Abstract
Previous studies disagree as to whether heritage bilinguals demonstrate loss of knowledge of Spanish grammatical gender. As phonetic variability is known to affect the acquisition of certain grammatical markers, we examine whether bilinguals’ gender difficulties relate to bilingual contact-induced phonetic variability, namely, reduction in the inventory of word-final unstressed vowels. We analyzed narratives from children in the United States (n = 49, ages 4–12). All NP s (n = 1415) were analyzed for structure, noun class, and morphology. Word-final vowels were sub-selected for acoustic analyses. Morpho-syntactically, group results show high accuracy with gender (95%), but with wide individual variation (44%–100%). Speakers also show individual variability and substantive numbers of vowel misclassifications (6%–33%) with higher variability for /a/ and /o/. We found bilingual effects in both domains but no association between phonetics and gender accuracy. These findings have implications for the relationship between phonetics and grammar, and for the morphosyntax of Spanish gender.
1 Introduction
There is little consensus as to the status of heritage speakers’ use of Spanish grammatical gender. Descriptions in the child language literature range from virtually error-free to moderately or strongly impacted, and possibly undergoing attrition. Studies that investigate bilingual delays or difficulties in gender systems find association with speaker-external input variables, such as timing, intensity, or length of exposure (Montrul & Potowski, 2007; Martínez-Nieto & Restrepo, 2023). Other studies consider language-level factors, such as transparency of the morphological systems (Eichler et al., 2013), or speaker-internal factors, such as lexical development (Unsworth, 2013). We propose to consider the potential influence of a different speaker-internal variable, i.e., contact-induced underspecification of unstressed vowels. To the extent that this phenomenon is frequent in their communities, bilingual children growing up as heritage speakers have more opaque input than monolingual children. Our question is: to what extent is variability in the realization of unstressed vowels associated with the vulnerability of grammatical gender? Additionally, the existence of underspecified word-final vowels, including in nominal categories, raises the possibility that previous analyses of gender agreement may have led to overestimated accuracy, as suggested by Colantoni et al. (2020). We contribute to these issues by analyzing narrative data in novel ways, where we 1) combine a morphosyntactic analysis with acoustic analyses of unstressed /a e o/ in word-final position; and 2) by focusing on outcomes at the individual level. Before we present our study, we review the status of gender in Spanish-English bilingual children, the possible role of phonetics in grammar development, and bilingual effects in Spanish heritage vowels.
2 Learning Spanish Gender as a Bilingual
Monolingual children acquire gender agreement in Spanish early. Before producing determiners, children frequently produce a prenominal vowel matching noun ending (López Ornat, 1997). The realizations of nouns are variable in form at that stage. López Ornat finds gender harmony subregularities in the sequences of vowel + noun, which can be interpreted as representing a step towards the full morphosyntax of DP s and help explain why gender emerges so early in monolinguals. By age two, accuracy is high, including determiner-noun-adjective sequences (Snyder et al., 2001). Older children’s errors mostly reflect lack of lexical knowledge of gender-opaque nouns, and show clear biases towards phonological cues (Pérez Pereira, 1991).
Bilingual children behave differently. Gender is acquired later (Barreña, 1997; Larrañaga & Guijarro Fuentes, 2012; Eichler et al., 2013). They also differ in the use of early prenominal vocalic fillers that match the noun ending (López Ornat, 1997) and precede the use of gender marked determiners. These forms might be absent (Kuchenbrandt, 2005) or variably reduced in young bilinguals (Ticio Quesada, 2018). Elicitation studies indicate bilingual differences in preschool-aged children. Five-year-old U.S. bilinguals from Mexican families were 76%–91% accurate with noun-adjective agreement (Bedore & Leonard, 2001), far behind monolinguals in Mexico (Grinstead et al., 2009). Morgan et al.’s (2013) direct comparison of monolinguals and bilinguals showed that Southwest U.S. bilinguals displayed significantly more gender substitution errors than their Mexican monolingual counterparts. Type of word matters: more errors appear with so-called opaque nouns. Cuza and Pérez-Tattam (2016) studied gender agreement in DP s with opaque nouns in children between the ages of 5–10, and found high error rates.1 Two thirds of the children gave predominantly mismatched responses to lexically feminine nouns, (e.g., la nube ‘the cloud’), and almost half of the children did the same to masculine nouns (e.g., el tren ‘the train’). This finding shows nouns with non-canonical endings are not learned well. Cuza and Pérez-Tattam’s data for gender concord between determiners and adjectives is even more problematic, since such mismatches suggest that the syntax of gender is not in place. Most bilingual children in their study had chance levels or lower numbers of concord for feminine; and half of the children had chance levels for concord with for masculine. Some of these children seem not to have learned gender.
Corpus studies report fewer gender mismatch errors than elicitation studies (< 4% in Balam et al. 2021; 2.8% vs. 1.7% for monolinguals in Goebel-Marhrle & Shin, 2020). Montrul and Potowski (2007) compared elicited and narrative data from the same children. Narrative data showed fewer errors (5%), concentrated in feminine nouns produced by younger and simultaneous bilingual children. The elicitation task had higher error rates; also concentrated in feminine nouns (34% and 60% errors for sequential and simultaneous bilinguals, respectively). Accuracy improved with age. Martínez-Nieto and Restrepo (2023) found similar age effects: overall gender errors dropped from around 28% in preschool, to about 16% by age 7–8. Accuracy was lower with nouns with opaque, non-canonical endings.
Why are bilingual children less accurate? The general answer, less exposure, clearly applies. Studies that consider the effect of exposure (Gathercole, 2002) or age of onset (Montrul & Potowski, 2007) find that these factors are significant predictors of bilinguals’ performance with Spanish gender. A more concrete answer is that children with less exposure have smaller lexicons. Gender is inherently a lexical phenomenon, and lexical development predicts bilingual gender performance (Nicoladis & Marchak, 2011; Kaltsa et al., 2019). Unsworth (2013) shows that gender is sensitive to both cumulative and current bilingual exposure. Children with smaller lexicons may have learned words but mapped them to the wrong gender. It is also possible, as suggested by an anonymous reviewer, that there are children who have lexical items for which the lemma is not associated with a gender feature.
One possibility not yet explored is that due to consequences of language contact at the level of phonetics, the system to which bilingual children are exposed may be less transparent than for children growing up in monolingual contexts. The so-called canonical -a/-o word markers are final unstressed vowels. Heritage bilingual speech shows substantive contact-induced reduction in the unstressed vowel inventory in Spanish in contact with English (Colantoni et al. 2020). Word final vowel underspecification ([aˈ
Why should vowel phonetics matter? Formal characteristics of the gender system itself may play a role on whether bilingual acquisition is late. Whether gender is acquired early or late is thought to rely on the transparency and consistency of phonological cues to lexical gender, in monolingual and bilingual children. Eichler et al.’s, (2013) cross-linguistic comparison of bilinguals suggests that transparency and reliability of form-based cues in the gender system of a language determines overall age of acquisition. These authors describe Spanish as more transparent than French and other languages. Similarly, Van Heughten and Johnson (2011) argue that phonological transparency and saliency of French gender leads to an earlier acquisition and earlier use of the gender cue in infant word recognition when compared to Dutch. Kupisch et al. (2021) show that phonological cues are as important to bilingual children as to monolinguals.
Within a single language, segmental and prosodic variability can impact the acquisition of grammatical categories. Children’s comprehension and production of English plural -s (Ettlinger & Zapf, 2011) and third-person singular -s (Song et al., 2009; Sundara et al., 2011) are affected by coda complexity and sonority. Similarly, in Spanish, the complexity of the plural allomorph -es (relative to -s, c.f., casas vs. trenes) creates difficulties for monolingual children learning plural forms (Arias-Trejo et al., 2014). Some studies show global delays in language varieties with more phonetic variability. Two-year-olds exposed to multiple English accents recognize familiar words spoken in the regionally dominant variant slower than children in homogeneous linguistic contexts (Buckler et al., 2017). Optional realization delays the acquisition of third person agreement -s in some dialects of English (de Villiers & Johnson, 2007). One especially relevant study is Miller and Schmitt (2012), which examined the acquisition of plurals in Spanish varieties differing on the realization of word-final /s/. Working class Mexican children, learning a dialect with low /s/ lenition, acquired plurals 2–3 years ahead of Chilean children, where -s is variably reduced to [h], or omitted. By controlling for social class across regional dialects, this study establishes a unique contribution of phonetic variability in grammatical acquisition.
We do not know whether vowel variability impacts gender acquisition. Our goal is to explore whether a reduction of the inventory of unstressed vowels due to contact with English is related to the observed bilingual divergence with gender. First, let us consider the properties of the Spanish gender system and its expression.
3 Spanish Gender and Its Expression
Gender is a morphological category that divides nouns into classes and is also reflected in associated categories (Corbett, 1991, p. 1). Spanish nouns bear a feminine or masculine gender feature, which is arbitrary for inanimate nouns, but often reflects the sex specification of some animate/human referents. Both types instantiate a single category: grammatical gender. Alongside number, grammatical gender partakes in the syntactic operation of agreement, through which every element inside the noun phrase (determiners, quantifiers, and adjectives) matches the features of the noun they modify. Long distant dependents (pronouns and predicative adjectives) also match the features of their antecedents, as shown in (1). The same agreement mechanisms rule nouns without canonical endings, as illustrated in (2).
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Experimental studies about gender in Spanish and beyond often make reference to the transparency of the Spanish gender system (or at least part of it) based on two aspects: 1) the correlation between the final vowels -o/-a found in some nouns, and masculine and feminine gender, respectively; and 2) the extent of this pattern (Kuchenbrant, 2005; Alarcón, 2011; Goebel-Mahrle & Shin, 2020; Cuza & Pérez-Tattam, 2015; Martínez Nieto & Restrepo, 2023, Kupisch et al., 2021). A corollary of this type of approach, espoused in some of the previous works (e.g., explicitly assumed: Alarcón, 2011; tacitly assumed: Kuchenbrant, 2005), is the treatment of -o/-a as gender morphemes. This treatment, however, is problematic from the perspective of form/meaning pairing (Harris, 1991). The characterization of the Spanish system as transparent is also problematic when evaluated in the light of other types of nouns found in the language beyond the group of so-called canonical nouns (those whose gender specification and final vowel align). It is a fact that this system also includes so-called opaque nouns (i.e., those that end in -e and in consonants), nouns that show a mismatch between their final vowel and their gender specification, animate nouns that have one form for both genders like estudiante ‘student’ (their gender specification is retrieved via accompanying elements), and epicene nouns such as persona ‘person’ that cannot express the sex specification of the referent.
Based on the complexity that emerges from the brief inventory along with the fact that every noun bears a gender feature in Spanish, we follow the theoretical position that gender is best understood as an abstract property, rather than the addition of a morpheme or an exponent to a root. Specifically, we follow Acquaviva (2020), who suggests that gender is as a property of pieces of nominal morphosyntactic structure related by agreement, rather than a word property shared between morphemes in a structure (pp. 52, 49). With respect to -o/-a endings, we assume Harris’ (1991) view that they are markers of form or class, that have no meaning associated and can appear with either gender (p. 59). Finally, regarding so-called canonical nouns, our position is that the distributional regularities that they display are cues that may aid the acquisition of the gender system in Spanish.2 The impact that these vowels may have on the acquisition of the Spanish gender system is precisely one of the questions that we explore in this paper.
4 Spanish Vowels in Contact
Spanish and English vowel inventories differ in size and on the effect of stress. Spanish’s five-vowel system (/a e i o u/) remains contrastive in unstressed positions (Hualde, 2014, p. 121).3 Monolingual Spanish children learn their vowel inventory early, by age 2;0 (Schnitzer & Krasinski, 1994; Goldstein & Pollock, 2000). Spanish unstressed vowels are significantly shorter than stressed ones but retain quality (Navarro Tomás, 1970; Delattre, 1965; Hualde, 2014). Even in varieties where vowel reduction is common, such as Mexican (e.g., Dabkowski, 2018) or Andean Spanish (Delforge, 2008; Moreno Fernández, 2000), vowel reduction involves devoicing and shortening rather than centralization (c.f., Romanelli et al., 2018, for Argentinean Spanish). Moreover, vowel reduction depends on phonological context (i.e., following voiceless obstruents; Moreno Fernández, 2000), and segment (i.e., /a/ is reduced less than the other vowels; Dabkowski, 2018).
In contrast, the inventory of stressed vowels in North American English is both variable and large, including up to 16 phonemes if gliding is considered (e.g., Labov et al., 2006; Ladefoged, 2001). In unstressed position, the English vowel inventory drops drastically, likely involving two categories [ɪ ə]. Schwa, the most frequent unstressed vowel (Rogers, 2000), lacks a fixed acoustic target. Some authors identify up to seven different types of schwa (Lass, 2009), others just two (Flemming, 2009; Flemming & Johnson, 2007), namely, word-final schwa (China) and pre- or post-tonic word-medial schwa. Consequently, Heritage Spanish-English bilinguals must not only acquire two very different vocalic systems, but must also learn that, in Spanish, but not in English, vowels retain the same quality in stressed and unstressed positions.4
The evidence suggests that the two vowel systems interact in bilinguals. Spanish-English bilingual children show centralization of unstressed vowels in Spanish (Gildersleeve-Neumann et al., 2009; Menke, 2010; Lease 2022).5 It is unclear if they do so before or after entering the school system, when contact with the majority language intensifies (Gildersleeve-Neumann et al., 2009). Various patterns of phonetic influence persist into adulthood, attested in data from New Mexico (Willis, 2005), North Carolina (Boomershine, 2012), the Midwest (Ronquest, 2012, 2013; Shea, 2019), Florida (Alvord & Rogers, 2014), and El Paso (Colantoni et al., 2020). Fronting of /o/ and lowering and fronting of /u/ have also been reported (Willis, 2005; Boomershine, 2012; Ronquest, 2012, 2013; Lease 2022), suggesting skewed vowel spaces, and possible interactions with duration (Boomershine, 2012; Ronquest, 2012). Greater overlap in the vocalic space was also reported in the stressed vowels produced by adult heritage Spanish speakers (Shea, 2019). Such patterns of overlap differ among individuals, with Spanish proficiency accounting for more variability than Spanish dominance. Interestingly, one study (Mazzaro et al., 2016) found differences in the perception of unstressed vowels in bilinguals vs. monolinguals, which suggests that perceptual categorization may underpin production differences.
5 The Study
5.1 Questions and Hypotheses
Although Spanish gender is not a fully transparent system, and gender cannot be reduced to an exponent, there is a clear association between noun endings and noun gender, and some evidence of very early knowledge of this association in monolingual children under age two, and possibly in some bilinguals (Ticio, 2018). The evidence also suggests that, at least for some children, learning gender is not guaranteed. These children might either have lexical gaps for non-canonical nouns or fail to develop a gender system altogether. So, our first question is: do all children acquire a grammar of gender? Our second question is: do other language-internal components, specifically, unstressed vowel quality, determine bilingual gender acquisition? Colantoni et al. (2020) propose the modular interaction hypothesis, which holds that bilinguals’ difficulties with the morphosyntax of gender relates to cross-linguistic influence at the level of phonetics. Vowels are robust cues in monolingual acquisition, which children use alongside with determiners to learn gender. Is gender acquisition in child heritage bilinguals associated with the status of their word-final unstressed vowel system?
We assume that there will be phonetic divergence in our participants, due both to existing variability in the speech community, as suggested in previous literature, and to cross-language influence in the children themselves. This phonetic divergence (to be directly tested in our study) might have a causal relationship to potential divergence in gender grammars. These assumptions are schematized in Figure 1.
Possible modular interactions for bilingual grammars
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
In one scenario, phonetics determines gender acquisition: speakers with blurred vowel boundaries would be the ones with high rates of gender errors. However, this scenario may be too strong. After all, children learn systems that are far less transparent, such as the one in French, and Spanish-speaking children learn the gender of opaque nouns. Thus, the second possibility is that there is limited or no impact of phonetics. This is in line with the idea that gender is an abstract property shared by elements in the nominal domain. In this alternative scenario, loss of grammatical gender does not result from an impoverished unstressed vowel system; other linguistic factors (e.g., the size of the lexicon), or a combination of both factors, will determine grammatical divergence.
In sum, the modular interaction hypothesis predicts that the reduction in the bilingual unstressed vowel inventory will contribute to gender loss. The null hypothesis is that phonetic variability does not influence grammatical success in child bilinguals. Such outcome will have different implications for our understanding of the Spanish gender system and its expression.
5.2 Participants
We analyzed spoken narratives from 49 Spanish-English bilingual children. Most parents were either born in Mexico (n = 33), or from Mexican background (n = 12). Two had one Mexican and one Peruvian parent; the rest did not report. All but two of the children were simultaneous bilinguals: most were U.S.-born (n = 45); two emigrated by age two, the other two emigrated at ages four and seven.
Narratives were elicited in two separate data-collection efforts. One set of samples (n = 28) was collected in 2012 in Lafayette, Indiana. The more recent data (n = 21) were collected in 2019 in Albuquerque, New Mexico (n = 8) and Lafayette (n = 13). Language history questionnaires (from Pérez-Leroux et al., 2011; Cuza et al., 2021) were used to obtain children’s overall fluency ratings in both languages, in a scale from not fluent (= 1) to fully fluent (= 4). Dominance was calculated by subtracting Spanish fluency minus English fluency. Use of Spanish in the family context was a composite of estimates from family members, on a scale of 0 (not used) to 5 (all the time). Children were administered the Test de Vocabulario de Imágenes Peabody (TVIP), a standardized Spanish receptive vocabulary measure (Dunn et al., 1997). Due to availability, not all participants received the questionnaire and the TVIP.
Group language profile
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
5.3 Materials and Task
All participants were invited to retell the story of Little Red Riding Hood (Sebastian & Slobin, 1994), using a wordless picture version (Rhys & Randall, 2008). The children were tested at school or home (Colantoni et al., 2016; Cuza, 2010; Rojas & Iglesias, 2013).
5.4 Data Analyses
Narratives were audio-recorded and transcribed by a native speaker of Spanish. Transcriptions were subsequently verified by two different researchers, also native speakers. Three analyses were conducted. We calculated individual measures of language development from the narrative samples. All NP s were extracted and coded for syntactic structure, noun lexical semantics and morphology. Formant frequencies (F1 and F2) of all final unstressed /a e o/ vowels (adverbs, prepositions, clitics, verbs, in addition to NP s) were measured with Praat (Boersma & Weenink, 2021).
5.4.1 Narrative Measures
Narrative analysis is a standard strategy for assessing development of vocabulary and grammar (Restrepo, 1998; Castilla-Earls et al., 2021). Following standard approaches, we excluded sentence fragments, abandoned sentences, and sentences with unintelligible portions. Mixed sentences were included if mixing was limited to a single word or NP in English. The following measures were obtained:
(1) Narrative Length (NarrLen) Total number of sentences per narrative, reflective of overall language productivity.
(2) Sentence Complexity:
(a) MLU5. Mean length of utterances in words for the five longest utterances (MacWhinney, 2000).
(b) Subordination Index (SubInd). Average number of clauses per sentence over the total sample per individual (Restrepo et al., 2010).
(3) Grammaticality (Gram). Proportion of grammatical sentences over the total number of analyzable sentences per individual (Simon-Cereijido & Gutiérrez-Clellen, 2007). Grammaticality was judged jointly by two native speakers syntacticians. Errors were judged with attention to the language variety and verified with speakers of the dialect and field workers familiar with it. Given the purpose of the study, gender errors were excluded from this count.
(4) Lexical Diversity. As verbal lexicon is an optimal predictor of grammatical development, we employed the number of different verb types (roots) (NDV) as a vocabulary measure (Hadley et al., 2016).
5.4.2 NP Analysis
All NP s in the narratives (tokens = 1415) were extracted and coded for: NP structure, gender and number features of nouns and associated categories (determiners and adjectives), target gender (based on the story for animates), and gender match/mismatch.
Following Colantoni et al. (2020), nouns were coded for lexical-semantic class (human/animate/concrete/abstract/event), countability (individual/ mass/collective/ambiguous) and type of gender morphology (nouns with canonical endings, nouns with gender-specific derivational suffixes, opaque nouns).
NP s were analyzed for constituent structure and coded as bare nouns/simple NP s/complex NP s containing modifiers (adjectives, prepositional phrases or relative clauses). Determiners were classified as definite, indefinite, possessive, demonstrative, quantifier, or absent. Finally, NP s were classified as to whether they contained visible gender information (cf. examples (1) and (2), above). Determiners and modifying adjectives were also coded for gender and number features. Overall accuracy was scored at the phrasal level. An NP was accurate if the noun form matched the target lexical value, and showed no mismatch between agreeing elements (i.e., masculine el abuelito ‘grandfather’ was coded as an error on the basis of story-level information; el casa ‘the house’, for the mismatch between noun and determiner).
Finally, we identified all unstressed final /a e o/ vowels in nominal and verbal categories. Two coders classified them as target /a e o/ vs. centralized or reduced. The proportion of reduced/centralized vowels was calculated for each individual participant.6 This analysis adds a perceptual perspective to the acoustic realization of the canonical vowels and interacts with the perceptual analysis of gender accuracy. Tokens, which were initially coded as non-target in gender but were subsequently identified as having a reduced final vowel, e.g., caperucita as [kapeɾusitə], were recoded as target. The reason for this is that reduced vowels do not offer unambiguous evidence of error, and might be perceived as target by other transcribers (Colantoni et al., 2020). The eleven tokens reclassified constituted the only source of coder disagreement, which amounted to less than 1%.
5.4.3 Acoustic Analyses
All word-final unstressed /a e o/ were labeled and selected. We included only vowels that were either in absolute word final position or in word-final position followed by another word starting with a consonant to avoid coarticulatory effects. Only vowels without overlap in conversation or background noise were acoustically analyzed (N = 1969).7 F1–F2 values in Hz were extracted at five points but we report the central point values to minimize coarticulation effects.8 The F1–F2 values obtained for each individual and each token were submitted to a K-mean cluster analysis. This method is ideal to determine whether each participant’s production of the three underlying vowels can be grouped into one, two or three different categories, independently of the concrete vocalic space of each vowel.
The K-means algorithm (MacQueen, 1967; Hartigan, 1975) is widely used in cluster analysis, due to its characteristics, such as speed of convergence. The goal of the K-means algorithm is to partition the input data (that belong in an n-dimensional space) into K clusters with minimal within-cluster sum of squares. Namely, the aim is to find the cluster centers (ck) and the location for each data point in only one cluster to minimize the within-cluster sum of squares of distances between the data points and center (Bishop, 2006; Murphy, 2012). The algorithm assumes that the number of K clusters is known a priori, achieving homogeneous groupings within each cluster. In our case, the input data is in a 2-dimensional space (F1, F2), and the number of clusters is 3, one per vowel. As such, the K-mean cluster analysis is used to determine the proportion of correctly classified vowels. We assume that for each pair (F1, F2), the true vowel is known a priori based on the sentence context.9 The true vowels are compared with vowel classification obtained using the algorithm where each point in the plane (F1, F2) belongs to a homogeneous group. Once again, we are not interested in vocalic spaces by group, and hence the lack of normalization of formant values, but rather on an analysis by participant, since previous studies (e.g., Shea, 2019) have shown a great degree of variability in patterns of overlap in heritage speakers of Spanish. In addition to the K-means statistics, we will report mean F1–F2 values by participant and by target vowel. Finally, and since we are interested in testing the hypothesis of whether vowel production at the individual level is a predictor of accuracy in gender marking, we will only include a subset of the data (i.e., only vowels obtained from noun phrases) when we discuss the combined grammar-phonetic analysis.
6 Results
6.1 Overall Development and Narrative Measures
The narrative measures extracted revealed a broad range of language skills across lexical diversity (NDV), sentence complexity (MLU5 and SubInd), and accuracy (proportion of grammatical sentences, excluding gender errors and accuracy for gender in NP s, as detailed below).
Descriptive statistics (range, median, mean, and standard deviation (SD)) from individual measures of vocabulary (number of different verb roots (NDV)), general grammar (mean length for five longest utterances (MLU5)), Subordination Index, proportion of grammatical sentences, and gender accuracy (in NPs)
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Pearson correlations between Age, Spanish Fluency, Narrative Length (NarrLen), Vocabulary (TVIP, NDV), general grammar: MLU5, Subordination Index (SubInd), proportion of grammatical sentences (Gram), gender accuracy (Gender)
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
To assess the overall language development, we examined the correlations between narrative measures and two extra-linguistic variables, Age and Spanish Fluency.
We found moderate correlations of age for narrative length and utterance length. With age, children produce longer sentences and tell longer stories. In contrast, their vocabularies (receptive and expressive) are not growing significantly, and their sentences are not becoming more complex or more grammatical. By the same token, we find no evidence of loss in vocabulary or grammatical abilities with age.10 The correlations between language internal measures appear in the expected directions: sentence productivity, productive vocabulary (NDV), and utterance length are significantly correlated; subordination was correlated with receptive vocabulary (TVIP) and with utterance length. Grammaticality and gender accuracy were significantly correlated with each other, and with productive vocabulary.
Only a few external variables were correlated to the global language measures. Spanish fluency was correlated with SubInd (r = .349) and MLU5 (r = .468). Dominance was correlated with NDV (r = .374); and Family Use was correlated with TVIP (r = .565). None of the external variables, including age, was correlated with either gender accuracy or vowel centralization.
6.2 NP Gender Accuracy
We excluded 77 tokens without visible gender, i.e., where neither noun nor determiner provided gender information, as in (3):
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Following our assumption that gender is a property of the entire DP, we also excluded 95 tokens with codeswitching inside the NP (i.e., switched nouns and nouns accompanied by novel mixed determiners), even in cases where the article indicated analogical knowledge of gender, as in (4a), or the noun has the correct vowel realization (4b).
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
A total of 1243 tokens remained for analysis. We observed few identifiable errors (n = 60), with different types of determiners and NP configurations, as shown in (5).
Proportion of accurate gender in NP s, classified by determiner type (frequencies in parenthesis)
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Feminine nouns were more frequent in these narratives than masculine nouns (876 to 366).11 Unlike in previous studies, children were not more accurate with masculine (93%) relative to feminine nouns (96%). At the group level, we found no identifiable effects of morphological and semantic parameters in accuracy rates. Children were scarcely more accurate with nouns in gender alternations vs. non-alternating nouns (carro/*carra) (95.6% vs. 94.4%); transparent vs. opaque nouns (96% vs. 92%) or animate vs. inanimate nouns (95.6% vs. 94.5%).
A key indicator of gender within the NP is the determiner, but only singular definite and indefinite determiners offer segmental information beyond vowel quality. In Table 4 we report accuracy rate for NP s grouped according to determiner type.12 A comparison across NP types shows that overall accuracy is lower in plurals than singulars, and lower in definite and indefinites than in other determiner types. No errors were observed with nouns not accompanied by determiners, or in demonstrative NP s. Overall gender accuracy can be considered relatively high.13
6.3 Phonetic Analysis
To determine whether children can distinguish the relevant vowels in production, we started by excluding children with insufficient data, which precludes estimating the degree of misclassification in vowel pairs. This eliminated four participants who did not produce all three vowels, and five more who produced less than five tokens per vowel. The mean F1–F2 values for all children analyzed are reported in Appendix B. The data reveal prototypical examples of participants who distinguish the three vowels (e.g., BC13), who merge pairs of vowels (e.g., BC11, BC21) or who have a centralized vocalic space with minimal overlap (e.g., BC14). Figure 2 plots the vocalic spaces of these 4 participants (Appendix C displays the vocalic spaces of all the participants analyzed).
F1–F2 values for the vowels /a e o/ by participant. Note: Ellipses include values that are within 1 s.d. from the mean. Vowel spaces were plotted with the R package phonR (McCloy, 2016)
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
The K-means cluster analysis (see Fig. 3 and Appendix A) includes the 32 participants who produced at least five tokens of the three vowels.14 The horizontal axis displays the underlying vowels, while the vertical axis indicates the percentage of tokens that were classified and the bars show the patterns of misclassification. For example, underlying /o/ tokens for participant BC13 (Fig. 2a) were mostly classified as /o/; a small proportion of underlying /a/ tokens were also classified as /o/ and an even smaller proportion of /e/ tokens were grouped into the /o/ category. Participants were equally distributed among those whose vowels had relatively good classificatory accuracy (i.e., if all three vowels were accurately classified at 50% or higher; Fig. 3a) and poor classificatory accuracy (Fig. 3b)).15
Results of K-mean classification of intended target segment /a e o/ per participant. Stacked bars represent the proportion of the classification response in the output of the algorithm (3a): Participants with more than 50% accurate classification for all vowels (3b): Participants with less than 50% accurate classification for any of the vowels
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
These results (see Appendix A) reveal a great degree of individual variation in classificatory accuracy. For many speakers one vowel was classified as another (e.g., CH10, [o] for [a]) or two vowels were classified as a third one (e.g., CH7: [e o] misclassified as [a]). The most frequently misclassified vowel was /a/, which may have implications for feminine nouns (see section 5.4). The opposite pattern ([o] tokens classified as [a]) was also attested (four participants whose vowels pairs displayed over 40% of misclassification). Of these participants, BC2 represents an interesting example, since over 50% of both [a] and [o] tokens were classified as the opposite vowel. In sum, these two analyses indicate many of these heritage speakers produce final unstressed Spanish vowels that occupy overlapping acoustic areas, with blurred phonetic boundaries.
6.4 Combined Analysis
The grammatical and the phonetic analyses demonstrate a substantive individual variation, as shown in Table 5. The distribution of speakers was positively skewed for gender accuracy: most children were highly accurate, but a few had low accuracy, as shown below in Figure 4. Perceptual analyses also showed substantive variation, with a negatively skewed distribution. As for classificatory accuracy (K-means), we found more individuals around the mean, but with the rest distributed across the variation space.
Descriptive statistics: K-means classificatory accuracy, gender accuracy, and perceived quality of final vowels
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Top panels represent histograms of individual participants classified according to Gender Accuracy, Proportion of Centralized/Reduced Vowels, and Noun Classification Accuracy. Bottom panels represent the correlations of gender accuracy with Age, Vowel Centralization (perceptual), and Vowel Classification Accuracy (K-means).
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Having established that variation is present, we examine whether there is a potential association between the grammatical and the phonetic parameters. These correlations are given in Table 6. The perceptual classification of vowels has significant negative correlations with age, vocabulary, and gender accuracy. Children identified as reducing their unstressed vowels were younger, had smaller vocabularies, and were less accurate with gender. In contrast, the acoustic classification was not associated with age, vocabulary, or gender. The strongest predictor for gender accuracy was vocabulary diversity, NDV, also included here.
Correlation between age, vocabulary, gender, and individual proportion of reduced/centralized vowels as per perceptual coding
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
This result is indeterminate. It is possible that classificatory accuracy was not significant because of the data loss associated with requirements of these acoustic analyses (e.g., individual exclusions, exclusions of nouns ending in a vowel followed by another word beginning with a vowel, which were retained in the perceptual analysis). The perceptual measure included all individuals, and more tokens by individual than the acoustic analysis. Another possibility is that the perceptual analysis, although less precise, might be capturing other phonetic parameters or combinations of parameters, beyond formant values. The third possibility is that overall gender accuracy does not provide a sufficient characterization of the individual grammars. In Table 7, where we characterize individual performance, we see that overall gender accuracy masks several atypical patterns of gender use.
Type of speakers classified according to their use of gender
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
We observe distinct behaviors. Three children who produced English or hybrid determiners (the, da) had high gender accuracy, but low overall grammaticality. These children produced some gender-marked nouns, but no matching between determiner and noun. They may know the form of given words, but it is not clear that they know the morphosyntax of gender. Other children (Feminine bias, Masculine default) appear to be simplifying the formal system. One child variably matched articles of different genders to the same noun.
Using this individual analysis, we now ask whether children who are in these atypical gender categories are more likely to have low acoustic classificatory accuracy (a proxy for high overlap between vocalic spaces) and/or low levels of target vowel realization as per the perceptual analysis (i.e., high rates of centralized or reduced vowels). We used the median K means classification scores to divide children into high (median or above) and low accuracy groups (below the median) as shown in (Table 8); similarly for vowel realization according to the perceptual coding (Table 9). The results show no evidence of association between vowel realization and patterns of gender use.
Cross-classification of individuals by pattern of gender use, and vowel classificatory accuracy (K-Means)
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Cross-classification of individuals by pattern of gender use, and levels of target vowel realization as per the perceptual analysis
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
These results show various behaviors. Sometimes, phonetics and grammar went together, as was the case for BC02, who has low gender accuracy and low vowel classificatory accuracy (but high perceptual accuracy). We observed the same for the masculine default child, BC04. One child, BC11, is at ceiling for gender, but shows low vowel classification accuracy. Conversely, CH05 exhibited low gender, but high vowel classification accuracy. That some participants did not produce tokens of the three final vowels deserves further thought. In some cases, it may be interpreted as an avoidance strategy. The case of CH24 is striking: this child only produced lexical items with underlying /a e/ vowels; all determiners were feminine and expected words ending in /o/, such as lobo, were substituted with the English equivalent ‘wolf’.
It is important to note that while some of these grammatical patterns are instantiated by few individuals, the main point is that the individuals do exist, and they are no more likely than those with relatively intact gender grammars to be those individuals with low phonetic accuracy. Our data show no discernible association between phonetic parameters and development of a relatively typical vs. atypical gender grammar. Neither perceptual nor acoustic classificatory accuracy can explain the grammatical patterns obtained across the different children.
7 Discussion
We analyzed data from Spanish heritage children for difficulties with grammatical gender, and for vowel realizations. We subsequently explored whether bilingual effects at the level of grammar (reduction of accuracy in gender realization and/or reorganization of gender grammars) were associated with bilingual effects at the level of phonetics (lack of separation in the vocalic spaces).
Was the grammar of gender affected by the heritage context? The answer is yes. Although many children were highly accurate, and many more have few errors, a full one third of the children are not yielding evidence of gender contrasts. Although a finding of 95% overall accuracy would suggest that gender is an invulnerable domain in heritage acquisition, the individual analysis indicates that this is not universal. Our overall results are not entirely atypical, with error rates comparable to those of Montrul and Potowski’s (2007) narrative data. They diverge from studies reporting minimal error rates for determiners, nouns, and adjectives such as Goebel-Marhle and Shin (2021). There are various explanations for some of these differences. First, our grammatical analysis revealed that many NP s in narratives are not informative for gender. Here we count not only those with opaque nouns and opaque determiners (e.g., mi tren ‘my train’), but also those where the only expression of gender are thematic vowels (e.g., los lobos ‘wolves’), which our acoustic and perceptual analyses reveal to be uncertain in nature. Another issue is that most studies (to the exception of Goeble-Mahrle & Shin, 2021) do not indicate lexical exclusion criteria. It is unclear whether accuracy counts could include NP s without visible gender. Finally, previous studies (i) transcribed the expected underlying vowels (i.e., /a, e, o/); (ii) do not include a centralized vowel option; and (iii) only report perceived error. Our results suggest that naïve transcribers (i.e., not trained phoneticians) are likely coding many vowels realized as schwa, according to their lexical expectations. Given the number of NP s whose only gender indication is vowel quality, we consider it possible that our previous results (including our current perceptual coding) overestimate grammatical accuracy.
The phonetic analysis also revealed the predicted cross-linguistic influence. Both analyses conducted over measures of vowel quality (i.e., the K-means classificatory accuracy and grouping of each of the vowels in F1–F2 acoustic space) concur. For many speakers, the three unstressed word-final vowels do not consistently differ in quality. This was corroborated by our perceptual coding.
As for our main question, the combined analysis shows that acoustic analyses do not predict the status of an individual’s gender system. The small correlation of perceptual coding with overall gender accuracy was driven by a small subset of younger children, who reduced vowels, had small vocabularies and low gender accuracy. In contrast to the low explanatory value of our phonetic measures, vocabulary diversity had the strongest (negative) association with gender errors, as predicted by Unsworth (2013). Vocabulary was also correlated with overall grammaticality.
To further explore the role for phonetics in grammar, we divided children into those that were highly accurate and those that were not. Again, we failed to find support for modular interaction, which would predict individuals with more undifferentiated vowel spaces would also be the ones with divergent gender system. Evidently, cleaner and more controlled data are needed to further evaluate these conclusions. For now, our results are most compatible with the null hypothesis, that there is not an association between phonetics and the grammar of gender. Most of our children are highly accurate in gender, but many have very indistinct vowel spaces. Other children have clearly divergent gender patterns, but their vowel realizations are not particularly impoverished.
If these results replicate, we should ask why gender is different from plural and verbal agreement final -s. One difference could be vowels vs. consonants. Another possibility is the lexical nature of gender, a factor that is less central to the other categories. This would affect the developmental timelines: an early phonetic effect might introduce a delay but disappear once children learn enough vocabulary.16 These important possibilities lie beyond the scope of the current study.
Alternatively, this null result is the answer. The Spanish gender system is more complex than just the final vowels. From our perspective, the cues to gender are noun endings and article form (or more broadly, concord patterns, with a central role for bootstrapping from the article).17 Gender systems are abstract and tolerate a fair amount of poverty of the phonological cues (for similar observations, see Kupisch, 2021). Learners can learn gender systems with few overt phonological cues, as in French. Spanish heritage bilinguals with great overlap in vowel space can still learn gender, much like French-speaking children do.18 A bilingual child with a reduced unstressed vowel inventory will still have a powerful distributional cue to gender, that is, determiners.19 For such learners, the system might be less transparent, but still learnable.20 Our study also suggests that listeners and transcribers at times fail to perceive these realizations as divergent. When the coding is restricted to /a/ vs. /o/, listeners and transcribers find few errors. But when invited to identify central or reduced realizations, more possibilities are detected; and the frequency of reduced vowel detection increases even more when the transcriber is a phonetician. Vowel perception offers plenty of room for accommodation. Perhaps this is what distinguishes this case from the case of final -s morphemes, where phonetic variability does predict individual and group level performance in grammar.
In conclusion, current results do not support the modular interaction hypothesis (i.e., for the interaction between phonetics and gender grammar), although they provide support for the role of vocabulary. Our results highlight the need for research that links the acquisition of grammatical morphemes and their phonetic realizations. They also demonstrate the importance of individual analyses in the study of heritage language acquisition.21 There are various ways of being a heritage speaker, and heritage grammars may have different shapes, representing interesting clusters of solutions to grammatical differences between the two languages of the heritage speaker. Such diversity might remain invisible in reports of group results. Our analysis of narrative data confirms there is more ambiguity in gender marking in spontaneous speech than previously thought. Methodologically, this shows that coding approaches impact estimates of accuracy. Finally, at a theoretical level, our results raise important questions about the learnability of the Spanish system, the role of vowels usually associated with gender, and the abstract and robust, but not invulnerable nature of the system.
Acknowledgement
This research was supported by an Insight grant from the Social Sciences and Humanities Research Council of Canada (SSHRC), awarded to L. Colantoni and A. T. Pérez-Leroux for the project “The impact of phonetic variability on bilingual morpho-syntax: The case of Spanish-English bilinguals” (2020–2025). Direct correspondence to A. T. Pérez-Leroux, Dept. of Spanish and Portuguese, University of Toronto. Room 208 Victoria College, 91 Charles St. West, Toronto, ON M5S 1K7, Canada. Email: at.perez.leroux@utoronto.ca.
Notes on Contributors
Ana T. Pérez-Leroux is Professor of Spanish and Linguistics at the University of Toronto. Her work on child language acquisition investigates the developmental interactions between syntax and other domains, and between the two languages of bilingual children.
Yadira Álvarez López (PhD, University of Toronto, 2019) is Research Associate at the University of Toronto. Dr. Álvarez López’s research is in theoretical syntax, specifically, argument realization and the interaction between argument structure and information structure in the structure of the Spanish clause. She has published work on weather predicates, copulas, and subjects.
Miguel Barreto is a PhD candidate in the Department of Statistical Sciences at the University of Toronto. His research focuses on statistical computing, Bayesian models, and machine learning.
Alejandro Cuza is Professor of Spanish and Linguistics at Purdue University. His research focuses on the acquisition of Spanish morphosyntax and semantics among child and adult heritage speakers of Spanish.
Irina Marinescu (PhD Hispanic Linguistics, University of Toronto, 2012) specializes in phonetics and experimental approaches to the acquisition of second language phonology. She has published on realization and perception of vowels across native and non-native varieties, with particular interest in synchronic and diachronic variation in Spanish and other Romance languages.
Jierui Yang has an undergraduate degree in Cognitive Science and Computer Science from the University of Toronto, and a Masters in Computational Linguistics from the University of British Columbia.
Laura Colantoni is a Professor at the Department of Spanish and Portuguese at the University of Toronto. Her research focuses on sound change and categorization and the second language acquisition of variable phonetic parameters.
Appendices
Percentage of /a e o/ vowels classified as non-target by participant and vocalic contrast
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Mean F1–F2 values (in Hz) by participant and vowel
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
Vocalic space for the vowels /a e o/ by participants (Note: Ellipses include values that are within 1 s.d. from the mean)
Citation: Heritage Language Journal 20, 1 (2023) ; 10.1163/15507076-bja10017
References
Acquaviva, P. (2020). Gender as a property of words and as a property of structures. Catalan Journal of Linguistics, 19, 49–74. https://doi.org/10.5565/rev/catjl.313.
Alvord, S., & Rogers, B. (2014). Miami-Cuban Spanish vowels in contact. Sociolinguistic Studies, 8(1), 139–170. https://doi.org/10.1558/sols.v8i1.139.
Arias Rodríguez, I. (2016). Cálculo de frecuencias de aparición de fonemas y alófonos en español actual utilizando un transcriptor automático [Frequency of occurrence of phonemes and allophones in contemporary Spanish as calculated by an automatic transcription system]. Loquens, 3(1), Article e029. https://doi.org/10.3989/loquens.2016.029.
Arias-Trejo, N., Abreu-Mendoza, R. A., & Aguado-Servín, O. A. (2014). Spanish-speaking children’s production of number morphology. First Language, 34(4), 372–384. https://doi.org/10.1177/0142723714544411.
Balam, O., Lakshmanan, U., & Parafita Couto, M. (2021). Gender assignment strategies among simultaneous Spanish/English bilingual children from Miami, Florida. Studies in Hispanic and Lusophone Linguistics, 14(2), 241–280. https://doi.org/10.1515/shll-2021-2045.
Barreña, A. (1997). Desarrollo diferenciado de sistemas gramaticales en un niño vasco-español bilingüe [Differentiated development of grammatical systems in a bilingual Basque-Spanish child]. In A. T. Pérez-Leroux & W. R. Glass (Eds.), Contemporary perspectives on the acquisition of Spanish: Vol. 1: Developing grammars (pp. 55–74). Cascadilla Press.
Bedore, L., & Leonard, L. B. (2001). Grammatical morphology deficits in Spanish- speaking children with specific language impairment. Journal of Speech, Language, and Hearing Research, 44(4), 905–924. https://doi.org/10.1044/1092-4388(2001/072).
Bishop, C. M. (2006). Pattern recognition and machine learning. Springer-Verlag.
Boersma, P., & Weenink, D. (2021). Praat: Doing phonetics by computer (Version 6.1.43) [Computer software]. Retrieved from http://www.praat.org/.
Boomershine, A. (2012, October 23–28). What we know about the sound system(s) of Heritage Speakers of Spanish: Results of a production study of Spanish and English bilingual and heritage speakers [Conference presentation]. 12th Hispanic Linguistic Symposium, University of Florida, Gainesville, FL, United States.
Buckler, H., Oczak-Arsic, S., Siddiqui, N., & Johnson, E. K. (2017). Input matters: Speed of word recognition in 2-year-olds exposed to multiple accents. Journal of Experimental Child Psychology, 164, 87–100. https://doi.org/10.1016/j.jecp.2017.06.017.
Castilla-Earls, A. P., Pérez-Leroux, A. T., & Auza, A. (2021). Elicited vs. spontaneous language as methods for the assessment of grammatical development: The DEME assessment tool. Revista de Logopedia, Audiología y Foniatría, 41(4), 164–171. https://doi.org/10.1016/j.rlfa.2021.02.002.
Clegg, J. H. (2011). A frequency-based analysis of the norms for Spanish noun gender. Hispania, 94(2), 303–319.
Colantoni, L., Martínez, R., Pérez Leroux, A. T., & Rinaldi, N. (2020). A phonetic account of heritage bilinguals’ divergence with Spanish gender agreement. Languages, 5(4), 58. https://doi.org/10.3390/languages5040058.
Corbett, G. G. (1991). Gender. Cambridge University Press.
Cuza, A. (2010). The L2 acquisition of aspectual properties in Spanish. Canadian Journal of Linguistics, 55(2), 181–208. https://doi.org/10.1017/S0008413100001468.
Cuza, A., & Pérez Tattam, R. (2016). Grammatical gender selection and phrasal word order in child heritage Spanish: A feature re-assembly approach. Bilingualism: Language and Cognition, 19(1), 50–68. https://doi.org/10.1017/S1366728914000893.
Cuza, A., Shin, N., & Sánchez, L. (2021, October 7–9). ¿Los perros, perros or both? Definite plural subjects in child heritage Spanish? [Conference Presentation]. Hispanic Linguistics Symposium, Wake Forest University, Winston-Salem, NC, United States.
Dabkowski, M. (2018). Variable vowel reduction in Mexico City Spanish. [Unpublished doctoral dissertation]. Ohio State University.
Delattre, P. (1965). Comparing the phonetic features of English, French, German, and Spanish: An interim report. Julius Groos Verlag.
Delforge, A. M. (2008). Unstressed vowel reduction in Andean Spanish. In L. Colantoni & J. Steele (Eds). Selected Proceedings of the 3rd Conference on Laboratory Approaches to Spanish Phonology (pp. 107–124). Cascadilla Proceedings Project.
de Villiers, J., & Johnson, V. (2007). The information in third-person /s/: Acquisition across dialects of American English. Journal of Child Language, 34(1), 133–158. https://doi.org/10.1017/S0305000906007768.
Dunn, L. M., Lugo, D. E., Padilla, E. R., & Dunn, L. (1997). Test de vocabulario en imágenes Peabody. American Guidance Service.
Eichler, N., Jansen, V., & Müller, N. (2013). Gender acquisition in bilingual children: French-German, Italian-German, Spanish-German and Italian-French. International Journal of Bilingualism, 17(5), 550–572. https://doi.org/10.1177/1367006911435719.
Ettlinger, M., & Zapf, J. (2011). The role of phonology in children’s acquisition of the plural. Language Acquisition, 18(4), 294–313. https://doi.org/10.1080/10489223.2011.605044.
Flemming, E. (2009). The phonetics of schwa vowels. In D. Minkova (Ed.). Phonological weakness in English: From Old to Present-Day English (pp. 78–98). Palgrave Macmillan.
Flemming, E., & Johnson, S. (2007). Rosa’s roses: Reduced vowels in American English. Journal of the International Phonetic Association, 37(1), 83–96. https://doi.org/10.1017/S0025100306002817.
Gathercole, V. C. M. (2002). Grammatical gender in bilingual and monolingual children: A Spanish morphosyntactic distinction. In K. Oller & R. Eilers (Eds.), Language and literacy in bilingual children (pp. 207–219). Multilingual Matters.
Gildersleeve-Neumann, C., Peña, E., Davis, B., & Kester, E. (2009). Effects on L1 during early acquisition of L2: Speech changes in Spanish at first English contact. Bilingualism: Language and Cognition, 12(2), 259–272. https://doi.org/10.1017/S1366728908003994.
Goebel-Mahrle, T., & Shin, N. L. (2020). A corpus study of child heritage speakers’ Spanish gender agreement. International Journal of Bilingualism, 24(5–6), 1088–1104. https://doi.org/10.1177/1367006920935510.
Goldstein, B., & Pollock, K. (2000). Vowel errors in Spanish-speaking children with phonological disorders: A retrospective comparative study. Clinical Linguistics and Phonetics, 14(3), 217–234. https://doi.org/10.1080/026992000298832.
Grinstead, J., de la Mora, J., Vega-Mendoza, M., & Flores, B. (2009). An elicited production test of the optional infinitive stage in child Spanish. In J. Crawford, K. Otaki, & M. Takahashi (Eds.), Proceedings of the 3rd Conference on Generative Approaches to Language Acquisition North America (GALANA 2008) (pp. 36–45). Cascadilla Proceedings Project.
Guirao, M., & García Jurado, M. A. (1990). Frequency of occurrence of phonemes in American Spanish. Revue Québécoise de Linguistique, 19(2), 135–149. https://doi.org/10.7202/602680ar.
Hadley, P., Rispoli, & Hsu, N. (2016). Toddlers’ verb lexicon diversity and grammatical outcomes. Language, Speech, and Hearing Services in Schools, 47(1), 44–58. https://doi.org/10.1044/2015_LSHSS-15-0018.
Harris, J. (1991). The exponence of gender in Spanish. Linguistic Inquiry, 22(1), 27–62.
Harris, J. (1992). The form classes of Spanish substantives. In G. Booij & J. van Marle (Eds.), Yearbook of Morphology 1991 (pp. 65–88). Springer.
Hartigan, J. A. (1975). Clustering algorithms. John Wiley.
Hualde, J. I. (2014). Los sonidos del español [The sounds of Spanish]. Cambridge University Press.
Kaltsa, M., Tsimpli, I. M., & Argyri, F. (2019). The development of gender assignment and agreement in English-Greek and German-Greek bilingual children. Linguistic Approaches to Bilingualism, 9(2), 253–288. https://doi.org/10.1075/lab.16033.kal.
Kuchenbrandt, I. (2005). Gender acquisition in bilingual Spanish. In J. Cohen, K. T. McAlister, Rolstad, K., & J. MacSwan (Eds.), Proceedings of the 4th International Symposium on Bilingualism (pp. 1252–1263). Cascadilla Proceedings Project.
Kupisch, T., Mitrofanova, N., & Westergaard, M. (2021). Phonological vs. natural gender cues in the acquisition of German by simultaneous and sequential bilinguals (German-Russian). Journal of Child Language, 49(4), 661–683. https://doi.org/10.1017/S0305000921000039.
Labov, W., Ash, S., & Boberg, C. (2006). The atlas of North American English: Phonetics, phonology, and sound change. Walter de Gruyter.
Ladefoged, P. (2001). Vowels and consonants: An introduction to the sounds of language. Blackwell.
Larrañaga, P., & Guijarro Fuentes, P. (2012). Clitics in L1 bilingual acquisition. First Language, 32(1–2), 151–175. https://doi.org/10.1177/0142723711403878.
Lass, R. (2009). On schwa: Synchronic prelude and historical fugue. In D. Minkova (Ed.). Phonological weakness in English: From Old to Present-Day English (pp. 47–77). Palgrave Macmillan.
López Ornat, S. (1997). What lies in between a pre-grammatical and a grammatical representation? In A. T. Pérez-Leroux & W. R. Glass (Eds.), Contemporary perspectives on the acquisition of Spanish: Vol. 1: Developing grammars (pp. 3–20). Cascadilla Press.
MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations. Berkeley Symposium on Mathematical Statistics and Probabilities, 5.1, 281–297. https://digitalassets.lib.berkeley.edu/math/ucb/text/math_s5_v1_article-17.pdf.
MacWhinney, B. (2000). The CHILDES Project: Tools for analyzing talk (3rd ed). Lawrence Erlbaum Associates.
Mariscal, S. (2008). Early acquisition of gender agreement in the Spanish noun phrase: Starting small. Journal of Child Language, 36(1), 143–171. https://doi.org/10.1017/S0305000908008908.
Martin, A. E., Monahan, P. J., & Samuel, A. G. (2017). Prediction of agreement and phonetic overlap shape sublexical identification. Language and Speech, 60(3), 356–376. https://doi.org/10.1177/0023830916650714.
Martínez-Nieto, L., & Restrepo, M. A. (2023). Grammatical gender in Spanish child heritage speakers. Incomplete or different acquisition? Linguistic Approaches to Bilingualism, 13(2), 267–297. https://doi.org/10.1075/lab.20042.mar.
Mazzaro, N., Colantoni, L., & Cuza, A. (2016). Age effects and the discrimination of consonantal and vocalic contrasts in heritage and native Spanish. In C. Tortora, M. den Dikken, I. Montoya, & T. O’Neill (Eds.), Romance Linguistics 2013: Selected papers of the 43th Linguistic Symposium on Romance Languages (LSRL), New York, 17–19 April, 2013 (pp. 277–300). John Benjamins.
McCloy, D. (2016). PhonR (Version 1.0–7) [Computer software]. Retrieved from http://drammock.github.io/phonR/.
Menke, M. (2010). Examination of the Spanish vowels produced by Spanish-English bilingual children. Southwest Journal of Linguistics, 28(2), 98–135.
Miller, K., & Schmitt, C. (2012). Variable input and the acquisition of plural morphology. Language Acquisition, 19(3), 223–261. https://doi.org/10.1080/10489223.2012.685026.
Montrul, S., & Potowski, K. (2007). Command of gender agreement in school-age Spanish-English bilingual children. International Journal of Bilingualism, 11(3), 301– 328. https://doi.org/10.1177/13670069070110030301.
Moreno Fernández, F. (2000). Qué español enseñar [What Spanish to teach]. Arco Libros.
Moreno Sandoval, A., Toledano, D., de la Torre, R., & Garrote, M. (2008). Developing a phonemic and syllabic frequency inventory for spontaneous spoken Castilian Spanish and their comparison to text-based inventories. In Proceedings of the International Conference on Language Resources and Evaluation, LREC 2008. European Language Resources Association (ELRA). http://www.lrec-conf.org/proceedings/lrec2008/pdf/283_paper.pdf.
Morgan, G., Restrepo, M. A., & Auza, A. (2013). Comparison of Spanish morphology in monolingual and Spanish-English bilingual children with and without language impairment. Bilingualism: Language and Cognition, 16(3), 578–596. https://doi.org/10.1017/S1366728912000697.
Murphy, K. P. (2012). Machine learning: A probabilistic perspective. MIT Press.
Navarro Tomás, T. (1970). La pronunciación del español [Spanish Pronunciation]. CSIC.
Nicoladis, E., & Marchak, K. (2011). Le carte blanc or la carte blanche? Bilingual children’s acquisition of French adjective agreement. Language Learning, 61(3), 734–758. https://doi.org/10.1111/j.1467-9922.2011.00660.x.
Pérez-Leroux, A. T., Castilla-Earls, A. P., & Brunner, J. (2012). General and specific effects of lexicon in grammar. Journal of Speech, Language and Hearing Research, 55(2), 313–327. https://doi.org/10.1044/1092-4388(2011/10-0004).
Pérez-Leroux, A. T., Cuza, A., & Thomas, D. (2011). Input and parental attitudes: A look at Spanish-English bilingual development in Toronto. In K. Potowski (Ed.), Bilingual youth: Spanish in English-speaking societies (pp. 49–175). John Benjamins.
Perez Pereira, M. (1991). The acquisition of gender: What Spanish children tell us. Journal of Child Language, 18(3), 571–590. https://doi.org/10.1017/S0305000900011259.
Restrepo, M. A. (1998). Identifiers of predominantly Spanish-speaking children with language impairment. Journal of Speech, Language, and Hearing Research, 41(6), 1398–1411. https://doi.org/10.1044/jslhr.4106.1398.
Restrepo, M. A., Castilla, A. P., Schwanenflugel, P. J., Neuharth-Pritchett, S., Hamilton, C. E., & Arboleda, A. (2010). Effects of a supplemental Spanish oral language program on sentence length, complexity, and grammaticality in Spanish-speaking children attending English-only preschools. Language, Speech and Hearing Services in Schools, 41(1), 3–13. https://doi.org/10.1044/0161-1461(2009/06-0017).
Rhys, A., & Randall, R. (2008). Treasured tales: Little red riding hood. Paragon Publishing.
Rogers, H. (2000). The sounds of language: An introduction to phonetics. Longman.
Rojas, R., & Iglesias, A. (2013). The language growth of Spanish-speaking English language learners. Child Development, 84(2), 630–646. https://doi.org/10.1111/j.1467-8624.2012.01871.x.
Romanelli, S., Menegotto, A., & Smyth, R. (2018). Stress-induced acoustic variation in L2 and L1 Spanish vowels. Phonetica, 75(3), 190–218. https://doi.org/10.1159/000484611.
Ronquest, R. (2012). An acoustic analysis of heritage Spanish vowels [Unpublished doctoral dissertation]. Indiana University.
Ronquest, R. (2013). An acoustic examination of unstressed vowel reduction in heritage Spanish. In C. Howe, S. Blackwell, & M. Lubbers Quesada (Eds.), Selected Proceedings of the 15th Hispanic Linguistic Symposium (pp. 151–171). Cascadilla Proceedings Project.
Schnitzer, M., & Krasinski, E. (1994). The development of segmental phonological production in a bilingual child. Journal of Child Language, 21(3), 585–622. https://doi.org/10.1017/S0305000900009478.
Sebastián, E., & Slobin, D. I. (1994). Development of linguistic forms: Spanish. In R. A. Berman & D. I. Slobin (Eds.), Relating events in narrative: A crosslinguistic developmental study (pp. 239–284). Erlbaum.
Shea, C. (2019). Dominance, proficiency and Spanish heritage speakers’ production of English and Spanish vowels. Studies in Second Language Acquisition, 41(1), 123–149. https://doi.org/10.1017/S0272263117000328.
Simon-Cereijido, G., & Gutiérrez-Clellen, V. F. (2007). Spontaneous language markers of Spanish language impairment. Applied Psycholinguistics, 28(2), 317–339. https://doi.org/10.1017/S0142716407070166.
Snyder, W., Senghas, A., & Inman, K. (2001). Agreement morphology and the acquisition of noun-drop in Spanish. Language Acquisition, 9(2), 157–173. https://doi.org/10.1207/S15327817LA0902_02.
Song, J. Y., Sundara, M., & Demuth, K. (2009). Phonological constraints on children’s production of English third person singular -s. Journal of Speech, Language, and Hearing Research, 52(3), 623–642. https://doi.org/10.1044/1092-4388(2008/07-0258).
Sundara, M., Demuth, K., & Kuhl, P. K. (2011). Sentence-position effects on children’s perception and production of English third person singular -s. Journal of Speech, Language, and Hearing Research, 54(1), 55–71. https://doi.org/10.1044/1092-4388(2010/10-0056).
Ticio Quesada, E. (2018). The emergence of nominal expressions in Spanish-English early bilinguals: Economy and bilingual first language acquisition. John Benjamins.
Unsworth, S. (2013). Assessing the role of current and cumulative exposure in simultaneous bilingual acquisition: The case of Dutch gender. Bilingualism: Language and Cognition, 16(1), 86–110. https://doi.org/10.1017/S1366728912000284.
Van Heugten, M., & Johnson, E. (2011). Gender-marked determiners help Dutch learners’ word recognition when gender information itself does not. Journal of Child Language, 38(1), 87–100. https://doi.org/10.1017/S0305000909990146.
Willis, E. (2005). An initial examination of southwest Spanish vowels. Southwest Journal of Linguistics, 24 (1–2), 185–198.
In Tables 4 and 5, Cuza and Pérez-Tattam (2016, pp. 63–64) report number of children per age group, classified according to their overall accuracy level. Many children were in the zero-accuracy group (no target responses) and low accuracy range (i.e., between 1 to 4 accurate responses out of 10 possible answers).
A study of high frequency inanimate nouns, Clegg (2011), reports that they are roughly split three ways based on their endings: -a: 31% (including the suffix -ma); -o: 32%, and the rest: 36% (nouns ending in –e, consonants, and consonantal suffixes). Assuming -o/-a ending nouns are masculine and feminine, respectively, leads to over 99% classification accuracy. Based on this, it seems plausible that -o/-a may function as cues that facilitate the acquisition of gender. Note, however, that a 36% of nouns do not include such vowels. They represent 32% and 40.6% of the total of masculine and feminine high frequency inanimate nouns, respectively, a non-negligible portion. Finally, these percentages should be contextualized: animate nouns need to be factored in, as they were excluded given their complexity.
The vowels /a e o/ are by far the most frequent sounds in Spanish (Guirao & García Jurado, 1990; Moreno Sandoval et al., 2008; Arias Rodríguez, 2016).
Cross-language differences in the processes affecting vowel sequences at word boundaries might add another layer of noise in the acquisition of gender-related vowels (Colantoni et al., 2020).
Studies have also reported that phonetic context has an effect on formant realizations, but lexical frequency does not appear to condition the reduction of the vocalic space in vowel production in Spanish-English bilinguals (Lease, 2022).
Five files were selected for a reliability check, roughly 10% of the data, to compare the consensus between the two syntacticians coders with the judgement of a trained phonetician. Inter-rater agreement was at 82%. Of the 204 tokens extracted for the reliability verification, we found 37 classification mismatches, mostly occurring when the syntacticians perceived a target vowel, which the phonetician classified as underspecified (31/37). An anonymous reviewer asks, given this level of disagreement, why was the data not entirely coded by a phonetician. The distance between [naïve] perceptual analyses and phonetic analyses is precisely part of what is at stake. One of our methodological goals was to show that there is undetected variability not being factored in in the current literature, where, to the best of our understanding, coding and transcription is not done by trained phoneticians.
Eight narratives were discarded because of the amount of overlap in the conversation and/or background noise.
Values were initially extracted with a script, but all the values at midpoint were manually checked independently by two authors. We acknowledge that using midpoint values minimizes but does not control for the effect of the phonetic context. To a large extent, vowels were extracted from the same words, given similarities between narratives.
Two anonymous reviewers suggested that we add schwa as a fourth vowel to the algorithm. This assumes that we know that participants have a schwa vowel in their Spanish inventories, something which needed to have been empirically determined.
The moderate negative correlation between narrative length and TVIP is peculiar. Some children may speak little but have good receptive vocabularies. Productive and receptive vocabulary measures do not always align; some studies find that productive vocabulary but not receptive vocabulary correlate with MLU (see references in Pérez-Leroux et al., 2012, i.a.).
Some tokens, including those containing noun ellipsis (e.g., las__dos ‘the two’ (BC14, 10;04)), were not classifiable for all parameters.
Bare nouns were included for comparability to previous studies, and because canonical nouns could exhibit potential mismatch between final vowel and lexical specification.
We also extracted other gender-marked categories, including predicative adjectives and various pronoun types. While the number of tokens was insufficient for statistical analyses, overall accuracy was in the range of various NP s types (3Sg accusative clitics = 89%, n = 123; demonstrative pronouns = 92%, n = 13; predicative adjectives = 94%, n = 17; and strong pronouns = 99%, n = 151)).
Simulated datasets with 20 samples per cluster demonstrate it is possible to achieve sufficient power (Dalmaijer, 2022). Not all our participants meet this criterion. Results should be thus interpreted with caution.
Monolinguals also have some overlap in vocalic space (see, Figure 1 in Colantoni et al. 2020); we do not anticipate 100% classificatory accuracy from this procedure.
A reviewer suggests an additional possibility, that there are more cues to gender than to number. Finally, we must consider that the data needed to fully assess modular interactions is vowel perception, not vowel production.
For Mariscal (2008), articles are a robust cue and the first locus of differentiation of prenominal vowels.
This reasoning presupposes identity between vowel realization, phonological representation and vowel perception, which may not be the case. Perception studies are essential to advance this question.
If so, mass nouns, which frequently appear without determiners, should be harder to learn. This prediction is borne out by data presented in Cuza and Pérez-Tattam (2015, p. 62), where Spanish monolingual children show lower proportion of accuracy rates of determiner selection for opaque mass nouns than for opaque count nouns.
For some speakers, the production of unspecified vowels might be an opportunity to hide gender behind an indistinct vowel realization. Further studies could explore whether heritage speakers use a higher proportion of NP s whose only mark for gender is vowel quality, compared to monolingual speakers.
Our study does not contribute to the question of whether differences in participant backgrounds can explain variance in the two dimensions studied, phonetics and grammar, as our focus is on the association between the two dimensions, which cannot be explained by extra-linguistic factors.
