Introduction

Mastery of the stress pattern of American English is predicated on a speaker’s ability to assign stress to particular syllables using his/her linguistic knowledge of the phonological patterns of the language [1]. Stressed vowels are produced with greater intensity and duration than their unstressed counterparts, and are more perceptually salient to listeners [2,3]. In unstressed syllables, vowels are prevented from extending to the periphery of the oral cavity where most meaningful vowel contrasts occur. As a result, schwas are often regarded as the neutralization of vowel quality contrasts [2–4].

As unstressed vowels, schwas have been described as the product of a collapsed vowel space caused by reduced subglottal pressure, decreased muscular effort, and lack of coordinated gestures [3, 5–9]. These decreases in articulatory effort have the combined effect of localizing schwa in English’s underutilized central vowel space [2,5]. Where measurable, acoustic analyses generally corroborate phonological descriptions of schwa as the midpoint of the audible vowel space for F1(height) and F2 (backness) dimensions [10]. For example, prototypical English schwas spoken by an adult male are expected to contain formant bands at 500/1500/2500 Hz at the midpoint of periodicity [5]. While it is possible for peripheral vowels to undergo truncation or spectral-quality adjustment due to phonetic factors such as faster speaking rate [5, 6, 10–12] or dialectal variation [13], schwa maintains a more centralized position than all other English vowels in phonologically-similar environments [2, 10].

Schwa is also more vulnerable to coarticulatory effects caused by proximity to nearby segments [14–16]. For example, proximity to alveolar consonants has been shown to decrease schwa’s F1 formants and increase F2 formants, thereby resulting in higher vowels [2, 8]. Certain spectral qualities of schwa may also be more vulnerable to co-articulatory effects than others [8, 14]. This vulnerability has been attributed to under-specification of schwa for particular vowel quality contrasts, such as height or backness [8, 16–17]. Through gestural modelling of articulatory phonetics, Browman & Goldstein used nonsense word sequences to test whether schwa is unspecified for tongue position. Non-words containing schwa were mapped temporally and spatially as a series of gestural trajectories across syllable boundaries [8–9]. Ultimately, their gestural modelling indicated that schwa tends to gravitate toward a central vowel height despite proximity to anterior consonants—an indication that schwa is specified for at least the F1 acoustic dimension [8, 16–17]. Subsequent research on the under-specification of schwa has indicated that schwa may be unspecified only for the F2 (anterior/posterior) dimension, while centralized vowel height remains targeted across most speech [8, 15–17].

One phonological environment where more of schwa’s centralized vowel qualities are likely to be intact is in word-final position (e.g. “sofa” [sofə], “Asia” [eʒə]) [12]. Word-final schwas have previously been identified as having more stable (i.e. reliable) F1 and F2 formant values than their word-internal counterparts [17–18]. Moreover, word-final schwa has been shown to maintain reliable F1 and F2 formants when it becomes word-internal due to the affixation of an inflectional morpheme such as the possessive {s} [17–18]. For example, Flemming & Johnson (2007) obtained more centralized formant values for schwa in “Rosa’s” [rozəz] than in the similar construction “roses” [rozəz] where schwa is phonologically-conditioned to occur only when the preceding morpheme ends in a sibilant. Though the second syllable of “Rosa’s” is closed due to affixation of the possessive morpheme, F1 and F2 spectral values were comparable to those in word-final position (i.e. open syllables) [17]. While others have found a three-way distinction in native English speakers’ F1/F2 formant values of schwa in the plural morpheme, before the possessive {s}, and word-finally, it does appear that word-final schwa and word-final schwa + possessive {s} belong to more robust categories of centralized vowels, whereas other types of syllable-final schwa have lower F1 values (corresponding to higher vowels) [18]. The discovery that schwa need not be entirely word-final to display stable spectral qualities supports our empirical assumption that word-final schwa in running speech can maintain centralized formant values due to its specification for certain dimensions in word-final position.

In addition to spectral qualities, phonological vowel reduction (not caused by superficial changes in speaking rate, amplitude, etc.) in English is characterized by shorter durations than those of stressed vowels [2, 6, 19]. As a general rule, reduced vowel durations are shorter word-internally than word-finally. However, schwas that are phonologically-conditioned for optional weak syllable deletion (e.g. between the stressed syllable and another syllable anchored by schwa) possess shorter mean durations than word-internal schwas not phonologically-primed for deletion [20–22]. Across all positions, schwa’s status as an unstressed vowel is highlighted by the large ratio of stressed vowel duration to unstressed vowel duration that characterizes English [6, 23–24].

The aforementioned studies examined the spectral and temporal qualities associated with vowel reduction by monolingual English speakers across a variety of phonologically-conditioned environments. Our study extends this line of investigation to explore the stability of spectral and temporal qualities of reduced vowels produced by early and late Spanish-English bilinguals living in the United States. For the purposes of this study, we adopt Grosjean’s (2010) succinct definition of bilinguals as “people who use two or more languages in their everyday life” [25]. While Spanish has traditionally been classified as a syllable-timed language, or one where syllable durations are of roughly equal length, lexical stress does create meaningful contrast [26]. Monolingual Spanish speakers’ ability to differentiate stress extends beyond meaningful contrasts, however, as evidence has suggested monolingual Spanish speakers are quite capable of differentiating non-words based on stress contrasts [27].

In speech production, certain monolingual varieties of Spanish tend to demonstrate slightly shorter vowel durations in unstressed syllables [28, 29]. However, centralization of unstressed vowels has found more empirical support in the speech of Spanish-English bilinguals than for monolingual Spanish speakers [30]. Moreover, atonic Spanish vowels have been observed to have more truncated durations and centralized spectral qualities in the Spanish of Spanish-English bilinguals than one typically records in monolingual Spanish [28–30]. For example, Spanish-English bilinguals living in the American Southwest have been shown to display greater centralization of unstressed vowels (i.e. lower F1 values) compared to monolingual Spanish speakers, notably in the case of /a/ raising to [ə] in atonic syllables [31].

Prior studies have also observed that both early and late Spanish-English bilinguals are capable of identifying English stress [21, 32–35]. That is, Spanish-English bilinguals have demonstrated sensitivity to the distributional patterns of English noun and verb stress that enable listeners to quickly classify real words by lexical category [33, 34]. However, neither early nor late Spanish-English bilinguals have reliably demonstrated speech output that approximates the difference between full and reduced vowel temporal qualities in accordance with the duration ratios that have been proposed for native-like American English speech rhythm [6, 36].

The underlying strategies used to produce stressed and unstressed English vowels may differ between English monolinguals and early or late bilinguals [32–35]. For early bilinguals, it may be the case that producing stressed and unstressed vowels with completely native-like temporal qualities is not necessary for intelligibility of speech. Hence, in some environments “compromise values” for duration may not impair intelligibility [18, 23]. This is more likely true in bilingual environments where the concept of “foreign accent” is gradient [30, 35]. For late bilinguals, on the other hand, non-native-like vowel reduction may be the result of transferring native-language (L1) Spanish stress rules to English [6]. For instance, Spanish stress has been shown to be tied to vowel duration over spectral qualities. As Spanish does not have vowels that contrast in both length and spectral qualities, as English does with its tense/lax distinctions, elongation of Spanish vowels serves only to enhance perceptual prominence (i.e. stress) [6, 23].

In many cases, singular attention to vowel duration will result in correct stress identification because increased duration, more peripheral spectral qualities, and prosodic enhancements (e.g. greater intensity) often co-occur [32, 37]. However, assigning stress to vowels based on duration alone can be problematic in languages that possess crowded vowel inventories, such as English, because differences in duration can be subtle and prone to variability. Vowel duration can be affected by neighboring segments, syllabic structure, or hyper/hypo-articulatory characteristics of the utterance [21, 35, 37]. While instances where intelligibility is compromised may be rare, over-reliance on one dimension of vowel stress can inhibit generalization of English stress assignment rules to new words.

Late Spanish-English bilinguals have demonstrated more success at assigning stress to English words when the target words are either frequently occurring or phonologically-similar to frequently occurring words [33, 35]. Though English has many rules for assigning and manipulating stress based on morphophonological alternations, the distributional frequency of various English syllable structures appears to influence the likelihood that Spanish-English bilinguals will correctly assign stress [6, 33, 35]. For instance, late bilinguals quickly master the stress-shifting rule that disyllabic nouns typically stress the penult syllable whereas verbs place stress on the ult, and are therefore able to contrast between the noun “IMport” and the verb “imPORT” without difficulty [21, 34]. However, words that have atypical syllable structures or occur in sparse phonological neighborhoods typically receive less accurate stress placement by late bilinguals [33, 35]. Guion, Harada, and Clark (2004) observed the effect of lexical category on accurate stress placement was also less robust for late bilinguals than for early bilinguals [35]. For example, early Spanish-English bilinguals produced L2 English words containing either marked or unmarked syllable structures with nearly native-like stress placement, with the exception of observed tendencies to avoid word-final vowel lengthening [35]. Late bilinguals, on the other hand, deviated more abruptly from correct English stress placement when the phonotactic probability of the syllable structure decreased [33, 35].

Issues concerning late bilinguals’ ability to accurately assign stress are significant in light of prior claims that accurate stress placement chronologically precedes accurate vowel reduction in speech production [21–23]. However, stress assignment is just one reason that spectral quality and temporal deviations from monolingual English speech can occur. Another potential contributive factor is lack of sensitivity to phonetic or prosodic features of the L2 that are not contrastive in the bilingual’s native language [21, 32]. In Spanish, there is no phonemic contrast for vowel duration [30]. English, on the other hand, does contrast vowels by duration under the tense/lax vowel distinction [38]. Therefore, maintaining categorical separation of target vowel durations is important for maintaining lexical contrast [38]. In general, Spanish-English bilinguals produce longer vowel durations in Spanish than their monolingual counterparts as well as shorter ratios of stressed-unstressed syllable durations than monolingual English speakers [28]. In addition to these sorts of “compromise values,” vowel durations in atonic syllables vary more widely for Spanish-English bilinguals than English monolinguals, which may indicate prosodic targetlessness that persists after segmental mismatches have been resolved [28].

Therefore, mismatches between the role of vowel duration in the L1 and L2 of Spanish-English bilinguals may contribute to longer unstressed vowels as well as smaller ratios for the duration of stressed-to-unstressed vowels than monolingual speakers typically produce.

As language learning is refined over the lifetime [39], early bilinguals’ prolonged experience perceiving and producing L2 English may result in more native-like L2 phonetic output [40]. Over time, L2 speakers may perceptually attune to differences between their articulatory output and phonetic properties of the L2. Heightened sensitivity to subtle differences between a bilingual’s L1 and L2 phonologies may stem from more exposure to native L2 speakers, more practice speaking in his/her L2, or explicit pronunciation feedback [41, 42]. This process is outlined by Flege’s Speech Learning Model and its subsequent revisions [39–41, 43–44]. Early in the L2 acquisition process, it is anticipated that speakers whose L1 contains neither central vowels nor mandatory stress assignment will substitute peripheral vowels in place of schwa [21, 39–41]. This tendency is magnified in any sort of reading task, orthographic representations of unstressed vowels correspond to peripheral vowels [44]. Refinement of unstressed vowel category boundaries (i.e. schwas) occur as a result of increased exposure to L2 speakers, more practice using the L2, or through explicit instruction such as contrastive analysis of L1/L2 phonetic properties [40–42]. Of these factors, none has attracted as much attention, or been contested as bitterly, as age of L2 learning [44–51].

While the presence of a foreign accent is likely the result of several factors working in unison, there is evidence that ‘age of L2 learning’ is one of the most reliable contributors to native-like L2 phonetic production [47–49]. A half-century ago, Lenneberg argued that there is a “critical period” which exists for language learning [50]. The idea that one’s opportunity for native-like L2 phonological acquisition closes after a certain biological window continued to gain support over subsequent decades [43, 45, 47, 51]. Patkowski described the critical period for L2 phonological acquisition as “… a period, ending around the time of puberty, during which it is possible, but not inevitable, for L2 speakers to acquire as an end-product of a naturalistic L2 acquisition process full native-like fluency in the phonological system of a second language, and after which such a possibility does not exist anymore” (see p. 206 of [47]). Studies have also observed a gradual decline in native-like L2 phonetic output as age of L2 learning increases [36, 38, 46, 52]. However, there does not appear to be either a biological impetus for decreased phonetic accuracy or an abrupt closing of the window where pre- and post-critical period boundaries can be delineated [10, 39, 48, 49].

The question of why L2 pronunciation appears to be hindered by advanced age of L2 acquisition when other language areas (such as vocabulary and syntax) do not has been offered mainly perceptually-based explanations [32, 54]. People who began learning their second language in adulthood may be disadvantaged when learning new phonetic categories because their L1 linguistic experience has accustomed them to grouping perceptually similar sounds as manifestations of the same phoneme [39, 40]. To produce native-like L2 sounds, late bilinguals must reshape their categorical boundaries based on L1 phonology to accommodate incoming L2 phones [40, 43, 54, 55]. To date, few people have suggested that even early bilinguals (who continue to use both languages) speak each language free from any cross-linguistic interference [10, 56, 57]. Studies that seem to contradict this statement have highlighted cases where bilinguals “pass” as native speakers; however, these case studies typically involved speakers of typologically-related languages (e.g. two Germanic languages) or accent judges who were unfamiliar with the phonetic features of certain “foreign accents” [56–58].

In recent years there has been resistance to the idea that late bilinguals face inevitable limitations on their capacity to achieve native-like L2 pronunciation [58–60]. A major criticism has been that, to date, there is not neurophysiological evidence of a causal relationship between cerebral maturation and inability to perceive and produce L2 sounds with native-like pronunciation [60–64]. Suggestions regarding loss of “neuroplasticity” in the brain [47] suggest that certain connective pathways (i.e. “white matter”) between phonological processing and motor speech planning regions do not form easily after the brain has reached maturity. Exactly how language learning differs from other types of learning in the brain has not been established to date [48]. For instance, studies of external neurological activity (electroencephalography) during classroom-style language instruction have indicated changes in electrical patterns in brain activity during the earliest stages of L2 acquisition [62]. In addition, some patients with aphasia (a neurological disorder characterized by difficulties using language) have experienced functional reorganization of connective pathways in their brains to circumvent damaged pathways following a cerebrovascular accident [62–64]. This discovery at the very least suggests that neuroplasticity is not controlled by exclusively biological factors [65]. In summary, physiological evidence of neuroplasticity after puberty indicates two important points regarding age and phonological learning: First, claims that a biological window that closes upon the opportunity to attain native-like pronunciation have yet to be concretely substantiated at either structural or functional levels of neural organization [61]. Second, we have learned from post-stroke patients that the brain is capable of reorganizing large-scale systems, including those for processing and producing language [63].

If reduced ability to accurately acquire L2 phonologies after puberty is not entirely biologically-driven, then the most likely explanation is that it is caused by either cognitive or experiential factors, or a combination of two. In recent years, cognitive models of working memory and information processing have been modified to incorporate the role of language and sub-vocal rehearsal in maintaining auditory information [65–66]. If we adopt the assumption that language acquisition is a form of information processing, then the native language necessarily creates a filter through which new languages are processed and organized [65, 67]. Cognitive abilities that can impact native-like L2 speech production include executive functions as pattern recognition, selective attention, and task switching [65, 67, 68]. A review of the role individual cognitive abilities play in L2 vowel production is beyond the scope of the current article; however, Abutalebi & Green (2007) provide a comprehensive review of the relationship between bilingualism, executive control, and functional imaging of language tasks in the brain [68].

Though age of L2 learning may be a comparatively reliable predictor of the likelihood of acquiring a native-like L2 phonology, experiential factors also play a role in predicting the outcome. Experiential factors include amount of L1 use [25, 57, 69–71], formal education in one or both languages [20, 72], and the native-speaker status of the L2 community [69]. These factors may overlap to varying degrees with nonstructural features such as linguistic attitudes [73], ethnic identity of the L2 speaker [74], and motivation to achieve native-like pronunciation in the L2 [75]. Of the experiential factors listed, one that has been shown to negatively predict native-like L2 speech production is amount of L1 use [57]. Using one’s native language more often is believed to strengthen its level of activation, thereby creating more opportunities for phonetic interference than one predicts in bilinguals who seldom use their native language [25]. For example, Italian-English bilinguals living in Canada who often speak Italian have been judged by monolingual English speakers to possess stronger foreign accents than their age-matched peers who speak Italian less often [41]. Children who spend a greater proportion of each day speaking in their L2 typically reflect less accented L2 pronunciation [71], as do adults who work in L2-immersive environments [57]. Amount of L1 use often overlaps with age of L2 learning because early bilinguals often have spent much of their day being educated in their L2 and interacting with L2-speaking peers—two avenues that have been suggested to motivate more native-like L2 speech output [53, 73–76].

Late bilinguals, on the other hand, have received at least some of their education in their native language. If this education included literacy training, as is likely the case, then late bilinguals likely possess some degree of L1 orthographic interference when reading L2 words. Only if a late bilingual is immersed in L2 writing and media for an extended period would we expect orthographic effects to diminish. As previously mentioned, interacting with peers is another source of motivation for improving L2 pronunciation. Late bilinguals may be less motivated to develop native-like L2 speech for a variety of reasons. For instance, native-like L2 pronunciation may conflict with a bilingual’s ethnic identity [74]. Late bilinguals are also more likely to have established relationships with people who share their native language, thereby dampening their motivation to approximate the speech of monolinguals who speak their L2 [75, 76]. Differences notwithstanding, even late bilinguals may alter their phonetic output as phonemic category boundaries change with experience [32, 39, 40]. Furthermore, “gestural drift” can occur in both early and late bilinguals when they move between monolingual L1 and monolingual L2 environments [70].

Various linguistic history questionnaires have probed bilinguals’ attitudes toward their languages using a variety of objective and subjective questions. Some questionnaires focus heavily on the amount of time engaging in L1 vs. L2 use and draw distinction between time engaging in expressive versus receptive language use [77–79]. Others have aimed to capture subjective measures, such as the level of importance the individual assigns to a language and the emotional attachment he/she feels toward one language versus the other [80]. Taken together, these measures contribute to the factor “bilingual dominance” that can predict how native-like we expect L1 and L2 speech output to be [80]. Bilingual dominance scores that also include subjective measures of linguistic attitudes can be instrumental in explaining why phonetic output varies within communities that are relatively homogenous according to many sociolinguistic assessments.

The current study examined vowel reduction in word-final position among groups of early and late Spanish-English bilinguals, as well as English monolinguals. To create a balanced stimuli set that accounts for possible co-articulatory effects due to the preceding consonant, many of our target words are cognates for Spanish and English. Cognates have previously been identified as more vulnerable to phonetic interference compared to non-cognates [80]. However, the direction of transfer is not unidirectional L1→L2 as one would predict if L1 literacy effects were the underlying cause. Amengual (2012) observed that Spanish-English bilinguals produce longer voice onset times (VOTs) for cognates than for non-cognates [80]. However, it has been noted that VOTs for /t/ are particularly unstable in Spanish-English bilinguals [56], and this segment showed the most cross-linguistic interference among voiceless stops [80]. Spanish-English bilinguals have also demonstrated clearer categorical separation of VOT values for cognates pronounced in Spanish and in English when stimuli were presented entirely in English or Spanish, compared to code-switching constructions [56, 81]. These findings suggest that placing a bilingual in a monolingual language mode can suppress cross-linguistic phonetic interference for bilinguals (at least those who are highly literate in the language being tested) even if the target words have cognate status.

All of our participants were residents of Miami, Florida who were enrolled as students at Florida International University, where courses are taught in English. Extensive experience with L2 reading/writing as well as monolingual L2 English language mode in the classroom (due to English-only instruction) positions both early and late bilinguals to produce English sentences with diminished L1 orthographic transfer. Our study further examined how well three linguistic history factors predicted native-like L2 vowel reduction. These factors were: Age of L2 learning, amount of L1 use, and self-reported bilingual dominance scores. To date, there are no studies establishing early and late bilinguals’ average schwa durations and vowel qualities in word-final position, where centralization of the unstressed vowel is more likely to be maintained than word-internal atonic syllables [2, 17]. Detailing vowel reduction patterns in these three groups of Miami-based speakers will also contribute valuable phonetic data to ongoing efforts by sociolinguists to categorize an emerging “Miami accent” [82].

We predicted that late bilinguals would vary their word-final schwa vowel qualities more than monolinguals and early bilinguals due to orthographic interference effects during our reading task. We also expected more evidence of an [a]/[ə] merger in late bilinguals’ speech from orthographic effects, while early bilinguals were predicted to produce compromise vowel qualities half-way between monolingual English speakers’ and late bilinguals’ vowel qualities due to having both languages heavily activated in the heavily bilingual city of Miami. In addition, we predicted longer word-final schwa durations for late bilinguals because they are expected to display smaller ratios for the duration of stressed and unstressed vowels compared to early bilinguals and monolinguals. Early bilinguals were predicted to produce longer vowel durations than monolingual English speakers, but shorter than late bilinguals, because interactions within their L1/L2 system may result in compromise values for all centralized vowels that are longer than [ə], but shorter than [a]. We did not predict that orthographic interference would affect our early bilinguals because most of them were not educated in Spanish beyond what any high school student receives, regardless of language background.

Before moving on, it is important to acknowledge that Miami-based English monolinguals may not produce the same vowel qualities and reduced vowel durations that one expects from general American English. For instance, English monolinguals living in Miami have recently been shown to deviate prosodically from general American English. Early characterizations of Miami-accented monolingual English include reduced duration ratios between stressed and unstressed syllables (giving the impression of a more syllable-timed English) as well as subtler variation in pitch for stressed syllables [83, 84]. However, we maintain that Miami English should be considered the “target English,” despite access to general American English in the media, because it is the source of L2 input and spoken across social classes without stigma.

We further predicted that within each bilingual language group, more native-like reduced vowel production could be predicted by the two linguistic factors “amount of L1 use” and “bilingual dominance score.” We were aiming to determine if certain factors predicted either vowel qualities, vowel duration, or both. Additionally, we sought to discover whether different factors predicted native-like schwa production in early bilinguals than in late bilinguals. Specifically, we predicted that speakers whose scores indicated more motivation to retain and improve their L2 would produce more native-like schwas than those who maintained a stronger social and emotional connection to their native language. We ranked the explained variance of our nonstructural variables differently for duration than for vowel qualities, which fulfilled our aim of exploring how nonstructural variables predict phonetic variation in bilingual speech.

Methods

This experiment was designed to test early and late bilinguals’ production of schwa in word-final position elicited from a reading task. It was approved by the Florida International University Review Board as IRB-011212-02-TR “Variation in the Production of Reduced Vowels by Monolingual and Bilingual Populations” as one of several proposed studies examining segmental variation. We aimed to determine if the two bilingual groups systematically produce different vowel durations and/or spectral qualities from Miami’s monolingual English speakers [17, 20]. We also examined three potential predictors of native-like L2 phonetic output including age of L2 learning, amount of L1 use, and bilingual dominance scores.

Stimuli

To examine schwa duration and vowel qualities between our language groups, an English-language reading task was created. Fifteen of the most frequently-occurring disyllabic words containing word-final schwa were selected as target words for the experiment [96]. The words we selected had also been previously rated as “highly familiar” by monolingual English-speaking students [97]. Target words were divided into three categories based on the point of articulation of the preceding consonant. These categories included: post-labial schwa, post-coronal schwa, and post-dorsal schwa. It is worth noting that a discrepancy exists between the point of articulation for English and Spanish stops /t,d/. English /t,d/ are typically located at the alveolar ridge (with dental and interdental allophones in certain phonological environments) [98, 99], whereas Spanish /t,d/ are typically produced at the back of the superior incisors [100]. For the purposes of this study, we emphasize that the three categories of preceding consonant that we have identified correspond to labial, anterior, and posterior sounds in both languages.

Each target word was embedded into the initial determiner phrase of each sentence in order to maintain similar prosodic rhythms and levels of focus across sentences. In addition, each target word was followed by an alveolar stop to avoid regressive co-articulatory effects brought on by anticipation of the following consonant. In addition to the target sentences, filler sentences of varying length and intonation patterns were added to the stimuli to prevent speaker boredom. These sentences did not contain any word-final schwas. Examples of filler sentences are “We’ll probably go shopping tomorrow” and “Don’t tell mom we’re leaving on Sunday.” Sentences were loaded into a PowerPoint presentation where one only sentence appeared per slide. Sentences are listed in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Sentences containing word-final schwa targets.

The “post-labial” category includes schwas that follow [m], [b], and [v]. “Post-coronal schwas” follow [s], [ʒ], [z], [ʤ] and [ʃ]. “Post-dorsal” schwas follow [k] and [g].

https://doi.org/10.1371/journal.pone.0175226.t001

Results

Spectral qualities

Prior to synthesizing the results of the models, descriptive statistics for the mean formant values and duration of schwa following labial, coronal, and dorsal consonants are presented in Table 2. Values reflect the estimated marginal means taken from a one-way multivariate analysis of variance (MANOVA) model, rather than raw means [28].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Duration and vowel quality means^a for word-final schwa produced by English monolingual, early Spanish-English bilinguals, and late Spanish-English bilinguals.

https://doi.org/10.1371/journal.pone.0175226.t002

Multivariate results indicated significant main effects for both independent variables “language group” (F (4, 1776) = 33.56, p <.001) and “preceding consonant” (F (6, 1776) = 10.83, p <.001). The interaction of language group * preceding consonant was not significant, p = .235. Results further indicated that a sizeable portion of the overall variance in vowel qualities that was present for variables F1, F2, and F2-F1 was explained by the predictors language group and preceding consonant (Adjusted R² = .400). Language group and preceding consonant were also significant predictors for each of the spectral variables when considered individually, p <.001. Within groups, results indicated that schwas were systematically influenced by the phonetic features of the preceding consonant for at least one formant. Of the individual spectral variables, language group explained the highest percentage of variance for F1 (η² = .11) with considerably less variance explained for F2 or F2-F1 (η² = .03 for both DVs). In general, LBs produced higher F1 values than either EMs or EBs, indicating a lower vowel position. EMs demonstrated the most homogenous spectral qualities for word-final schwa, while EBs differed markedly in their F2 values (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Raw spectral values of schwa for all language groups.

The x-axis shows F2 formants that correspond to a vowel’s position along the anterior-posterior dimension of the oral cavity. The y-axis shows F1 formants that correspond to vowel height.

https://doi.org/10.1371/journal.pone.0175226.g001

Pairwise comparisons further revealed that not every group differed significantly from the other two along every spectral measure. For example, F2-F1 values (overall vowel position in the oral cavity) indicated that EMs produce schwa at points of articulation that differ from either EBs (p <.001) or LBs (p = .019), whereas the difference between EBs and LBs failed to reach significance, p = .544. F1 differences between EMs and EBs were also less robust, although significant, p = .024 [Table 3].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Pairwise comparisons of F2-F1 formant values for English monolingual, early Spanish-English bilinguals, and late Spanish-English bilinguals.

https://doi.org/10.1371/journal.pone.0175226.t003

Duration

For duration, there was a significant main effect of language group (F(2, 891) = 82.46, p <.001). However, this effect was not equal across all groups. Pairwise comparisons revealed that LBs had significant duration differences compared to EMs (p <.001) and EBs (p <.001). EBs and EMs did not differ significantly with regard to duration, p = .401.

A second significant main effect of preceding consonant on schwa duration (F(2,891) = 25.45, p <.001) was found. Specifically, a lack of significant difference between EM and EB schwa production was affected by the occurrence of certain preceding consonants. In addition, the interaction between language group * preceding consonant was significant, p = .029. For EMs, mean schwa values were nearly identical when following either of the anterior consonants (i.e. labial and coronal), as summarized in Table 2. EBs and LBs, on the other hand, had mean schwa durations that categorically differed according to the preceding consonant. Post-labial schwa had the most separation between the groups.

In general, EMs and EBs produced schwa in post-coronal and post-dorsal position with nearly identical durations. The pattern can be summarized as follows: EMs produced post-labial and post-coronal schwas with nearly equal durations, while EBs produced post-labial and post-dorsal schwa with nearly equal durations. LBs, on the other hand, exhibited wide differences between the duration of schwa in post-labial, post-dorsal, and post-coronal positions. Across all preceding consonant categories, LBs produced schwa following all three places of articulation with longer durations than those observed for either EMs or EBs. As Fig 2 illustrates, both bilingual groups adhere to the following durational hierarchy for schwa: post-coronal < post-labial < post dorsal. English monolinguals do not contradict this pattern; rather, their overall shorter schwa durations have less range for categorical separation of durations following anterior consonants.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Mean durations for schwa by preceding consonant.

The X-axis identifies the place of articulation of each category of consonant that precedes schwa. The Y-axis shows duration in seconds. Each line represents a different language group.

https://doi.org/10.1371/journal.pone.0175226.g002

Nonstructural measures

As mentioned above, we explored the role three nonstructural variables had in predicting native-like vowel reduction. These variables were age of L2 learning, bilingual dominance score (BDS), and amount of L1 use. While it is commonly accepted that age of L2 learning is a solid predictor of native-like L2 production (although not a guarantee), the other role played by bilingual dominance and amount of L1 use is less clear. We performed two multiple multivariate linear step-wise regressions to explore whether these three nonstructural measures explained the variance in schwa duration for EB and LB groups. One multiple regression was performed for duration and the other to predict the variance for spectral qualities of schwa.

Nonstructural predictors of schwa duration.

Taken together, the three predictors (AoA, amout of L1 use, and bilingual dominance score) explained a sizeable portion of the variance for schwa duration (R² = .54, p <.001). Of the three predictors, age of L2 learning was the greatest predictor of word-final schwa duration (R² = .51, F(1, 39) = 39.27, p <.001). Amount of L1 use had a marginally significant main effect F(1, 39) = 3.26, p = .08, R² = .08. Bilingual dominance score was not a significant predictor of schwa duration. Essentially, as age of L2 learning increased, word-final schwa duration also increased (Fig 3). However, the individual data points in Fig 3 also show considerable overlap between certain EB and LB speakers for word-final schwa duration.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Age of second language learning predicts word-final schwa duration.

The x-axis shows the age when second language learning began (in years). The y-axis shows the duration of word-final schwa in seconds.

https://doi.org/10.1371/journal.pone.0175226.g003

Bilingual dominance scores were the other significant predictor of the variance in word-final schwa durations (adjusted R² = .10, F (1,38) = 5.25, p = .028). Recall that bilingual dominance scores (BDS) were calculated on a scale of -30 to 30, where -30 is an L1 heritage language monolingual (in this case Spanish), 30 is a dominant language monolingual (English), and zero is a completely balanced bilingual [79]. Therefore, any subject who has a negative BDS prefers to speak, or is dominant in, L1 Spanish. Any subject who has a positive BDS prefers to speak English or is dominant in that language. It should be noted that separating subjective linguistic preference from empirically-validated linguistic dominance is beyond the scope of the current study. Results from the regression indicate a negative relationship between duration and L2 dominance—as bilingual dominance shifts toward L1, mean schwa durations increase. EBs demonstrated more within-group homogeneity than LBs for this measure (Fig 4). Bilingual dominance scores were highly correlated with both amount of L1 use (r (38) = -.54, p <.001, and age of L2 learning r(38) = -.55, p <.001). Age of L2 learning and amount of L1 use were less correlated r(38) = .24, p = .064. For word-final schwa duration, amount of L1 use was not a significant predictor, p = .079.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Early age of L2 learning predicts higher bilingual dominance scores and lower word-final schwa durations.

The x-axis shows the range of reported bilingual dominance scores, where negative scores indicate L1 dominance and positive scores indicate L2 dominance. The y-axis shows word-final schwa durations in seconds.

https://doi.org/10.1371/journal.pone.0175226.g004

Nonstructural predictors of spectral qualities.

Taken together, our nonstructural predictors significantly predicted the spectral qualities of schwa, defined as the variable F2-F1 F(3,39) = 6.23, p = .002, R² = .34. Age of L2 learning was again the single greatest factor for explaining variance in F2-F1 spectral qualities for word-final schwa (R² = .34, F (1,38) = 19.65, p <.001). BDS did not quite explain a significant portion of the variance, F(1, 39) = 3.78, p = .06, R² = .09. Amount of L1 was again not a significant predictor, p = .378. A closer inspection of the spectral qualities reveals that for the dependent variable F1 (height), age of L2 learning also explained the most significant percentage of the variance (R² = .37, F(1, 38) = 23.65, p <.001). Bilingual dominance scores also explained a significant portion of the variance for bilinguals’ vowel height (R² = .12, F(1,38) = 6.03, p = .019. As with vowel duration and vowel position (F2-F1), amount of L1 use was highly correlated with age of L2 learning but not a significant predictor on its own.

Likely driven by the heterogeneity observed in F2 formants (anterior-posterior position), all of our predictors had less explanatory power than for duration, F2-F1, and F1 values. Age of L2 learning was the only significant variable for explaining variance along this dimension (R² = .07, F (1,38) = 5.71, p = .022). Neither BDS (p = .306) nor amount of L1 use (p = .037) were significant predictors of native-like vowel qualities. Lack of good fit between our nonstructural predictors and F2 values is consistent with the heterogeneous spread of F2 values by EBs observed in Fig 1.

Discussion

Supporting information

Acknowledgments

Special thanks to Dr. Paulette Johnson for her help in preparing the statistical analysis for this manuscript.