Specificity of Phonological Representations in U.S. English-Speaking Late Talkers and Typical Talkers

Participants

All procedures involving human subjects in this study were approved by the Institutional Review Board at Northwestern University (STU000208022). The present study was conducted according to guidelines laid down in the Declaration of Helsinki, with written informed consent obtained from a parent or guardian for each child before any assessment or data collection. Participants were recruited through online advertisements, flyers distributed in the community, and through existing research databases. Participants were eligible if the child was between the ages of 22 and 30 months of age, was born at 37 weeks gestation or later, and if the caregiver reported that their child only heard English at home, had no concerns about their child’s hearing, and had no concerns that their child may have an autism spectrum disorder. Additional screening for autism and non-verbal cognitive delay was conducted during research visits, as described below.

Our original recruitment goal was to enroll 40 LTs and 40 TTs, yet recruitment and data collection were curtailed due to the COVID-19 pandemic. In total, 57 participants were enrolled in the study. Twelve participants were excluded from the analysis: six toddlers met screening criteria for autism; two toddlers met screening criteria for non-verbal cognitive delay; one toddler’s caregiver did not complete the MCDI; and three toddlers did not contribute at least 4 trials in each condition during the Looking-While-Listening task (a criterion set in the pre-registered analysis plan). A total of 45 toddlers were retained for analyses. Twenty-three toddlers were classified as TTs, and 22 as LTs, using the definitions below.

Participant characteristics are given in Table 1. Although our recruitment strategies included advertising widely in the community and using recruitment materials including images of racially and ethnically diverse families, the final sample consisted of families who were predominantly White and of high socioeconomic status, which is not representative of the wider population. Toddlers in the TT and LT groups were similar on most characteristics. However, toddlers in the LT group were significantly younger than toddlers in the TT group (t(42.24) = 2.60, p = 0.01). Consistent with our analysis plan, we include child age as a covariate in all models, together with race (White vs. not White), biological sex, and caregiver education (college or less vs. post-secondary degree). A detailed discussion of how the age difference between groups might impact the results is presented in the discussion.

Screening Measures

Screening Tool for Autism in Toddlers and Young Children (STAT; Stone & Ousley, 2008). The STAT is a Level 2 autism screener for children between the ages of 24 and 36 months, with studies demonstrating its validity in younger children as well (Stone et al., 2008). The STAT assesses three broad domains: Play, Social Communication, and Imitation Skills. Assessors use a series of activities throughout the assessment, including rolling a ball back and forth with the assessor, giving the child a sealed jar of snack that the child cannot open themselves, and giving the child a bag of toys. The STAT classifies children receiving a score of 2 or higher as having a higher likelihood of having autism. Six toddlers scored a 2 or higher on the STAT and were excluded from the study; these toddlers were also referred for further assessment through Illinois’ Early Intervention System.

Childhood Autism Rating Scale, Second Edition (CARS-2; Schopler et al., 2010). Two toddlers in the study did not complete the STAT because their second study visit was cancelled due to the 2020 Covid-19 pandemic. To screen for the presence of autism in these participants, we utilized the CARS-2. The CARS-2 is an observational coding system. Observers rate children on 15 items, including items that assess how the child relates to others, how the child uses objects, and how the child adapts to change. The CARS-2 can be scored based on observations of the child in a variety of settings. For the current study, a trained rater scored the CARS-2 based on recordings of a 15-minute caregiver-child interaction. Both toddlers scored in the “minimal to no risk of autism spectrum disorder” category and were included in the study.

Mullen Scales of Early Learning, Visual Reception Scale (MSEL-VR; Mullen, 1995). The MSEL is a norm-referenced standardized assessment of early cognitive development in children from birth to 68 months of age. This subscale assesses children’s non-verbal cognitive skills. Children complete a range of activities, such as completing a foam-board puzzle, making object associations, and matching shapes. The Visual Reception scale produces a T-Score based on the child’s age. In the current study, toddlers needed to achieve a T-Score of 35 (1.5 standard deviations below the mean) or higher to be included in the study. Two children obtained a T-Score lower than a 35 and were excluded from the study.

Expressive vocabulary assessment and classification into TT and LT groups

The MacArthur-Bates Communicative Development Inventory – Words and Sentences (MCDI; Fenson et al., 2006) was used to measure toddlers’ vocabularies and to classify children into the Later Talker (LT) and Typical Talker (TT) groups. The MCDI is a standardized and norm-referenced caregiver report measure of children’s communication development, including expressive vocabulary and syntax. Toddlers who received a score at or below the 15^th percentile (Beckage et al., 2011; Hodges et al., 2017; Lu et al., 2020), and/or children who had no reported two-word combinations were classified as LTs. All other children were classified as TTs.

It is important to note that we utilized a definition of LT status that included two possible criteria: a child was classified as an LT if either they had a vocabulary score on the MCDI at or below the 15^th percentile or they were not combining words. Using low expressive vocabulary or lack of word combinations to define LTs is consistent with definitions used in previous studies and in clinical settings (Beckage et al., 2011; Hodges et al., 2017; Lu et al., 2020; Rescorla et al., 1997; Rescorla & Schwartz, 1990). However, the youngest children in the current study were 22 months of age, which is 2 months younger than the typical criterion of not combining words by 24 months. Thus, one might worry that classifying children as LTs for not combining words at 22 months is too inclusive a definition for LTs. However, norming data from the MCDI shows that in a sample of 911 toddlers, over 90% were combining words by 22 months (Fenson et al., 2000; Hoff et al., 2012). These data support our criteria of classifying toddlers as LTs if they were not yet combining words by 22 months, as this places them in the bottom 10% relative to their same-aged peers, consistent with our criteria for vocabulary size.

Looking While Listening Mispronunciation Task

Twelve monosyllabic words that are acquired early in development (Fenson et al., 2006) were selected. Six yoked pairs were created by matching word pairs on age of acquisition as derived from the US English data on Wordbank, an online repository of MCDI data (Fenson et al., 2006; Fernald et al., 2013; Frank et al., 2017; Thal et al., 2013), lexical frequency in child-directed speech as derived from the North American English collection in CHILDES (MacWhinney, 2014; Sanchez et al., 2019), and phonological neighborhood size from the Clearpond database (Marian et al., 2012). For each word, a mispronunciation was created by altering the initial phoneme’s voicing (6 words) or place of articulation (6 words). See Table 2 for stimuli descriptions. Previous work has shown that toddlers as young as 19 months of age are sensitive to changes in both voicing and place of articulation in the mispronunciation paradigm (White & Morgan, 2008). In this design, sensitivity to mispronunciation refers to the proportion of frames looking to the target vs. the distracter image in the correct pronunciation condition relative to the proportion of frames looking to the target vs. the distracter image in the mispronunciation condition. This is captured by the ‘condition’ term in the models.

Each yoked pair was presented 4 times during the experiment. Each item was presented once as a target with the correct pronunciation, once as a target with the mispronunciation, and twice as a distractor for its yoked word. Four pseudo-random counterbalanced orders were created. In each order, side of target presentation was counterbalanced. Orders were adjusted so that the first item was a correct pronunciation and no more than two mispronunciations were presented consecutively. Additional filler trials with colorful images and cartoon characters were interspersed every 3 to 4 trials to maintain toddlers’ attention.

On each trial, two images were presented for a total of 7000 ms. After 1500 ms, toddlers heard the speaker say “Look!” At 3000 ms, the target stimulus word was played (e.g., “Couch!”). The images remained on the screen for another 4000 ms in silence. The trials thus consisted of a pre-naming phase (0-3000 ms) and a post-naming phase (3000 – 7000 ms). See Figure 1.

Audio stimuli were recorded by a female Midwest American English speaker who used infant-directed prosody. Multiple examples of each word were recorded. A group of adults then listened to both correct and mispronunciation word tokens in isolation. Only tokens on which all listeners correctly identified the initial phoneme were included as stimuli in the experiments. Words were on average 0.7315 (sd = 0.0856) seconds in duration. The correct pronunciation and mispronunciation word sets did not differ in their length (t(21.784) = −0.378, p = 0.709), mean pitch (t(21.719) = 1.028, p = 0.315), or listeners’ average clarity ratings (t(21.997) = 1.481, p = 0.153).

Two sets of images were selected for each yoked pair of stimuli (e.g., there were two image pairs for the duck-bed pair, two for the foot-book pair, etc.). We used two image pairs to maintain toddlers’ interest and attention during the experiment (see Supplemental Materials for an example). Images were resized and centered in a 600 x 600 pixel area. The narrower of the two images was adjusted so its maximum length would be 600 pixels. The other image was then resized so that it had approximately the same number of pixels. Images were placed on a 1920 x 1080 gray background.

Procedure

All toddlers completed a brief warm-up upon arriving at the lab and played with a set of toys with the research staff. All testing was completed by a PhD student receiving clinical training in speech-language pathology. After completing the warm-up, toddlers completed the eye tracking experiment and other testing procedures, including completing the Mullen and the STAT while their parent completed questionnaires.

The LWL experiment was presented using Tobii Studio Pro. Toddlers either sat in a highchair or in their caregiver’s lap in a sound-attenuating area of the testing room. Caregivers were given a pair of occluded sunglasses to wear and were instructed not to repeat any of the target words and to only speak to redirect their toddler’s attention to the stimuli. Visual stimuli were presented on a 24” ASUS VS248H-P screen. Auditory stimuli were presented at 65 dB through Bose Companion 2 Series III Multimedia Speakers on either side of the screen. Toddlers’ gaze was captured using a Tobii X3-120 corneal-reflection eyetracker using a 120 Hz sampling rate and via video recording. The entire LWL procedure took approximately 5 minutes to complete.

Eye Tracking Data Preparation

Raw gaze data was exported from the Tobii Pro Studio software. The average gaze coordinates provided by Tobii Pro Studio were used in analyses. Gaps in data less than 75 ms, most likely representing blinks or brief periods of track loss, were interpolated using scripts from Unger et al. (2020). An area of interest (AOI) was drawn around each image.

Data from six participants were hand coded using peyecoder (v 1.1.13; Olson et al., 2020). Two of these participants had extreme track loss, despite optimal viewing position and angle. The first and last author reviewed plots that aggregated fixations from each toddler in a cartesian plane to identify calibration drift. Plots were reviewed in a random order, and identifying information was removed from each plot. The authors were blind to LT status, and plots contained data from both correct pronunciation and mispronunciation conditions. From this inspection, four toddlers were identified as having significant drift. Upon reviewing the videos for these toddlers, it was observed that these toddlers frequently shifted their head position during the experiment, despite remaining attentive. These toddlers’ data were hand-coded in peyecoder as well. A separate participant with optimal data from the Tobii eye tracker was also hand coded, and the Tobii gaze data was compared with hand-coded data. The correlation between proportion of looks to the target image in 100 ms bins was 0.93, demonstrating that the two methods gave very similar results.

Data were processed in R (v 4.0.3) (R Core Team, 2014) using the eyetrackingR package (Dink & Ferguson, 2015) and the lme4 package (Bates et al., 2015). Looks to neither AOI (shifts, looking at parts of the screen outside the AOI, trackloss, looks away from the screen, etc.) were removed. In order to reduce autocorrelation among adjacent timepoints in growth curve analyses, temporally adjacent samples from the eye tracking data were collapsed into 100 ms bins, following standard eye tracking analysis procedures (Mirman, 2014). For window analyses, all gaze data from the analysis window were collapsed into a single bin.

Analysis Plan

Window Analysis

Looks to the target and looks to the distractor were entered into a logistic multilevel model with fixed effects of condition (dummy coded as 0=correct, 1=mispronunciation), talker type (dummy coded as 0=TT, 1=LT), and a condition x talker type interaction, as well as random effects for participant, participant by condition, lexical item, and lexical item by condition. Both LTs and TTs were included in this model. Results for this analysis are given in Table 3.

Because TTs are the reference group, the intercept in this model represents TTs’ looks to the target image in the correct pronunciation condition. Typical talkers looked more to the target than the distractor image in the correct pronunciation condition (Wald = 5.479, p < .001). Across both conditions, LTs looked less at the target than TTs, although the effect of talker type was only marginally significant (p = 0.058), see Figure 3. There was a significant effect of condition (Wald = −3.793, p < .001), indicating that toddlers looked more to the target image in the correct pronunciation condition compared to the mispronunciation condition. Finally, the LT by Condition interaction was not significant (p = 0.256), indicating that the effect of Condition did not differ between TTs and LTs. These results suggests that across both LT and TT groups, toddlers were sensitive to mispronunciations and looked less to the target image when words were mispronounced than when they were correctly pronounced.

Growth Curve Analyses

In order to capitalize on the temporal precision offered by the LWL methodology, we also analyzed the timecourse data (illustrated in Figure 4) using growth curve modeling. Growth curve analyses were conducted using logistic fixed effects models. The main model contained orthogonal polynomials for linear and quadratic time (beginning 300 ms after stimulus onset to 1800 ms), as well as their interaction with talker status and mispronunciation condition and previously stated covariates. The model was first run with maximal random effects of intercept, linear time, and quadratic time nested in both participants and participants by condition. This model did not converge, however, and so the random effect of linear time nested within participant was removed, and the intercept and quadratic time random effects nested in participant were set to be uncorrelated (Barr et al., 2013). Results are shown in Table 5.

This analysis revealed a main effect of condition (Wald Z = 7.437, p < 0.001), confirming the mispronunciation effect found in the window analysis. There was also a main effect of talker type (Wald Z = −2.025, p = 0.043), indicating that LTs were less accurate than TTs overall. The main coefficients of interest were the two-way interaction between condition and talker type, and the three-way interactions between condition, talker type, and linear and quadratic time. None of these interactions were significant (talker by condition: Wald Z = 1.177, p = 0.239; talker by condition by linear time: Wald Z = 0.055, p = 0.956; talker by condition by quadratic time: Wald Z = −0.587, p = 0.557). As in the window analysis, these results revealed that both TTs and LTs looked less to the target in the mispronunciation condition relative to the correct pronunciation condition. In these analyses, the effect of talker status was statistically significant, revealing that LTs looked less to the target image overall, but this was not modulated by mispronunciation condition.

Exploratory Analyses: Effect of item phonotactic probability

Previous literature has suggested that words’ phonotactic probability affect toddlers’ sensitivity to mispronunciations (MacRoy-Higgins et al., 2014). In particular, when assessing newly-learned words, MacRoy-Higgins et al. (2013) found that typical talkers showed greater sensitivity to mispronunciation on high-phonotactic probability words compared to low-phonotactic probability words, while phonotactic probability did not influence late-talkers’ sensitivity to mispronunciations. Indeed, the results from our mixed effects models suggested considerable item-level variability in the magnitude of the mispronunciation effect (see Figure 5). Thus, we considered whether phonotactic probability might have a differential impact on TTs’ and LTs’ sensitivity to mispronunciations. Because the study was not designed to assess the impact of phonotactic probability, we also included other item-level covariates: age of acquisition, frequency in child-directed speech, number of phonological neighbors, and mispronunciation type (change in voicing or change in place). Random effects of participant, participant by condition, target word, and target word by condition were entered as well. The main term of interest was the 3-way-interaction between talker type, phonotactic probability, and mispronunciation condition.

Results revealed that this 3-way-interaction was significant: phonotactic probability altered the mispronunciation effect, but did so differentially for LTs and TTs. TT’s appeared to be sensitive to mispronunciations in words of varying phonotactic probabilities, whereas’ LTs’ sensitivity to mispronunciations depended on the phonotactic probability of the word. However, since this is an exploratory analysis and the study design was not designed to test the effect of phonotactic probability, this issue merits further study.

Limitations and Constraints to Generality

A limitation of the study is that the group of late talkers was significantly younger than the group of typical talkers. Though we controlled for age in all statistical models to account for this, one might nevertheless worry that differences between the groups could be explained by age rather than late talker status. However, since we found that both groups were sensitive to mispronunciations and did not find evidence that late talkers differed from typical talkers in the magnitude of their sensitivity to mispronunciations, it is unlikely that the age difference between groups influenced the main findings.

This study focused on monolingual English-speaking toddlers living in the U.S. However, a significant limitation of the study is that the sample is not representative of the population of English-speaking U.S. children. Though not by design, the majority of toddlers in the study were white (80% in our study vs 48% of 0 to 4 year-olds in the United States population as of 2020; (The Annie E Casey Foundation, 2022), and all but one toddler’s caregiver had completed a bachelor’s degree or above, compared to 35% of adults aged 18 years or older in the U.S (U.S. Census Bureau, 2022). Future studies should strive to include a more diverse sample of participants to increase the generalizability of findings.

In addition, toddlers in the current sample passed a screening test for ASD, scored within 1.5 SD of the mean on a test of nonverbal cognitive abilities, and were not reported to have been born prematurely. Thus, we expect these results to generalize to monolingual English-speaking U.S. toddlers with similar characteristics but may not generalize to toddlers learning languages other than English, toddlers learning more than one language, toddlers at risk for ASD or general cognitive delay, and toddlers who were born prematurely.

The study was interrupted by the COVID-19 2020 pandemic, which curtailed our recruitment and testing efforts and resulted in a reduced sample size than originally intended. While interaction terms assessing the difference in sensitivity to mispronunciations between LTs and TTs failed to reach statistical significance, they were in the expected direction of LTs showing smaller mispronunciation effects. With a larger sample, these effects may have been statistically significant. Although the current study leaves open the possibility that there may be differences in the magnitude of the mispronunciation effect between TTs and LTs, it is quite clear from the present findings that LTs – as a group – are sensitive to mispronunciations.

Conclusion

Using a looking-while-listening mispronunciation paradigm, we found that toddlers with typical expressive vocabularies and those who are late talkers showed sensitivity to mispronunciations. This suggests that in both groups, stored phonological representations of familiar words contain sufficient detail to detect minor changes in pronunciations, at least in word-initial position. This stands in contrast to findings showing delayed expressive phonological development in LTs. Future research should explore LTs’ sensitivity to mispronunciations of phonemes in non-initial positions and further probe the impact of phonotactic probability on LTs’ processing of mispronunciations.