Respect the surroundings: Effects of phonetic context variability on infants' learning of minimal pairs

One of the major challenges young infants face when learning their native language(s) is to find out which acoustic properties of speech are relevant in the phonological system of the language they hear and which are not. Two main approaches on how children may solve this problem have been put forth: one emphasizes the role of the emerging lexicon in the establishment of phonological categories and the other emphasizes the role of phonetic properties in the speech input. The goal of the present study is to further explore the role of the speech input by investigating the so far understudied role of phonetic context variability.

Exploiting properties of their speech input, infants' highly efficient mechanisms for distributional analysis may add to the formation of phonological categories. Here, the basic idea is that infants cluster sounds onto an acoustic space defined by a set of phonetic dimensions: exemplars whose phonetic parameters have a high frequency of occurrence serve as central instances of a category with less frequent but acoustically similar exemplars grouped around such distributional peaks (e.g., Maye et al., 2002). This reconstruction of the acoustic space paves the way to language-specific phonological categories, which in turn leads to the emergence of language-specific effects on speech perception (e.g., Werker and Tees, 1984). This perceptual attunement happens during the first year of life, which suggests that infants may rely on a relatively stable system of native phonological categories for word learning.

Seminal work by Stager and Werker (1997), however, showed that 14-month-olds struggle in a word learning task that requires the processing of minimal pairs while they succeed when the words are phonetically more distinct. Learning of minimal pairs in this paradigm is not impossible but has been demonstrated when the task enhances the referential character of the labels (Fennell and Waxman, 2010), when the labels are presented by multiple speakers (Höhle et al., 2020; Rost and McMurray, 2009) or when the critical sound contrast has specific enhancing acoustic properties (Archer and Curtin, 2018; Curtin et al., 2009). These findings cast doubt on the assumption that phonological categories are ready for use in word learning at the beginning of the second year of life and gave a new impetus to considering the role of the emerging lexicon itself as a source for phonological categories. One influential hypothesis has been that lexical growth, especially learning of words from dense lexical neighborhoods, leads to stable representations of phonological categories (e.g., Metsala and Walley, 1998; Storkel, 2002). However, an essential role of phonological neighbors (i.e., minimal pairs) has been questioned, as minimal pairs are scarce in children's early vocabulary and their presence in the child's lexicon is not predictive for their ability to detect mispronunciations in a minimal pair contrast (e.g., Swingley, 2003). Alternatively, it has been proposed that words which are not minimal pairs but show a higher degree of phonological distinctness may play a more effective role on the path to a more mature phonological system (Feldman et al., 2013; Thiessen, 2007).

Thiessen (2011) tested this proposal using the same procedure as the original studies by Stager and Werker (1997) but extended with an initial familiarization phase. In this design, infants were first familiarized with either a minimal pair (/daɡu/, /taɡu/) or a non-minimal pair with a larger segmental difference (/dabo/, /taɡu/) without the simultaneous presentation of an object as a potential referent. Following this, infants entered a habituation phase, in which they were presented with an object-label association involving only the label /ta/. Finally, infants were tested with either this trained association or with the same object but presented with a minimally different label /da/. Results showed an increase in looking times in the mismatch trial only for those infants who had been exposed to the non-minimal pair during familiarization. In a second experiment, Thiessen (2011) tested whether this kind of familiarization would also generalize to a vowel context that had not been presented during the familiarization. In one condition, infants were familiarized with only /dabu/ and /taɡu/, while in the other condition, four additional items with quite variable segmental shape were presented (/dusil/, /tukʊl/, /difo/, /tila/). The label used in the habituation phase was /dιv/, which was contrasted with /tιv/ in the test phase. Only infants exposed to the more variable familiarization condition showed a mismatch effect. These results lead to the question of whether it is the overall segmental diversity in the variable familiarization that gives rise to this supportive effect or the specific variation of the vowels following the crucial consonants.

The latter option is based on findings from a long tradition of phonetic research elaborating on the acoustic signatures of consonants and their dependence on their vocalic context. In a series of landmark studies, Liberman and colleagues demonstrated that cues to segmental contrasts are highly context dependent. More specifically, in consonant-vowel combinations as in /bV…/ versus /dV…/, formant transitions to the vowel crucially depend on the preceding consonant (Liberman et al., 1952; Delattre et al., 1955). For example (see Fig. 1), consider the contrast between /ba/ and /da/. For voiced stops, an important acoustic cue to place of articulation is the initial portion of the second formant (F2) transition from the release of the stop into the vowel. If this formant transition rises, then the stop is a/b/, but if it falls, it is a /d/. Moreover, the direction of the second formant transition varies with the following vowel, for instance, it rises slightly in /bu/ but falls dramatically in /du/. The notion of context-dependent acoustic manifestations to place of articulation informed subsequent signal processing methods focusing on short-term spectra. Integrating spectral information over time with such methods as linear predictive coding (Blumstein and Stevens, 1979) has led to robust statistical recognition results on consonantal place of articulation [see, especially, Forrest et al. (1988) for discriminant analyses based on pooled data from different vowel contexts]. Others have pursued specific dynamic information as in locus equations which emphasize formant value relations at different time points, as in F2 at the release of the consonant and F2 at the middle of its following vowel (Sussman et al., 1991; Iskarous et al., 2010). Shared among the above approaches is the demonstration that the signatures of consonants are better revealed when they are embedded across varying adjacent vowels. For perception, the idea that variable contexts enhance recognition is not new. Across all sensory modalities, perceptual systems are maximally sensitive to change and presentation of the same event in different contexts enhances the distinctiveness of that event (Kluender et al., 2003; for a review, see Winn and Stilp, 2019).

In continuity with this broad range of findings and recalling that in the classic Stager and Werker (1997) result on infants' struggling with /bih/ versus /dih/ the consonantal contrast was presented before only a single vowel, we hypothesized that presenting a /b/ versus /d/ contrast in different vocalic contexts would enhance the consonantal distinction and lead to better learning outcomes. Using the same method as Thiessen (2011), 14-month-olds were exposed to one out of two conditions: they were familiarized with /bVk/ and /dVk/ syllables that only varied in the vowel next to the crucial locus of contrast (e.g., /biːk/, /buːk/, /diːk/, /duːk/) or with /buC/ and /duC/ syllables that shared the vowel but varied in their final consonant (e.g., /buːk/, /buːn/, /duːk/, /duːn/). Both conditions include contextual variability with the difference that in the latter condition the locus of variability is not in the transition from the initial consonant to the vowel, the CV part, but rather in the non-proximal to the crucial locus of the contract VC part of the signal. It is in the CV part where rich information about the identity of the consonant is localized. We thus hypothesized that learning outcomes will be better in the former than in the latter condition.

The analysis sample comprised of 40 typically developing 14-month-old monolingual German-learning infants. Age and gender were evenly distributed between the two familiarization groups: variable vowel condition (mean age: 14.0 months, range: 13.1 to 15.0 months; 10 girls), variable Consonant condition (mean age: 14.1 months, range: 13.2 to 15.0 months, 10 girls). 19 additional infants were tested but excluded because of fussiness and crying (N = 15), not meeting recruitment criteria (N = 3), and experimenter error (N = 1). Informed consent was obtained from the parents before the experiment. The study was approved by the ethics committee of the University of Potsdam.

The syllable used during habituation was /buːk/. The contrasting syllable, /duːk/, in the test phase only differed with respect to place of articulation (labial versus coronal). Two different sets of CVC syllables were prepared for the two familiarization conditions. In the Vowel condition, all syllables contained the consonant /k/ in the coda and either the voiced labial /b/ or the voiced coronal /d/ in the onset. Five different long vowels occurred in the syllable nucleus, with each of the vowels paired once with the labial and once with the coronal, resulting in ten syllables that formed five minimal pairs (/baːk/, /daːk/, /beːk/, /deːk/, /biːk/, /diːk/, /boːk/, /doːk/, /buːk/, /duːk/). In the Consonant condition, syllables shared the same long vowel but had variable final consonants (/buːp/, /duːp/, /buːt/, /duːt/, /buːn/, /duːn/, /buːm/, /duːn/, /buːk/, /duːk/).

The syllables were produced by a female native speaker of German in a mild child-directed manner. One token of each syllable was selected. For the familiarization phase, the ten different syllables were presented twice with a stimulus-onset asynchrony of 2 s yielding a duration of 40 s. For the habituation and test trials, the syllable /buːk/ (or /duːk/) was duplicated such that ten identical tokens occurred within a sound file with a stimulus-onset asynchrony of 2 s yielding a duration of 20 s. For the novel trial at the end of the experiment, a previously not presented syllable /løːk/ was used.

Formant transitions in our familiarization stimuli exhibit the expected vowel-dependent patterns; see Fig. 2. For example, in the /i/ context, the only context used in Stager and Werker (1997) and where no learning was reported, the formant transitions are comparable across the labial versus coronal consonant. F2 rises in both /bi/ and /di/, albeit with a more pronounced rise in /bi/ just as in the classic studies on American English (e.g., Delattre et al., 1955). When one considers other vowels, more salient differences become evident. For example, in /ba/–/da/, we find the expected clear contrast in F2 transition: in /ba/, F2 rises towards the vowel but falls in /da/. In /bu/–/du/, differences are also rather salient and so on.

Three unfamiliar objects of different shapes and colors (see Fig. 3) were selected from the NOUN database (image nos. 2061, 2002, 2015; Horst and Hout, 2016). The green and the red object served as referents during the habituation and test phases; the blue object served as a new object for presentation during the test phase as per the usual practice in the paradigm we adopt and describe in Sec. 2.3.

The procedure was the same as that in Thiessen (2011), which in turn is a variant of the habituation-switch paradigm (Stager and Werker, 1997), and was implemented using habit 2 (version 2.1.25) (Oakes et al., 2019). Infants were seated on their caregiver's lap in front of a screen. Caregivers were instructed to close their eyes and to avoid any interaction with their child during the experiment. The experiment started with a 40 s long familiarization phase during which infants were exposed to syllables in one of the two familiarization conditions while the screen displayed a static, colorful checkerboard.

Directly after familiarization, the habituation phase started. During habituation, one object (counterbalanced between infants) was paired with the syllable /buːk/ (see Fig. 3 for an illustration of the procedure). For each trial, infants' looking duration to the screen was coded online by the experimenter. When the infant looked away for more than 1.5 consecutive seconds, the trial was stopped and the next one started. The habituation criterion was a 50% drop in the average looking duration in three consecutive trials compared to the mean of the first three trials. The habituation ended after the infant reached the habituation criterion or after maximally 30 trials and was immediately followed by the test phase.

In the test phase, the object presented during habituation appeared in two types of trials. In Same trials, the object was presented together with the same syllable (/buːk/) as in the habituation phase. In Mismatch trials, the object was presented together with the minimal pair syllable of the habituated label (/duːk/). Overall, three Same alternated with three Mismatch trials (the first trial type was counterbalanced across participants). The final trial of the test phase was a so-called Novel trial combining a previously unseen blue object with the novel syllable /løːk/ as a check for attention. Looking times were measured in the same way as in the habituation phase. The whole experiment lasted between five and eight minutes.

Infants in the Vowel condition took on average 105 s (45–315) or 10.1 trials (5–21) to habituate. In the Consonant condition, the averages were 137 s (46–448) and 12.0 trials (5–27 trials). This difference is not significant according to a two-sample t test (duration: t = 1.09, p = 0.285, number of trials: t = 1.06, p = 0.300).

Looking times of the seven test trials (Fig. 4) were analyzed in a linear mixed-effects model using package lme4 (Bates et al., 2015) in r (version 4.0.2) (R Core Team, 2020). The within-participant factor Trial type was dummy-coded such that Mismatch and Novel trials are compared to Same trials (= reference level). The between-participant factor familiarization condition was effects-coded (±1/2) such that the simple effects refer to the grand mean. Trial position was entered as a continuous predictor with the reference value set to the first trial, i.e., the three Same and Mismatch trials were coded as 0, 1, 2, with the Novel trial as 0. These three fixed effects were allowed to interact. The random component structure consisted of random intercepts for participants and objects (green or red) as well as individual adjustments for the effects of trial and position including their interaction. The model explained 60.4% of the variance. Across both familiarization conditions, the difference between Same and Mismatch trials (size: 4.2 s, t = 4.40) as well as between Same and Novel trials (size: 10.5 s, t = 8.58) were significant. While the familiarization condition did not interact with Same (t = 1.12) or Novel trials (t = 1.61), it did interact with Mismatch trials: the difference between Same and Mismatch trials was by 4.9 s larger in the Vowel group than in the Consonant group (t = 2.58). A follow-up analysis with the Consonant condition as reference level confirmed that a difference between Same and Mismatch trials was not present in this familiarization condition (size: 1.7 s, t = 1.28). Trial Position interacted neither with Same (t = 0.79) nor with Mismatch trials (t = 1.26) across both familiarization conditions, and it did not interact with Familiarization condition (t = 1.60) either. There was, however, a three-way interaction between Mismatch trials, position, and familiarization (t = 2.02) indicating that the mismatch effect was differently affected by trial position in the two familiarization conditions.

We tested the hypothesis that infants benefit from information provided by variable vocalic contexts in establishing consonantal contrasts. This hypothesis was supported by our results as infants familiarized with the varying vowel context but not infants familiarized with the varying final consonant context responded with longer looking times in the test phase when presented with a label in which the two consonants had been exchanged.

Several aspects of our findings are specifically notable. Most importantly, our findings demonstrate that, in detecting the phonologically relevant sound distinctions in their language, infants can make use of specific acoustic properties in their speech input. In our case, the acoustic characteristics of the context in which a sound occurs are informative about the phonetic properties expressing a phonological contrast and, as evidenced in our results, infants as young as 14 months can exploit this information. This sheds light on another dimension about the potential role that speech input plays in the formation of phonological categories during early first language acquisition—in addition to the potential role of frequency distributions of sound exemplars (Maye et al., 2002) or specific acoustic properties of the sounds themselves (Archer and Curtin, 2018; Curtin et al., 2009). Further, our results confirm that variability in infants' speech input can be useful for mastering the acquisition of language, as previously shown, for example, for variability deriving from multiple speakers in minimal pair word learning (Rost and McMurray, 2009; Höhle et al., 2020) or in learning of non-adjacent dependencies (Gómez, 2002). Importantly, our results suggest that not all types of contextual variability have the same supportive effect: only infants exposed to a variable vowel context next to the locus of the relevant consonantal contrast showed evidence for learning. Variability in the final consonant, non-proximal to the location of the crucial contrast, had no such effect. We cannot exclude that the variability provided by the final consonant was not beneficial to the infants, as previous research has shown that final consonants may be more difficult for infants to discriminate than initial ones (Archer et al., 2016). Nonetheless, the difference in the results between our two conditions suggests that the beneficial effect observed in the Vowel condition cannot be reduced to the repeated exposure to forms that exhibit the initial place of articulation contrast during familiarization. If this were the case, a beneficial effect should have also occurred in the Consonant condition.

The effect of vowel variability is consistent with broad evidence, across sensory systems from all modalities, about the context-dependency and the sensitivity of perception to change. In a CVC form, it is in the initial CV part (not the final VC part) of the signal where information about the identity of the initial consonant lies and presenting consonants in different vowel contexts enhances the distinction between these consonants. Such presentation of two consonants across different vocalic contexts in our experiments helped infants identify the contrast between the initial consonants, which then transferred to an ability for better detecting the mismatch between the label presented in the mismatch trial and the label learned for the object during habituation.

A further relevant finding of the current study is that pre-exposure to minimal pairs can indeed have a beneficial effect on sensitizing infants to the pertinent sound contrasts supporting different lexical items. Recall that some approaches that have considered lexical development to be crucial for the formation of phonological categories propose that words with less phonological overlap than minimal pairs are specifically useful for the detection of the phonologically relevant contrasts (e.g., Feldman et al., 2013). The results from Thiessen (2007, 2011), who did not find beneficial effects by familiarizing infants with minimal pairs but only by familiarizing with forms with less phonological overlap, were consistent with this approach. The results of our present study highlight that beneficial effects can also be obtained by pre-exposing infants with minimal pairs and not only with items that are phonologically similar but overlap less than a minimal pair (Thiessen, 2011). However, our results also indicate the crucial qualification that not all types of minimal pairs seem to be effective to the same degree, as indicated by the different effects of the two familiarization conditions in our study.

Overall, our study adds to the findings that acoustic properties of the speech signal modulate infants' ability to recognize the contrastive sound dimensions during word learning. These findings go beyond previous research by showing that exposure to contextual vocalic variability, where the formant transitions contain valuable information about the place of articulation of the preceding consonant, enhances infants' sensitivity to this dimension in word recognition. Although the acoustic stimuli were not presented together with any potential referents during the familiarization, our findings may also support the assumption that lexical neighborhoods promote infants' phonological development. Considered in this way, our study points to close interactions of the developing sound system with the developing lexicon as, for instance, suggested by Swingley (2009) or Werker and Curtin (2005). Further studies are needed to investigate which specific phonetic properties of speech support infants' detection of contrasts and which do not or do so to a lesser extent and to complement our understanding of the roles that the emerging lexicon and input-driven mechanisms play in the formation of phonological categories in early first language development.

Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Projektnummer 317633480–SFB 1287, Project C03.

Mixed model Formula and output:

Linear mixed model fit by REML ['lmerMod']

Formula: look∼1 + trial * fam + (1 | id)

 Data: dat

REML criterion at convergence: 2276.1

Scaled residuals:

   Min     1Q  Median    3Q   Max

−1.63507 −0.56771 −0.04903 0.58275 2.34355

Random effects:

 Groups  Name      Variance  Std. Dev.

 id    (Intercept)  7467737  2733

 Residual        17888092  4229

Number of obs: 120, groups: id, 40

Fixed effects:

             Estimate Std.  Error  t  value

(Intercept)         5784.4    796.2    7.265

trialMismatch        4452.4    945.7    4.708

trialNovel          10317.5    945.7    10.910

fam2 − 1            1021.7    1592.4    0.642

trialMismatch:fam2 − 1   −4429.7    1891.5    −2.342

trialNovel:fam2 − 1     −2429.5    1891.5    –1.284