Variability, accuracy, and cross-linguistic transfer in bilingual children speaking Jamaican Creole and English

Abstract

Due to the lack of normative data about bilingual speech development and limited availability of diagnostic tools optimized for this population, bilingual children under consideration for speech-language services are at an elevated risk of misdiagnosis. In the absence of validated assessment tools, speech-language pathologists may use measures of accuracy and variability of speech production to diagnose suspected speech sound disorders in bilingual children. Research in general motor development suggests that variability and accuracy may trade off in the course of maturation, whereby movement variability spikes before the transition to a more mature stage of motor control. Such variability-accuracy tradeoffs have been described in monolingual speech development but are understudied in bilingual populations, where cross-linguistic transfer occurs. This study aimed to examine variability, accuracy, and cross-linguistic transfer in the speech of 20 bilingual children speaking Jamaican Creole and English. We hypothesized that children who showed higher accuracy in their productions would also exhibit more variable speech, indicating a variability-accuracy tradeoff. The Word Inconsistency Assessment from the Diagnostic Evaluation of Articulation and Phonology was administered to measure accuracy and variability in the English context, where misdiagnosis is likely to occur. Contrary to hypothesis, we observed that individuals with higher accuracy tended to be less variable in their productions. Future research should examine longitudinal trajectories of accuracy and variability and consider a more culturally-appropriate definition of ‘accuracy’ in documenting bilingual speech sound development.

Introduction

Speech sound disorders (SSD) are defined as significant deviations from typical speech sound development. SSD have been shown to have lasting effects on academic, social, and behavioral outcomes in adolescents and adults (Felsenfeld, Broen, & McGue, 1994). Thus, prompt and effective treatment of SSD represents a high priority for speech-language pathologists (SLPs). However, bilingual children under consideration for speech therapy services for SSD face additional obstacles; an overwhelming majority of SLPs in the USA do not have the resources and/or training to provide adequate speech therapy services to bilingual individuals. In fact, only 8% of SLPs identify themselves as bilingual service providers (ASHA, 2021).

When assessing for the presence of SSD in children, SLPs often consider the variability or inconsistency of speech sound production. Variability in speech production can be measured in multiple ways. Acoustic or articulatory measures can be used to quantify fine-grained phonetic fluctuations across productions of the same word or sound. Phonetic transcription using the International Phonetic Alphabet (IPA) can be used to characterize productions as consistent or inconsistent at the lexical level (i.e. across repeated productions of the same word) or at the segmental level (i.e. across different phonetic contexts, such as in onset versus coda position).

Here we focus on a commonly used clinical measure, token-to-token variability¹ of repeated productions of the same word (Holm, Crosbie, & Dodd, 2007) based on differences in broad transcription. Higher rates of token-to-token variability have been observed in different clinical populations such as children with SSD (sometimes attributed to a specific subtype termed inconsistent phonological disorder; Dodd, 2014) and Childhood Apraxia of Speech (Iuzzini-Seigel et al., 2017). However, there is controversy as to whether token-to-token variability always indicates disordered speech, as previous studies of typically developing monolingual talkers have reported conflicting results. For example, Holm et al. (2007, 2022) found highly consistent speech in typically developing children aged 3;0 to 6;11, positing that a high degree of variability may be suggestive of SSD. In contrast, other researchers observed substantial token-to-token variability in typically developing monolingual children in the same age range and suggested that such variability should not be treated as indicative of SSD (Sosa, 2015; Macrae & Sosa, 2015). The discrepancy between the studies may relate to the method of transcription (online versus offline). In the current study, we use transcription-based token-to-token variability to measure bilingual speech production for two reasons. First, this approach has been tested in previous studies of bilingual populations (Abu El Adas et al., 2021; Holm et al., 1999; Preston & Seki, 2011). Second, variability measures have the advantage that the child is compared against themselves across repeated productions, rather than drawing comparisons with an adult target, where issues of cross-language transfer come into play (Abu El Adas et al., 2021).

Predictors and correlates of variability in speech production

Various factors such as age, expressive vocabulary, and exposure to multiple languages have been shown to influence variability in speech production. Previous studies have also reported conflicting results regarding the relationship between age and variability in the speech of typically developing children. Holm et al. (2007) reported an association between age and consistency of productions, in that younger participants produced significantly more variable speech. However, Macrae and Sosa (2015) found that age was not a significant predictor of token-to-token variability in a model that controlled for expressive vocabulary. The observation of higher rates of consistency in older children may be due to the expansion of vocabulary that occurs with age. A possible mechanism for this association can be explained by the Lexical Restructuring Model proposed by Metsala & Walley (1998). This model states that as children’s vocabularies grow, their lexical representations change to include increasingly fine-grained phonetic detail to differentiate between similar lexical items. This attention to segmental-level details can be associated with greater consistency of speech productions (Macrae & Sosa, 2015). It has been suggested that children with more fine-grained, segmental representations of words might also be expected to produce more accurate speech (Walley et al., 2003).

Exposure to multiple languages may also impact production variability. Given that bilingual children receive more variable input across two languages, they may show even greater variability in their productions than monolingual children (Byers-Heinlein & Fennell, 2014). However, some studies have reported opposing results; Abu El Adas et al. (2021) found similar rates of production variability between monolingual and bilingual children aged 3;4–5;1. This finding aligns with previous findings by Preston & Seki (2011) that suggested that token-to-token variability may be useful in diagnosing bilingual children with SSD. Because the nature of phonological and lexical overlap in a given language pairing may impact rates of production variability, more research is needed to further investigate how interactions between language pairings and contexts can affect production variability in bilingual speakers.

Measures of accuracy in bilingual speakers

In addition to measures of variability, clinicians often use standardized tests of articulation and phonology to assess how accurately a child produces speech compared to their typically developing peers (Davis, 2007; Fabiano-Smith, 2019). Several studies provide evidence for variant speech development patterns in bilingual children that may prevent them from being fairly compared relative to norms for monolingual speakers (Gildersleeve-Neumann, Kester, Davis, & Peña, 2008; Fabiano-Smith & Hoffman, 2018; Fabiano-Smith et al., 2021). For example, consonant accuracy can present as delayed or accelerated in bilingual speakers depending on factors such as whether the phoneme in question is shared between the two languages (Fabiano-Smith & Goldstein, 2010). Using normative data derived from monolingual speakers to assess bilingual children puts bilingual children at risk for misdiagnosis (Castilla-Earls et al., 2020) which may include underdiagnosis, in which a communication disorder is not properly identified, or overdiagnosis, in which typically developing speech is mistaken for disordered speech. Nevertheless, in a national survey that evaluated SLPs’ assessment practices, a majority of clinicians reported using informal or English-only assessments when evaluating bilingual individuals (Skahan et al., 2007). In part, this may be because few tests have been formally established in languages other than English and Spanish (Yavas & Goldstein, 1998; Wright Karem & Washington, 2021). There is a clear need for assessments that evaluate language capabilities for bilingual speakers in both of their languages.

Due to the limited availability of standardized diagnostic tools for bilingual speakers, criterion-referenced assessments such as the Percentage of Consonants Correct-Revised (PCC-R) metric are widely used (Gildersleeve-Neumann et al., 2008; Fabiano-Smith & Hoffman, 2018; León et al., 2021, 2022; Holm et al., 2022). The original measure, Percentage of Consonants Correct (PCC), reflects the number of accurately articulated consonant sounds in a conversational sample divided by the total number of consonants in the adult target forms (Shriberg et al., 1984). The updated PCC-R metric takes substitution and omission errors into account and excludes distortion errors (Shriberg et al., 1997). Fabiano-Smith and Hoffman (2018) found that PCC-R provides a good indicator of phonological ability in bilingual children aged 5;0 and above. In addition to segment-level measures such as PCC-R, whole word analyses have been used to measure accuracy in children as they tend to be word-oriented in their acquistion of language (Ingram & Ingram, 2001). Whole word accuracy assesses a child’s ability to produce entire words relative to an adult target (Ingram & Ingram, 2001). Small-scale studies have suggested that whole word accuracy could be a valid indicator of phonological abilities in bilingual children with SSD (Burrows & Goldstein, 2010).

Despite these positive findings, there are limitations of accuracy-based measures for the diagnosis of SSD in bilingual children. The results in Fabiano-Smith and Hoffman (2018) suggested that PCC-R may be less sensitive with younger bilingual children, creating a higher risk of misdiagnosis. Recent studies on bilingual populations suggest that the likelihood of misdiagnosis can be lowered by using accuracy measures from both languages (Fabiano-Smith et al., 2021; León et al., 2022). For example, Fabiano-Smith et al. (2021) found that Spanish-English bilingual children ages 3–6 were more accurately diagnosed when PCC-R measures were derived from both languages.

The influence of transfer on accuracy for bilingual speakers

It is important to consider how bilingual input might influence children’s performance on accuracy measures (Washington et al., 2017). Previous research has examined how similarities between two languages may affect a speaker’s representation and production of speech, particularly in terms of shared phonemes across languages (i.e. sounds occurring in both language inventories) versus unshared phonemes (i.e. sounds occurring only in one language). Fabiano-Smith & Goldstein (2010) investigated interactions between languages in the speech of Spanish-English bilingual individuals and found that instances of transfer – defined as the incorporation of properties from one language to the other that may occur bi-directionally (Paradis & Genesee, 1996) – systematically reflected the phonetic rules of their other language. Because transfer is a typical feature of bilingual speech development, it poses a challenge when attempting to assess the accuracy of speech production in a given language for diagnostic purposes (Genesee et al., 1995; Yow et al., 2018). A culturally-appropriate approach avoids the automatic counting of non-target productions as ‘errors’ and qualitatively examines substitution, deletion, or insertion patterns in relation to the possibility of interactions between two language systems (Wright Karem & Washington, 2021).

Variability-accuracy tradeoffs in motor development

As discussed previously, some research on variability in speech development associate a high level of variability with underdeveloped or disordered speech. However, evidence from other contexts suggests that variability may play a beneficial role in learning (Sternad, 2018). In the context of general motor development, studies have shown evidence of variability-accuracy tradeoffs, where a spike in variability occurs before a child transitions to a new stage of motor development (Thelen et al., 1993; Gliga, 2018). These variability spikes become more common as the availability of information increases (Gliga, 2018). For example, infants who are exposed to objects with novel characteristics (e.g. temperature or texture) will show an increase in variability in object-directed movements, displaying an exploratory approach to learning in a contextually appropriate environment (Gliga, 2018). Such findings suggest that variability in movement may be beneficial or even necessary for children to achieve mature motor control.

While these findings come from the nonspeech motor learning literature, they may also have relevance to the speech domain. Speech production is unique in that it not only involves the motor act of controlling the articulators, but also engages abstract layers of representation such as phonemes and syllables. However, we subscribe to the theoretical notion that articulation and phonology are inextricable constructs (Fey, 1992), and that factors relevant for articulatory motor control may also play an active role in shaping phonology (Namasivayam et al., 2020; McAllister Byun et al., 2016). This is possible when the phonological grammar is constructed over the course of the speaker’s experience (McAllister Byun & Tessier, 2016). In this theoretical framing, the behavior of the general motor control system has the potential to translate directly to influences on phonology.

In the speech domain, a study by Goffman et al. (2002) measured changes in a child’s articulatory movements before and after receiving a cochlear implant. The researchers observed an increase in variability of articulatory movements for six months post-implantation before the child reached a degree of consistency comparable to control participants. He was able to maintain a high level of accuracy in his speech productions, indicating that the variable movements did not have an adverse effect on speech output precision. This increase in variability can most likely be attributed to the new auditory stimuli the child received and offers evidence for the benefits of using exploratory learning strategies when exposed to new information. This also suggests that variable productions of target words across repeated elicitations may not be indicative of a disorder, but rather a sign of ongoing speech development.

In typical speech development, numerous longitudinal studies have qualitatively observed that children show intermediate periods of increased variability as they transition between stages of stable production (Becker & Tessier, 2011, citing Smith, 1973, Menn, 1976, and Fikkert, 1994). Describing monolingual speech development, Vihman & Greenlee (1987) proposed that some children reuse a limited set of well-practiced speech-motor plans, resulting in a ‘systematic-stable’ pattern of errors, while others fit an ‘exploratory-variable’ profile in which more challenging targets are attempted but inconsistently achieved. A similar dynamic is embodied in the A-map model (McAllister Byun et al., 2016), a theoretical framework positing that children’s speech reflects a balance between accuracy (i.e. pressure to approximate the acoustic target modeled in the environment) and precision (i.e. pressure to use a speech-motor routine stabilized through extensive practice). Findings from these studies suggest that speech development may resemble general motor development in showing accuracy-variability tradeoffs, in which an increase in the variability of speech production is associated with an increase in accuracy.

Contrastively, other studies have reported that consistency in speech production shows a positive association with age and vocabulary size (Holm et al., 2007; Macrae & Sosa, 2015). If local variability-accuracy tradeoffs coexist with a general tendency toward increased consistency over the course of maturation, it is difficult to predict what association will prevail in a cross-sectional study. The present study examines the relationship between accuracy and variability in speech acquisition in bilinguals, a context where there is a lack of previous literature. Bilingual speech development provides an appropriate context to investigate questions about variability because bilingual learners may be exposed to particularly high variability in their language input, which may translate to increased variability in production (Byers-Heinlein & Fennell, 2014).

Bilingual speech development in under-studied language contexts

The risk of misdiagnosis due to the lack of normative data in bilingual speech development represents an increasingly pressing problem as the bilingual population in the USA continues to grow (Fabiano-Smith et al., 2021). Most of the available research on bilingual speech development has been conducted on Spanish-English speakers (Gildersleeve-Neumann et al., 2008; Fabiano-Smith & Goldstein, 2010; Fabiano-Smith & Hoffman 2018). Given the large size of this bilingual population in the US, this emphasis is not inappropriate. However, there are many other languages present in the multilingual population of the US, and these languages may interact with English differently. Without representation of the full range of linguistic typology, there is a risk of misunderstanding the breadth of possible profiles of bilingual experience. Jamaican Creole (JC), the focus of the present study, is an example of an understudied but widespread language, spoken in Jamaica and throughout the Jamaican diaspora (Washington et al., 2017).

JC is one of the main Atlantic English-lexifier creole languages (Harry, 2006), meaning that most of its vocabulary is derived from English. JC and English share extensive phonologies and cognate words but also have distinct phonological characteristics (Washington et al., 2017, 2019, 2021). For example, voiced and voiceless interdental fricatives exist in the phonemic inventory of English but not in JC, and vowel length is phonemically contrastive in JC but not in English. There are also phonotactic sequences that may behave differently in one language than the other: for instance, English features minimal pairs that contrast in the presence or absence of word-initial /h/ (e.g. eat, heat), whereas it regularly undergoes deletion (e.g. /ɛlikɑptɑ/, /hɛlɪikɑptɑ/ for helicopter) and insertion (e.g. /hɔnɑ/, /ɔnɑ/ for honor) in JC (Harry, 2006). Most of the available research on transfer focuses on differences in phonetic inventory between languages, and more research is needed to explore how phonotactic differences may affect patterns of realization in each language context.

Studying bilingual speakers of JC and English has the potential to provide valuable insight on the broader range of bilingual speech patterns, particularly for other pairings between languages with extensive phonological and lexical overlap, such as other creole-lexifier pairings or acquisition of distinct dialectal variants of the same language. In the long term, this will allow better practices of diagnosis and treatment of bilingual populations.

Research Questions

The primary objective of this study is to investigate whether there is evidence of a variability-accuracy tradeoff in bilingual children acquiring JC and English. Following previous literature, we measured variability based on differences in broad transcription across repeated productions of the same lexical item (Holm et al., 2007; Macrae & Sosa, 2015) and measured accuracy using PCC-R (Fabiano-Smith & Hoffman, 2014; Fabiano-Smith et al., 2021) and whole-word accuracy (Ingram & Ingram, 2001). We hypothesized that children who produce more accurate speech will also produce more variable speech, indicating the presence of a variability-accuracy tradeoff in developing bilingual speakers. This is based on research in general motor development that report elevated levels of movement variability as learners make a transition to more mature patterns of motor control (Thelen et al., 1993; Gliga, 2018; Goffman, Ermer, & Erdle, 2002). If such tradeoffs are observed, it would support the idea that children show increased production variability as they transition to a new stage of motor development. We chose to focus on bilingual speakers because it has been posited that elevated variability in bilingual input could lead to greater levels of variability in typical speech production, making it easier to observe an association with accuracy. However, when examining variability-accuracy tradeoffs in bilingual speech, it is crucial that ‘accuracy’ be defined relative to a standard, and a monolingual standard may not be appropriate for speakers acquiring multiple languages. Thus, we additionally examined how accuracy-based scoring in our bilingual sample is impacted when the potential for cross-linguistic transfer is taken into account. We hypothesized that the children’s productions will be scored as more accurate when transfer is considered. Finally, in light of previous research documenting a relationship between vocabulary and production variability (Holm et al., 2007, Macrae & Sosa, 2015), we controlled for expressive vocabulary in our analyses of variability-accuracy tradeoffs.

Methods

Ethical Approval

Approval and/or permission for this study and related activities was obtained from the Institutional Review Board of the University of Cincinnati, the Early Child Commission in Jamaica, and the Board for the participating early learning centers.

Participants

Twenty bilingual children ages 3;4 to 5;1 (12 females, eight males) from two preschools in Kingston, Jamaica were recruited as part of a larger project investigating variability and accuracy in bilingual development.² All children were simultaneous bilinguals who were exposed to both JC and English from a young age at home and school; classroom instruction in English begins from age three for children in Jamaica. To be considered bilingual, participants were required to have at least 20% input and output in JC and English (Pearson et al., 1997) based on parental report. Children were also required to have no history of speech, language, or hearing impairment based on reports by their parents and a teacher or SLP.

All children passed a hearing screening at 25 dB HL for octave frequencies 1kHz, 2kHz, and 4kHz. They were also required to perform within normal limits on the Primary Test of Nonverbal Intelligence (PTONI; Ehrler & McGee, 2008) and met age-based criterion on the Oral Motor subtest of the Diagnostic Evaluation of Articulation (DEAP; Dodd et al., 2006). Information about the demographics and standardized test results of the participants are summarized in Table 1.

Table 1.

Participant information and test score statistics.

	Range	Mean	Standard Deviation
Age	3;4–5;1	4;4	0.53
PTONI Standard Scores	91–149	124.7	13.66
PTONI Percentiles	27–99	88.8	17.96

(Age is listed in year;months); PTONI = Primary Test of Nonverbal Intelligence (M=100, SD=15); Ehrler & McGhee, 2008)

Procedure

The DEAP Word Inconsistency Assessment (WIA) was used as the basis for measurements of accuracy and variability of the preschoolers’ speech. This assessment measures how consistently children produce words across multiple repetitions of the same word (Holm et al., 2007; Macrae & Sosa, 2015). The DEAP WIA consists of 25 words elicited three times using colored picture stimuli. Each participant completed the DEAP WIA in JC and English as part of a separate study (Abu El Adas et al., 2021). For the purposes of the current study, only targets in English were analyzed. The administration of the word list was timed for five minutes, and between each elicitation of the list, the participants were given five minutes of free play.

The data were recorded using an H4N or H6N Zoom recorder and MOVO LV4-C XLR unidirectional lavalier microphone. The microphone was secured onto a vest worn by each participant to minimize background noise and maximize the acoustic quality of the recording. Assessments took place in closed classrooms in the participants’ school. Speech samples were digitized using a sampling rate of 22kHz Hz and 24-bit encoding.

Measurement

Eleven of the 25 words from the DEAP WIA were analyzed for this study. The DEAP has not been normed for bilingual speakers of JC and English, but these eleven words were previously validated for the Jamaican English context and age group of the participants (Washington et al, 2017). The list of words is available in Supplementary Materials.

The eleven words were marked off in the acoustic record using Praat software (Version 6.1; Boersma & Weenink, 2001) and linked to records in Phon (Version 3.0.5; Hedlund & Rose, 2019), a computer software designed for transcription and analysis of speech productions. Six undergraduate and graduate students with previous phonetic training performed broad transcription of the participants’ productions. Prior to transcribing speech in the Jamaican context, students were required to complete specific training, including reviewing the sound system of JC, watching a video of the third author (a native speaker of JC) transcribing a child producing the target items from the WIA, and receiving feedback on practice transcriptions.

The six students transcribed a total of 660 English tokens (11 words x three elicitations x 20 participants) following a transcription procedure similar to that used by Macrae & Sosa (2015). Two students independently transcribed each recording using the Blind Transcription feature on Phon, which prevents one transcriber from viewing the other’s transcriptions. Students were instructed to provide a broad transcription of each production but were allowed to include diacritic marks for salient sub-phonemic differences. The students then jointly viewed both of their transcriptions in one file using Phon and resolved any discrepancies with reference to a consensus procedure outlined by Shriberg, Kwiatkowski, and Hoffman (1984). A consensus was reached after listening to the target words a maximum of three times per disputed phoneme and determining which of the two transcriptions most closely matched the participant’s actual production of the target word.

Coding

To measure variability, the standard scoring procedure for the DEAP WIA was used, such that each production was marked as consistent (1) or inconsistent (0) based on the broad transcriptions of each item across all three repetitions. Productions were counted as inconsistent if at least one token differed from the others by at least one phoneme. Tokens that differed only by the presence or absence of diacritics were marked as consistent. Variability across the repetitions was measured independent of accuracy measures (e.g. three repetitions of /tit/ for teeth was marked as consistent). The percent of items produced consistently was computed across the productions of all target words. These data were taken from a companion study examining the same sample of participants (Abu El Adas et al., 2021).

For accuracy analyses, whole-word accuracy and PCC-R (operationalized below) were calculated for each participant in two ways: 1) based on English adult target transcriptions using the automated functions in Phon, and 2) with reference to JC adult target transcriptions calculated manually (Washington et al., 2017). Obtaining accuracy measurements relative to JC adult productions allowed us to examine whether cross-linguistic transfer affects variability-accuracy tradeoffs.

To calculate whole-word accuracy based on English adult targets (Ingram & Ingram, 2001), every phoneme in an adult target production was compared to those of the participant’s production. An item was coded as an ‘Exact Match’ if every phoneme in the produced word matched the target form, or an ‘Exact Mis-Match’ if the produced form differed from the target by at least one phoneme. The production most similar to the adult target across the three repetitions was used to determine whole-word accuracy for each item. A percent of words produced fully correctly was calculated for each participant.

To calculate PCC-R, the number of consonants that matched an adult-like production of a given target item was divided by the number of correct consonants plus the number of incorrect consonants (i.e. number of consonants substituted, deleted, or inserted), as defined by Shriberg and Kwiatkowski (1984). That number was multiplied by 100 for the percent of consonants correct in that word. For example, if a participant said /ɛlɪfan/ for elephant, the PCC-R for this token would be 75% (3 correct / (3 correct + 1 deleted) x 100). PCC-R was calculated for each production, and the most accurate production across the three repetitions was selected. These values were averaged across the 11 words to generate a PCC-R score for the whole session.

To analyze for transfer from JC to English, each item marked as an ‘Exact Mis-Match’ was compared to transcribed forms considered accurate in the Jamaican context (Washington et al., 2017). Items deemed a possible case of transfer were reclassified as correct (1) and items deemed a true error (i.e. not attributable to language transfer) were marked as incorrect (0). Using the aforementioned example, the production /ɛlɪfan/ was rescored a (1) with a PCC-R score of 100% in the Jamaican context. The percent of possible cases of transfer was calculated for each participant and averaged across all 20 participants. The percent of words produced fully correctly and PCC-R scores were recalculated for each participant after instances of transfer were reclassified as correct.

As noted above, previous research has shown significant associations between speech production variability and vocabulary size (Holm et al., 2007, Macrae & Sosa, 2015).³ To control for vocabulary size when examining the relationship between accuracy and production variability, we administered the Expressive Vocabulary Test–Second Edition (EVT-2; Williams, 2007). The order of administration of the EVT-2 relative to the DEAP WIA was counterbalanced across participants. We used raw rather than standardized scores because the EVT-2 was normed on monolingual speakers of American English, rendering any normative comparisons with our bilingual participants invalid. While even raw scores must be treated with caution when applying a test across cultural and linguistic contexts, we used EVT-2 raw scores as the best available alternative given the lack of standardized vocabulary assessments for JC or English in the Jamaican context.

Data analysis

To examine how accuracy and variability trade off in bilingual children and whether cross-linguistic transfer affects variability-accuracy tradeoffs, four multiple linear regression models were fit using the Stats package (R Core Team, 2021) in RStudio (RStudio Team, 2016). Figures were created using ggplot2 (Wickham, 2016).

All four models included percentage of words produced consistently for the 11 target words as the dependent variable. Because percentage scores are bounded above and below by 0 and 100, percentages were arcsine transformed to better meet the assumptions of linear regression. Two of the four models examined accuracy measures based on the English context only before transfer was taken into account, while the other two examined accuracy with transfer from JC considered. The first model included average whole-word accuracy scores and raw scores from the EVT-2 as fixed effects for each participant, and the second model included average PCC-R scores and raw EVT-2 scores as a fixed effects. The other two models were similar except that whole-word accuracy and PCC-R scores reflected the reclassification of errors considered possible instances of transfer as correct. For all models, both accuracy scores and raw EVT scores were scaled by subtracting the mean and dividing by the standard deviation to increase the interpretability of model outputs.

Results

In the analyses using accuracy measures based only on the English context, the first model using whole-word accuracy scores revealed no significant effect of whole-word accuracy (β = .004, SE = .031, p = 0.891) or EVT scores (β = .009, SE = .031, p = 0.771) on consistency. The second model, which used PCC-R as the measure for accuracy, revealed a significant effect of PCC-R on consistency (β = .080, SE = .029, p = .011). However, the direction of the observed effect was contrary to our original hypothesis (i.e. participants who produced more accurate speech also produced more consistent speech). Figure 1 shows the relationship between PCC-R and percent of words produced consistently in the present sample. There was no significant association between raw EVT score and consistency in the second model (β = −.034, SE = .028, p = .250).

Figure 1.

Percent of tokens produced consistently versus PCC-R across all participants based on the English context only.

Figure 2 shows the percentage of possible instances of cross-language transfer from JC to English for each participant out of the total number of productions coded as errors in English. For 12 of the 20 participants (60%), a majority of the productions marked as errors in English were identified as possible cases of transfer from the Jamaican context. Across all 20 participants, over half of the total number of differences scored as errors qualified as possible instances of transfer.

Figure 2.

Instances of possible transfer from JC to English across all participants. The light grey color shows productions coded as possibilities of transfer while the dark grey color shows productions coded as true errors.

For the analysis using accuracy measures that accounted for possible instances of transfer, the model investigating whole-word accuracy continued to show no significant association between consistency and accuracy (β =.045, SE = .028, p = .129) or EVT scores (β = −.002, SE = .028, p = .929). In the model investigating PCC-R, the association between PCC-R and consistency remained significant in the opposite direction of what was hypothesized (β = .073, SE = .024, p = .008), and EVT scores continued to show no association with consistency (β = −.011, SE = .024, p = .645). Figure 3 shows the relationship between PCC-R and percent of words produced consistently after reclassification of possible instances of cross-linguistic transfer.

Figure 3.

Percent of tokens produced consistently versus PCC-R across all participants after errors reclassified as possibilities of transfer were marked correct.

Discussion

The primary aim of this study was to explore the relationship between accuracy and variability in bilingual children acquiring JC and English. Based on the current literature, we predicted that children who produced more variable speech would also produce more accurate speech, indicating the presence of a tradeoff. The association between PCC-R and percent of words produced consistently was significant, but children who produced more accurate speech also produced more consistent speech. The association between percent of whole words correct and percent of words produced consistently was nonsignificant. These results were unchanged when accuracy scores were recalculated to take possible instances of cross-language transfer into consideration. While these findings are not consistent with our original hypothesis, they have interesting resonance with previous work examining variability in typically developing monolingual children (Macrae & Sosa, 2015).

Variability and Accuracy Measures

Two findings from our study offer insight into ongoing discussions on the role of accuracy and variability in speech development. Firstly, as described in a companion paper (Abu El Adas et al., 2021), the low percentage of words produced consistently amongst all participants in the present study is in line with previous research that reported high rates of variability in typically developing preschool-aged children (Macrae & Sosa, 2015). Secondly, previous work has shown that age and vocabulary size are associated with higher rates of consistency in speech productions, whereby older children and those with larger vocabularies produce more consistent speech (Holm et al., 2007; Macrae & Sosa, 2015). The Lexical Restructuring Model (Metsala & Walley, 1998), which holds that children’s lexical representations become more segmentally specified as their vocabularies grow, may explain this association: greater attention to segment-level detail is thought to support greater consistency of speech productions. Children with more segmental representations of words might also be expected to produce more accurate speech. However, when Macrae and Sosa (2015) tested the association between consistency and accuracy of speech productions, they found no significant relationship. They suggested that the standardized test of articulation used to measure accuracy, a test designed to diagnose SSD, may not have been sensitive enough for typically developing children and proposed that a criterion-referenced measure should be used to further examine the relationship between accuracy and consistency (Macrae & Sosa, 2015). Consistent with this suggestion, the present study found a significant association between consistency and accuracy as measured by PCC-R. This association was significant while controlling for English vocabulary size.

Although this study failed to demonstrate evidence of a variability-accuracy tradeoff, there are several reasons that such tradeoffs may be observed using a different methodology. In motor development, variability-accuracy tradeoffs are thought to occur immediately before the transition to a more mature stage of motor control (Gliga, 2018). Because this study examined the participants’ speech samples at a single point in time, it is difficult to definitively state whether or not the participants were transitioning between developmental stages of speech acquisition. Moreover, examining average values of variability and accuracy across participants who vary in age and across phonemes that are acquired at different stages of development may have prevented us from observing time-specific variability-accuracy tradeoffs. It remains possible that variability-accuracy tradeoffs could be observed through longitudinal tracking of speech development, particularly through focusing on individual trajectories of change over time.

It is also possible that the measure of variability used in this study, segmental consistency across repeated productions based on broad transcription, is not ideal for capturing variability-accuracy tradeoffs of the type described in motor development. Previous research has shown that transcription-based measures may fail to identify distinct sources of variability at different levels of processing, such as at the level of motor planning (Goffman et al., 2007). Preston and Koenig (2011) suggested that acoustic measures may better capture differences in speech motor processing, while segmental measures reflect phonological processing. Future studies that consider acoustic and/or articulatory measures of variability may be more successful in identifying variability-accuracy tradeoffs in speech development. Another possible direction would be to explore variability of a single phoneme across different phonetic contexts, such as in onset versus coda position.

Cross-linguistic Transfer

Upon qualitatively examining English productions marked as ‘inaccurate’, we labeled more than half of all errors as instances of possible transfer from JC to English. For example, many participants produced /tit/ for the target word /tiθ/ and /tankju/ for the target /θæŋkju/. The phoneme /θ/ does not exist in the Jamaican phonetic inventory and it is likely that the substitution for /t/ was influenced by transfer from JC to the English context. The high rate at which ostensible errors reflect possible instances of transfer is consistent with findings from other JC-English bilinguals (Washington et al., 2017). It is also consistent with models of phonological acquisition (e.g. Vihman & Greenlee, 1987; McAllister Byun et al., 2016) suggesting that child language learners show some preference to reuse a limited set of well-practiced speech-motor plans. In such models, children navigating the acquisition of two sound systems may favor phonemes that are shared across both of their languages, since those items have had more opportunities to stabilize with practice. Because /t/ is acquired before /θ/ in the English context, it is plausible that some participants employed a well-practiced, shared phoneme instead of attempting to articulate a challenging phoneme that only exists in one of their languages. On the other hand, models such as McAllister Byun et al. (2016) also predict variability-accuracy tradeoffs, and the hypothesized associations in the present study were not observed even after reclassification of possible instances of transfer. This suggests that more research must be completed to explore the relationship between accuracy and variability in bilingual speakers.

Limitations and Future Directions

Several limitations of the present study should be considered for future studies of bilingual speech acquisition. The small sample size makes it difficult to generalize the results of this study to the wider population of children acquiring JC and English across geographic and linguistic contexts. Future studies should sample a larger population and compare bilingual speakers from different locations (e.g. JC-English bilinguals in New York City) to examine the effect of relative exposure and dominance of each language on patterns of variability, accuracy, and transfer. Production variability may have also been influenced by the participants’ exposure and proficiency in both JC and English. Though the participants in this study were simultaneous bilinguals, future studies should consider how individual, speaker-level differences may affect variability in each language. This study analyzed participants’ productions in English only, in part because it represents the language context where misdiagnosis is most likely to occur. Future studies should analyze responses in both languages to provide more comprehensive information about variability and accuracy in bilingual speech development. Similarly, it is a limitation that the transcribers in the current study were speakers of American English and lacked prior familiarity with JC or English in the Jamaican context. We believe that the number of errors reflecting interference from the transcribers’ first language was limited, albeit not eliminated completely, by the rigorous training procedure our transcribers were required to complete. Increasing the representation of speakers of JC in the field of communication sciences and disorders could address this issue while also supporting improved service delivery for JC-speaking communities. Finally, as noted above, collecting and analyzing data longitudinally will allow for better exploration of variability-accuracy tradeoffs in speech development.

Clinical Implications.

The high rate of possible transfer reported in this study emphasizes the importance of being culturally-appropriate in defining accuracy for bilingual individuals, especially in contexts like JC and English where the languages are closely related (Wright Karem & Washington, 2021). This finding has important implications for research and the clinical setting – it is a reminder that measuring accuracy in a single linguistic context without considering between-language interactions likely contributes to the overdiagnosis of SSD in bilingual children. The establishment of standardized assessments for bilingual speakers may help reduce the risk of misdiagnosis in this context (Fabiano-Smith & Hoffman, 2018). It is important to continue increasing the knowledge base on bilingual speech development, especially in understudied contexts like JC and English, with the ultimate goal of improving culturally-appropriate speech and language services for bilingual individuals.