of Nahualá as it has been shown to be a municipality with high levels of K’iche’-dominance among bilinguals (Baird 2015, 2017, 2018, 2019, 2021b).

3.2 Data collection and procedures

The sociolinguistic interviews were conducted in Spanish in Nahualá, Guatemala and ranged between 30 and 50 minutes in length. The sociolinguistic interviews included open-ended questions regarding themes such as cultural practices and personal histories to encourage naturalistic speech. All participants were recorded via a Marantz PMD661 digital voice recorder digitized at 16 bits (44.1 kHz) with a Shure SM10A head mounted microphone. The sociolinguistic interviews were conducted by K’iche’-Spanish simultaneous bilingual research assistants who were also from Nahualá and were the same biological sex as the participant and neither author was present during the interviews in order to reduce effects from the observer’s paradox (Labov 1972). The research assistants were trained by the first author in how to conduct a sociolinguistic interview.

3.3 Segmentation of tokens

Textgrids in Praat (Boersma and Weenink 2019) were created for each recording. A research assistant was trained to segment each word that contained an underlying /f/, resulting in 1,430 total tokens analyzed.^[3] The first author then segmented each /f/ token in Praat. Based on both auditory and acoustic cues, the tokens were coded as [f], [h], or [p]. Both [f] (Figure 3) and [h] (Figure 4) demonstrated frication in the waveform and spectrogram. However, [f] tokens present a much higher center of gravity (Hz) than [h] tokens as noted in the spectrograms as well as much more frication as noted in the waveform. Furthermore, tokens coded as [f] had a sound pressure level (dB/Hz) that continues past 3,000 Hz whereas tokens coded as [h] had a sound pressure level (dB/Hz) that drops before 3,000 Hz (Ladefoged and Disner 2012) (see Appendix I). Finally, [p] realizations (Figure 5) were coded as any tokens that presented an occlusion of at least 20 ms or more.

Figure 3:

Waveform, spectrogram, and textgrid of [f] token in fútbol ‘soccer/football’.

Figure 4:

Waveform, spectrogram, and textgrid of [h] token in fui ‘I went/I was’.

Figure 5:

Waveform, spectrogram, and textgrid of [p] token in enfermedad ‘illness/disease’.

3.4 Independent and dependent variables

For one of the extra-linguistic factors and several linguistic factors, the original coding included many more levels than the final statistical analysis. As several independent variables have many levels, and some of the levels had few tokens, this may result in a lack of convergence of a regression model, or result in Type I or Type II errors. To avoid these issues, data visualization was conducted for each independent variable to see if certain levels that behaved similarly and could be combined into one level. Thus, below we list the original levels for each factor, and then in some cases, the reduced number of levels.

Four extra-linguistic factors were included: sex, age, education, and language dominance. Age was considered a continuous variable. For education, participants were placed in the highest level of educational attainment: less than secondary education, secondary education, at least some college education, college education, and graduate education. Given the number of educational levels, we examined the data prior to conducting inferential statistics. The large cutoff in terms of linguistic practice was between those with secondary (or less) education and those with some college (or more) education.^[4] Based on these observations, we reduced these categories into two levels: secondary (or less) education, some college education (or more). It should be noted that the current study examines education not necessarily as a proxy for social class, but rather as a factor in its own right. Several variationist studies have utilized a socioeconomic index (Cedergren 1973; Labov 1966, 1972, 2001; Trudgill 1972; Wolfram 1969) that includes occupation, education, and income. Many studies have used education as a proxy for social class. However, the end goal of those studies is not to study social class, but rather that “class is meant to model the socioeconomic hierarchy of a community” (Tagliamonte 2012: 25), and not necessarily a “discrete set of identifiable classes” (Labov 2001: 113), to see how said socioeconomic hierarchy reveals “dynamics of language variation and change” (Mallinson and Dodsworth 2009: 265). It has been suggested that “linguists might treat education as a separate variable, independent of social class” (Ash 2013: 356) as previous studies demonstrate that education may not “contribute to the social stratification of speakers.” As previous /f/ studies (Mazzaro 2011: 164; Renaud 2014: 267) have indicated, more years of formal education is a direct connection between literacy, orthography, and more prescriptive phone-to-grapheme realizations. Thus, education will be treated solely as referring to education. There is most likely some collinearity with social class, as is common in many communities, but this study is more interested in the effect of formal education (i.e., literacy and orthography) on /f/ lenition and fortition. Finally, language dominance was measured based on a continuous BLP score (Birdsong et al. 2012) from −218 to +218. Prior to statistical analyses, both age and BLP were normalized and scaled using the scale() function in R so that they were centered at zero with a standard deviation of one.

For the extra-linguistic factors, prior to running any models, we examined any important differences or possible examples of collinearity that would confound the models. A Pearson correlation was conducted to examine collinearity between speaker age and BLP score. There was a close to (but not) significant weak negative correlation between age and BLP score (n = 40, df = 38, r = −0.298, R² = 0.088, p = 0.062). Thus, we included both independent variables within the analysis. However, it should be noted, given how close to significant this correlation is, that any age-related effects should be taken with caution. To examine if there were differences in BLP scores based on educational attainment we conducted a Paired Welch two-sample t-test and found that those with more some college (or more) education (M: −11.37, SD: 38.7) were overall slightly less K’iche’-dominant than those with secondary (or less) education (M: −37.18, SD: 37.62), p = 0.045. While barely significant, this would indicate that there is some amount of collinearity between the two variables. However, the authors decided to include them in the same models as educational attainment and language dominance are not two measures of the same variable. Finally, to examine whether there were any BLP score differences between the male and female speakers, a Paired Welch two-sample t-test was conducted and found that there were no significant differences in BLP scores between the female speakers (M: −29.70, SD: 29.03) and male speakers (M: −27.57, SD: 39.61) included in the sample, p = 0.853.

Four linguistic factors were also considered: previous segment, following segment, sentence position, and lexical stress. For previous segment, there were originally five levels, including: pause, vowel, nasal, liquid, fricative. Following previous research (Stivers et al. 2009), a pause was coded when there were 250 ms or longer of no speech prior to the segment. This was later reduced into three levels: other, pause, nasal. For following segment, it was originally coded with eleven levels including: [a], [e], [i], [i̯a], [i̯e], [l], [o], [ɾ], [u], [u̯e], and [u̯i].^[5] This was then reduced to three levels: [u̯e], [u̯i], and other. Position in the sentence included three levels: beginning, middle, and end. Lexical stress was a binary coding of tonic or atonic. Finally, speaker was considered as a random factor.

The dependent variable was the segmental coding of [f], [h], and [p]. As described below, these were treated as a binary dependent variable of [f] versus [h] (lenition) and [f] versus [p] (fortition).

3.5 Statistical analysis

Prior to regression modeling, a random forest was conducted using the cforest function from the party package (Hothorn et al. 2020) to determine the importance of each independent variable prior to conducting regression analyses following Tagliamonte and Baayen (2012). This is important as the order in which independent variables are placed into each model may affect the output. Thus, according to the results from the random forest, independent variables were placed into order of importance within the regression models.

The current study employed a Bayesian approach instead of a frequentist approach for two principle reasons. First, the Bayesian approach allows for mixed effects with logistic and multinomial regressions, while the frequentist only allows mixed effects for a logistic regression. Second, a common issue with frequentist approaches, particularly that of logistic regressions, is (quasi-)separation in which “one predictor completely or almost completely separates the binary response in the observed data” (Kimball et al. 2019: 231). This is highly problematic for logistic mixed effects models (i.e., glmer) as separation or quasi-separation can then “cause problems for model estimation that result in convergence errors or in unreasonable model estimates” (Kimball et al. 2019: 231) in which some estimates “become extremely- even infinitely- large” (Fonteyn and Petré 2022: 86). This particular issue was found in exploring the data with frequentists models, hence the authors employed a Bayesian approach to avoid (quasi-)separation issues following previous literature (Fonteyn and Petré 2022; Kimball et al. 2019).^[6]

Using the information from the random forest, a three-way ([f], [h], [p]) mixed-effects multinomial Bayesian regression models were conducted in R (R Core Team 2024) using the brm package (Bürkner 2017). However, after running several models, particularly those that included the most important independent variables (according to the random forest), there were too many error messages to continue (i.e., transition issues after warmup, R-hat values well above 1.1, too low of a Bulk Effective Sample Size, too low of a Tail Effective Sample Size). A multinomial regression, as compared to logistic regressions, requires significantly more data. Hence, the warnings for Effective Sample Size (Bulk and Tail) and R-hat value issues would indicate that we do not have enough data to run a mixed effects multinomial regression.

Thus, statistical analyses were separated for lenition ([f] and [h]) and fortition ([f] and [p]) into two Bayesian mixed effects logistic regressions using the brm function in R from the brms package (Bürkner 2017). One difference between a frequentist and a Bayesian approach is that of having constraints on the model parameters. In frequentist models, the values of β are assumed to be equally likely, ranging from “negative infinity to infinity” (Kimball et al. 2019: 238). Bayesian approaches, on the other hand, allow the researcher to begin with starting constraints, known as priors, which help avoid the over estimation of effects when there is quasi-separation or separation issues (2019: 238). Thus, rather than assume that estimates range from negative infinity to infinity, the current model begins with the assumption that most values will be near zero and thus the priors for all models were “weakly informative priors” using the following code: c(set_prior(“normal(0,1)”, class = “b”)). To determine which was the best model, we utilized the Bayesian leave-one-out cross validation (LOO-CV; Vehtari et al. 2017) method using the loo function in which the best model has the lowest LOOIC score. Finally, the visualization of the data was created with ggplot2 (Wickham 2016). Given that many interactions in each model would not allow the regressions to converge, a conditional inference tree using the cforest function in the party package (Hothorn et al. 2020) was conducted to examine possible interactions.

4.1 Lenition: [f] & [h] variation

The random forest results indicate that for [h]/[f] variation, the most important predictor is following sound, followed by educational attainment, speaker, scaled BLP score, syllabic stress, sex, scaled age, previous sound, and finally, sentence position (Figure 6). Of note, syllabic stress was not included in the regression modeling for [h]/[f] variation due to its collinearity with following sound (that is, [u̯i] and [u̯e] contexts are almost always in the tonic syllable).

Figure 6:

Random forest of [h]/[f] variation.

A Bayesian mixed effects regression model was conducted. The best fit Bayesian mixed effect logistic regression model is shown in Table 1. Of note, the reference level for realization was [f], coded as 0, and [h] realizations were coded as 1. The model selected is that with the lowest LOOIC vales (see Appendix III for loo() model comparison).

The model indicates a strong effect of following sound in which both [u̯i] and [u̯e] strongly favor [h] realizations in comparison to other following sounds. The next strongest predictor was the interaction between following sounds and education in both following diphthongs, [u̯e] and [u̯e], most favor [h] realizations among those with secondary (or less) education (Figure 7). Education as a simple predictor also indicated that those with some college education or more strongly disfavor [h] realizations compared to those with secondary (or less) education. The next strongest predictor was the interaction of education and BLP score in which those with some college or more, regardless of language dominance, show little to no [h] realizations (Figure 8). On the other hand, among those with secondary or less education, language dominance plays more of a role in which higher K’iche’-dominant speakers produce more [h] than speakers who are less K’iche’-dominant or more Spanish-dominant. The model also showed a relatively weak effect for the BLP score in which those with lower negative BLP scores (i.e., very K’iche’-dominant speakers), produce more [h] than those with less negative BLP scores or positive BLP scores (less K’iche’-dominant and Spanish-dominant speakers). Finally, there was a simple effect for sex in which women favored [h] more than men.

Figure 7:

Interaction of following sound and education for [h]/[f] variation.

Figure 8:

Interaction of education and BLP score for [h]/[f] variation (note: these are the actual BLP values as opposed to the scaled values in the model).

To explore the relationship between the fixed factors in the above model, a conditional inference tree was conducted. As seen in Figure 9, the model demonstrates that the most important predictor, similar to the random forest, is that of following sound, in which there are no [h] realizations for other following sounds but is only produced in the context of a following [u̯e] or [u̯i] diphthongs. Within the context of these diphthongs, it is then only those with secondary or less education who produce [h] realizations, while those with some college education or more generally realized an [f] even in the context of these following diphthongs. For those with only secondary or less education, the most K’iche’-dominant speakers (less than −71.65) produce more [h] than those who are less K’iche’ dominant. Among those with secondary (or less) education, women produce more [h] than men.

Figure 9:

Conditional inference tree of factors that predict [h]/[f] variation.

4.2 Fortition: [f] and [p] variation

The random forest results indicate that for [p]/[f] variation, the most important predictor is previous sound, followed by speaker, scaled BLP score, syllabic stress, educational attainment, following sound, scaled age, sex, and finally, sentence position (Figure 10).

Figure 10:

Random forest of [p]/[f] variation.

The best fit Bayesian mixed effect logistic regression model for [p]/[f] variation is shown in Table 2. Of note, the reference level for realization was [f], coded as 0, and [p] realizations were coded as 1. The model selected, that with the lowest LOOIC values (see Appendix IV for loo() model comparison) is more complex than what the frequentist models would allow.

The model revealed that previous sound is the strongest predictor of [p] realizations in which a previous nasal most favors [p] realizations, followed by a previous pause, as compared disfavor [p] realizations (Figure 11A). Another linguistic factor, syllabic stress, demonstrated a simple effect into which atonic syllables favored [p] realizations more than tonic syllables (Figure 11B). The model also showed an effect for the BLP score in which those with lower negative BLP scores (i.e., K’iche’-dominant speakers), produce more [p] than those with less negative BLP scores or positive BLP scores (i.e., Spanish-dominant speakers) (Figure 12A). However, the interaction between BLP and following sound demonstrates that those who are most K’iche’-dominant produce more [p] than those who are less K’iche’-dominant or Spanish-dominant in the context of a previous nasal and also with a previous pause (Figure 13). Finally, the model demonstrated a very weak effect for age in which older speakers produced more [p] realizations than younger speakers (Figure 12B). As mentioned previously, age nearly correlated with BLP scores. Between that, and the low estimate, this simple effect should be taken with caution.

Figure 11:

Simple effects of previous sound (A) and syllabic stress (B) for [p]/[f] variation.

Figure 12:

Simple effect of BLP score (A) and for Age (B) for [p]/[f] variation (note: these are the actual BLP scores and Age values as opposed to the scaled values in the model).

Figure 13:

Interaction of previous sound and BLP for [p]/[f] variation (note: these are the actual BLP scores as opposed to the scaled values in the model).

To explore the relationship between the fixed factors, a conditional inference tree was conducted. As seen in Figure 14, the model demonstrates that the most important predictor is previous sound. For previous nasal segments, the model shows this interacts with BLP score in which the model separates those with a score of −33.6 and even more negative (i.e., K’iche’-dominant speakers) as the speakers most producing [p] with previous nasal segments. Then following the tree, for previous pause, it interacts with syllabic stress in which overall there is more [p] with atonic syllables, but again, among the most K’iche’-dominant speakers. Then for other previous segments, this again interacts with BLP. However, it appears to segment off one K’iche’-dominant speaker (with a score of −78.8) who produced a high number of [p] and thus this particular partitioning should be taken with caution (the same model was run with age and this same speaker (age 46) was separated from the group).

Figure 14:

Conditional inference tree of factors that predict [p]/[f] variation.

5.1 Returning to the research questions

Here we review the results in connection to the research questions and previous studies, followed by theoretical implications. Regarding the first research question (What is the overall frequency of lenition ([h]) and fortition ([p]) realizations for /f/ present in the speech of simultaneous bilingual K’iche’-Spanish speakers?), the overall frequency of /f/ lenition and /f/ fortition is relatively low. Specifically, for lenition, [h] only accounted for 5.0 % of all realizations. This low frequency of [h] realizations supports Utgård’s (2010) findings that [f] is the more common realization with few [h] realizations in Guatemalan Spanish. Similarly, regarding [p] fortition, [p] realizations only accounted for 5.0 % of all realizations. While overall few in number, this data supports previous studies that /f/ fortition does occur in Guatemalan Spanish (Aleza Izquierdo 2010; García Tesoro 2008; Lentzner 1893), while contradicting claims that it no longer occurs in Guatemalan Spanish (Alvar 1980; Utgård 2010). However, the overall low number of [p] tokens among this particular population, which is quite K’iche’-dominant, may help explain these contradictions in the literature. It is certainly feasible that /f/ fortition would not be found at all among groups of speakers in Guatemala that are either monolingual in Spanish or Spanish-dominant bilinguals.

While frequency (Bardovi-Harlig 1987) can play a role in salience, a lack of frequency does not inherently indicate a lack of salience with these features. As Babel (2011) notes in cases of language contact, it is essential to understand the process of enregisterment (Agha 2005, 2007). Agha (2005: 38) defines enregisterment as “processes whereby distinct forms of speech come to be socially recognized (or enregistered) as indexical of speaker attributes by a population of language users.” As seen through Spanish-Quechua contact, Babel (2011: 59) states that enregisterment “can be described as the historical process by which people link linguistic signs to social meanings, producing relationships between signs and social categories (in other words, linguistic features become indices of social categories and stances).” These features tend to occur more in informal conversations than in more formal contexts (Babel 2011: 88). Of importance is that whether these linguistic features are frequent or not, the process of enregisterment allows for such features to be associated with social groups (Babel 2011: 88) so that even infrequent realizations may still index social meaning. For example, the celebrated Guatemalan author Flavio Herrera employed /f/ fortition for words like empermo ‘sick’ when depicting the spoken dialogue of Maya, but not Ladino, characters in his novels (1935: 145). Quantitatively, the results from Baird’s (2023a, 2023b) social perception studies examining the evaluations of stimuli with [f] and [p] realizations for /f/ also support that [p] realizations have undergone this process of enregisterment. Baird found that listener responses indicated that [p] realizations remain a stereotypical feature of the Maya and is highly stigmatized in Guatemalan Spanish. Thus, even if used infrequently, the process of enregisterment (Agha 2005, 2007) allows for [p] realizations to have salient semiotic associations for Guatemalan listeners, including for previous linguists’ descriptions of the Guatemalan Spanish. Furthermore, as orthography (Regan 2022: 503; Trudgill 1986: 11) has been shown to play a role in salience, the phone-to-grapheme mismatch may contribute to making [p] realizations salient, even if infrequent.

In terms of the second research question (What are the linguistic and extra-linguistic factors that condition the processes of lenition and fortition in the speech of simultaneous bilingual K’iche’-Spanish speakers?), we first discuss lenition. The strongest predictor was the linguistic factor of following sound in which /f/ followed by the diphthongs [u̯i] and [u̯e] were most likely to favor lenition. This finding that /f/ followed by /uV/ leads to lenition is supported by previous studies (Penny 2000; Utgård 2010; Van Buren 2017; among others). This particular phonetic context is one that has been in diachronic variation between [f] and [h] across many varieties of Spanish (Penny 2000: 72). The simple effect of education in which those with lower educational attainment favor /f/ lenition more than those with more educational attainment also supports previous studies (Mazzaro 2011; Predmore 1945; Renaud 2014; among others). As indicated by previous scholars who have analyzed /f/ lenition (Mazzaro 2011: 164; Renaud 2014: 267), more formal education indicates speakers have been exposed to more prescriptive phone-to-grapheme realizations. It is of interest, however, that following sound and education interact indicating that those with less educational attainment are most likely to produce the diachronic variation following [u̯i] and [u̯e].

The simple effect of BLP (language dominance), with a very weak estimate, suggests that those who are most K’iche’-dominant favor /f/ lenition more than those who are less K’iche’-dominant or Spanish-dominant. Of all the previous /f/ studies, only Van Buren (2017) examined the role of bilingualisms in /f/ lenition and found what may appear to be the opposite tendency in which Spanish-English bilinguals were less likely to produce [h] than Spanish monolinguals. However, there are several differences between Van Buren’s (2017) bilingual population and those of Nahualá. First, migration plays a significant role for Spanish-English bilinguals in the Pacific Northwest, while it is not a factor in Nahualá, Guatemala; all participants in this study were from Nahualá. Also, English, different from K’iche’, has the phoneme /f/, which does not make this a phonemic conflict cite among Spanish-English bilinguals as it does for K’iche’-Spanish bilinguals. Nevertheless, when considering the social status of these languages, the results of this study corroborate Van Buren (2017). In the Pacific Northwest of the United States and in Nahualá, Guatemala, individuals that speak or are more dominant in the language that symbolizes upwards mobility are less likely to lenite /f/ in Spanish, regardless of what the institutionally prestigious language may be: English in the United States and Spanish in Guatemala. Additionally, while the simple effect of language dominance is of interest here, it is more important to examine the interaction between language dominance and education, which demonstrated a larger estimate in the model. That is, the current study found that /f/ lenition is most common among those who are most K’iche’-dominant who have secondary or less education. On the other hand, K’iche’-dominant speakers with some or more college education demonstrated almost no /f/ lenition similar to Spanish-dominant speakers with some or more college. Thus, /f/ lenition cannot simply be attributed here to language dominance, but rather an example of diachronic variation that varies in large part due to phonetic context and educational level.

The simple effect of speaker sex demonstrated that women produce more [h] than men. The conditional inference tree indicates that this is most true among women with less formal education. This would present a counter example to the Principle 2 of the “Gender paradox” (Labov 1990: 210: 2001 266) that states that in stable linguistic variation, that in general, women produce more of the institutionally prescribed variant than men. However, given the few tokens in this study, this finding should be taken with caution.

Regarding fortition, the strongest predictor was previous sound in which previous nasals most favored /f/ fortition, followed by a previous pause, while other previous sounds disfavored /f/ fortition. These coincide with positional factors that promote fortition cross-linguistically. Specifically, it is more common for fortition to occur in what Ségéral and Scheer (2008: 135) refer to as “strong positions” in which the articulators are already found in positions that facilitate fortition. In Spanish, the most common “strong positions” tend to be immediately following a nasal or a pause. For example, /b, d, g/ are often produced as stops in these positions, or that fortition is maintained, whereas they tend to undergo lenition elsewhere (Hualde 2005; among many others). As such, the speakers in this study that produce /f/ fortition appear to be following a similar fortition/lenition allophonic distribution as in /b, d, g/. There was also a simple effect of syllabic stress in which atonic syllables slightly favored /f/ fortition, although this should be taken with caution as the model indicated low estimates for this simple effect.

The simple effect of BLP (language dominance) indicated that /f/ fortition was favored by those who are most K’iche’-dominant more than those who are less K’iche’-dominant or Spanish-dominant. This supports previous studies in Guatemala that attribute /f/ fortition to bilinguals (Aleza Izauierdo 2010; García Tesoro 2008; Lentzner 1893) as well as previous studies that document /f/ fortition in Spanish due to contact with other languages such as Chavacano in the Philippines (Lipski 1987; Quilis 1996), Wayuunaiki in Colombia (Rodríguez Cadena 2008), Miskitu in Nicaragua (Chappell 2021), and Yucatec Maya in Southern Mexico (Flores Farfán 2000). Finally, it should be noted that BLP interacted with previous sound in which those who were most K’iche’-dominant demonstrated more [p] more so with a previous nasal or a previous pause, indicating that even those who are K’iche’-dominant demonstrated very little [p] realizations with other previous sounds.

The simple effect of age is of note. This may suggest a change in progress (Labov 2001) in this community. However, this notion should be taken with caution for three reasons: (i) the very weak estimate in the model, (ii) the overall number of tokens, and (iii) the near close significant correlation between age and BLP score. Future studies should explore this in depth with more tokens and a larger population size.

5.2 Theoretical implications

The role of education warrants further discussion. In variationist sociolinguistics it has been common to include education, with both occupation and residence value to form a socioeconomic index (Labov 2001: 60–61) while other studies have used education by itself as a proxy for social class (Agüero Chaves 1962; Aleza Izquierdo 2010; Alvar 1980, 2013; Canfield 1981; García Tesoro 2008; Moya 1981; Penny 2000; Quesada Pacheco 2013; Resnick 1975; Zamora Vicente 1996). However, the current study has followed the suggestion from Ash (2013: 356) that sociolinguists examine education as a separate factor from social class. Surprisingly, within the sociolinguistic literature, there is a lack of discussion of the role of orthography within the educational system in sociophonetic variation. This may perhaps be due to the fact that much sociolinguistic theory is based on studies examining English (Docherty and Mendoza-Denton 2012: 47). In general, the English language has a poor phone-to-grapheme correlation, and as such may not be as much as a factor within English sociophonetic variation. The role of orthography, particularly in the context of a grapheme-to-phoneme correspondence, is fundamental to understand /f/ variation in the present study. Orthography has been shown to be key in sociophonetic variation in non-English languages. For example, previous studies have demonstrated a split of a phonetic merger is possible where there are grapheme-to-phoneme correspondences (Lin 2018: 41; Regan 2020: 184, 2022: 503). Thus, if one has increased educational attainment, they will be exposed to prescriptive phone-to-grapheme realizations. As stated by previous scholars (Mazzaro 2011: 164; Renaud 2014: 267), increased educational attainment indicates more exposure to the prescriptive phone-to-grapheme for /f/ realizations in the Spanish-speaking world.

Of importance in the discussion of education is that of language dominance. As much of sociolinguistic theory is based on monolingual varieties, it is important to note that language dominance must be given more consideration. Furthermore, it is noteworthy that education is not necessarily correlated with language dominance in any given community. For example, Baird (2019) found that in Nahualá, many relatively highly educated people make conscious efforts to be K’iche’-dominant. Accordingly, it is crucial in bilingual communities to not only look at possible simple effects of education and language dominance, but at the interaction between the two. In the current study, we see that /f/ lenition ([h]) is most predicted by educational attainment, especially among those who are most K’iche’-dominant, but not simply being K’iche’-dominant favors [h] production. However, then when we examine /f/ fortition ([p]) this does demonstrate a more direct relationship with language dominance.

While the overall frequency of fortition ([p]) and lenition ([h]) is the same, it is interesting to note that different independent variables govern these realizations. While we acknowledge the linguistic factors that govern these realizations (following sound for lenition and previous sound for fortition) as described below, here we briefly focus on the social factors. It is noteworthy that educational attainment was only significant for /f/ lenition. Both /f/ lenition and /f/ fortition present cases of phone-to-grapheme mismatch (read: prescriptively <f> = /f/).^[7] Previously, studies have noted that orthography (Regan 2022: 503; Trudgill 1986: 11) has been shown to play a role in salience. Given this salience, one would predict that more years in the educational system would favor the orthographically prescribed phone (/f/). The results for lenition, a common diachronic process across the (monolingual and bilingual) Spanish speaking world, are in line with this process (and of course in interaction with language dominance). What is interesting is the fact that education was not a significant simple effect or a significant interaction with language dominance. That is, this would indicate that /f/ fortition, in this contact site between K’iche’ and Spanish, is governed by language dominance. In this sense, those who are most K’iche’-dominant, regardless of educational attainment, even if infrequently, favor [p] realizations more than other groups. That is, following Babel’s (2011) use of enregisterment (Agha 2005, 2007), this feature has become associated with those who are K’iche’-dominant. This notion is supported by Baird’s (2023a, 2023b) aforementioned perception studies. This demonstrates the need to examine both educational levels as well as language dominance in situations of language contact.