Perceptual sensitivity to vowel quality and vowel length in the first year of life

There is little doubt that vowels play a pivotal role in the early stages of language development. At birth and some time before that, infants most probably perceive speech as a stream of tonal and loud sounds (vowels) alternating with noisier and quieter intervals (consonants). Since vowels represent the more salient parts of the speech signal, due to their relatively high intensity and long duration, and since prenatally the acoustic features of vowels are most probably better available than those of consonants, infants are assumed to have an initial preferential focus on vowels rather than consonants.¹ Vocalic portions of the speech signal display language-specific melodic and rhythmical patterns, as well as patterns characteristic of individual speakers, and their intentions or affective states. Knowledge of these patterns starts to be learnt prenatally, allowing infants already at birth, and perhaps even earlier, to recognize their mother's voice or rhymes and songs that she sang during pregnancy.² From about the 35th week of gestational development, humans discriminate spectral information at least up to 500 Hz, differentiating between the vowels /a/ and /i/,³ and soon after birth the perceptual processing of vowels seems to display language-specific traits.⁴ In contrast to spectral information, only partially available prenatally, the durational properties of the speech signal are fully available to fetuses,⁵ allowing them to learn not only the suprasegmental (rhythmical) features of the ambient language but hypothetically also temporal properties of individual speech-sound categories, such as those constituting vocalic length (also termed quantity) contrasts.

The developmental precedence of vowels over consonants is evident throughout the first year of an infant's life. For some vowels, the formation of phonological categories has been evidenced for the age of about 6 months,⁶ but it is later for consonants.⁷ However, the class of vowels itself comprises phonological contrasts based on phonetic properties differing in their acoustic salience. Since salience seems to play an important role in early language development, one can expect more salient vowel contrasts to be acquired earlier than less salient ones.^8–10 In all languages of the world, there are spectrally based vowel contrasts, with the values of the first two to four formants defining their quality. In some languages (e.g., Czech, Finnish, or Japanese), there are also vowel contrasts cued (primarily or even exclusively) by the acoustic property of duration, giving rise to phonological length distinctions between “short” and “long” vowels. Most infant studies interested in the acquisition of native vowels focused on spectral contrasts,^11–13 while fewer examined length contrasts.^14–16 Research on the acquisition of duration-cued contrasts is relatively scarce, and the findings are inconclusive. In behavioral experiments, infants acquiring English, in which duration is one of the cues (but not the primary one) to vowel identity, seem able to use duration contrastively from the age of 5 months.^15,17 On the other hand, behavioral research with Japanese infants,¹⁶ whose language—similarly to Czech—uses vowel duration as the primary cue to short-long phonemes, reports duration-based vowel discrimination only at 10 months of age (even though at 4 months these infants discriminate the maximally distinct and universal [a]–[i] spectral difference).

At the neural level, however, Japanese infants discriminate native length distinctions already at 3–4 months.¹⁴ An early sensitivity to length has been reported by a recent electroencephalography study¹⁸ also for infants acquiring Finnish (which, like Japanese or Czech, is a quantity language), who at birth had a reliable neural mismatch response to both a vowel length change in [a] vs [aː] and a vowel quality change in [a] vs [o]. German-learning infants at 2 months of age, too, show robust neural processing of durational changes in [ba] vs [baː] stimuli.¹⁹ How German- or Japanese-exposed infants pre-attentively process durational changes compared to spectral changes was not addressed in those studies.

In summary, the findings of the relatively few behavioral and neurolinguistic studies that, to date, focused on infants' perceptual processing of vowel duration do not provide conclusive evidence as to the age at which vowel length categories are acquired and particularly how the development of phonemic vowel length relates to that of vowel quality. It is not known whether 6 months is the age of acquisition of vowel quality for all infants, i.e., including those whose language employs contrastive length, or whether the exposure to both vowel quality and vowel length delays the development of either of the vocalic contrasts or both. In fact, at least one study documents that infants learning a language that has both vowel length and quality (Finnish) show linguistic processing of vowel quality at 12 but not at 6 months.²³ 

We thus aim to test whether quantity and quality contrasts are acquired at different moments in development. We predict that younger infants will be more sensitive to a vowel-length than to a vowel-quality contrast because duration, unlike spectral structure, is fully audible from an earlier age, i.e., already prenatally.⁵ This means that the perceptual sensitization to vowel length can start earlier than sensitization to vowel quality (see also Burnham's¹⁰ proposal that the salience of duration-cued contrasts predicts their early acquisition). If infants acquiring a quantity language develop spectral phoneme categories at the same age as infants from non-quantity languages, i.e., at 6 months,⁶ then the early advantage of vowel length over vowel quality may fade at around that time. However, based on Ref. 23, we expect that the early advantage of vowel length would disappear at an older age than 6 months, i.e., at the 10th month or perhaps even later.

In order to investigate the developmental path of spectral and length contrasts' acquisition, the present study employs a habituation-dishabituation paradigm. Infants' perceptual sensitivity to a native quantity and a quality contrast is assessed at the age of 4 months (before first vowel categories are typically formed⁶), 6 months (when first vowel categories are acquired⁶), 8 months (when a decline of sensitivity was found for some vowel contrasts^12,13), and 10 months (when most segmental categories are acquired, including vowel length in Japanese infants¹⁶). The time frame between 4 and 10 months of age would make it feasible to perform a longitudinal study, with the attractive possibility of referencing each child's experiment results to their own previous performance. However, it has been shown that infant familiarity or novelty preference in experiments may be affected by previous experience with experimental procedures.²⁴ To avoid this serious confound, we chose to use the cross-sectional design so that each child participated in an experiment for the first time. To avoid any potential effects of infants' varying prosodic or lexical development, we employ as stimuli simple consonant-vowel syllables. The tested infants are acquiring Czech, a language in whose vocalic system five contrasting vowel qualities specified in terms of height and backness combine with two degrees of length.^20,22

Both types of contrasts tested in the present study are native, which means that, throughout exposure to the ambient language, infants will attune their perception to both, although perhaps at different points for development for each contrast.

A total of 82 monolingual infants acquiring Czech, aged 4–10 months, participated in the experiment, and it was the first lab visit for all of them. They were divided into 4 age groups as follows: 21 four-month-olds (mean age 123 days, range 106–142), 15 six-month-olds (mean age 183 days, range 168–197), 26 eight-month-olds (mean age 245 days, range 227–257), and 20 ten-month-olds (mean age 303 days, range 290–317). All infants were born full-term, passed the neonatal hearing test, and did not have any hearing impairment as per parental report. The infants did not attend daycare and their parents were high-school or university educated, living in the urban area of the Czech capital. Fifty-nine additional infants were tested but excluded for the following reasons: seven failed to complete the habituation phase, 49 infants did not pass the habituation criterion, two other infants did not complete the first test round, and one additional infant turned out to be above the target age range (13 months old).

The stimuli were three consonant-vowel (CV) syllable tokens, [fa], [fɛ], and [fɛː]. The vowels /ɛ/ and /a/ were chosen for the experiment because they belong to the most frequent Czech vowels. Long vowels are less frequent than short ones; the long /ɛː/ was chosen because /ɛ/–/ɛː/ is a contrast that occurs in native Czech vocabulary, and it is cued almost exclusively by duration.^20,22 The phoneme frequencies were estimated from written corpora and impressionistic evaluation of infant-directed speech as no data on phoneme distribution in Czech IDS are currently available (but are being collected in our lab). None of the three CV combinations used in the experiment carries lexical content in Czech. All three stimuli were obtained from a single source recording of a hyperarticulated [fa] token produced naturally by a native Czech female speaker by editing in Praat²⁵ the vowel duration (using the overlap-add method) and spectral quality (using the method described in Praat's manual under “Source-filter synthesis 4. Using existing sounds,” additionally preserving the original intensity of the vowel and the original high-frequency resonances). The resulting [fa] stimulus had the first three formants shifted to 864 Hz, 1287 Hz, and 2831 Hz, and the duration to 220 ms; for the [fɛ] stimulus, the formants were shifted to 755 Hz, 1646 Hz, and 2710 Hz, and duration again to 220 ms; the [fɛː] stimulus had the same formant frequencies as [fɛ] but its duration was changed to 360 ms. These resulting stimuli were good exemplars of isolated Czech /fa/, /fɛ/, and /fɛː/ syllables, as assessed by the vowels' acoustic properties^20,21 as well as the impressionistic judgment by three Czech phoneticians.

The three auditory stimuli were used to assemble alternating/non-alternating trials, each containing 12 CV syllables with the onsets aligned at 1.09-s intervals. The non-alternating trials contained 12 repetitions of [fɛ]; these trials served as habituation trials as well as control test trials. In the alternating trials, [fɛ] switched 6 times with [fa] or with [fɛː], serving as vowel quality and vowel length test trials, respectively. The total trial length was 12.3 s.

Infants were tested in a central fixation, habituation-dishabituation paradigm. Each infant was seated on their parent's lap facing a 19-in. Dell screen (about 1 m away) in a sound-treated booth. Auditory stimuli were played from a PC (Creative Sound Blaster Audigy FX soundcard) at approximately 65 dB SPL through two loudspeakers (Truth B2030A) placed laterally to the monitor. The parent wore silencing headphones (Bose QuietComfort 35 II) with masking music and was instructed not to interact with or stimulate the child. The visual stimulus was a static black-and-white checkerboard during habituation and test trials, and a growing-shrinking colorful balloon as an attention-getter.

The session began with a habituation phase whose length was contingent on infant looking. After reaching the habituation criterion, i.e., after their looking time (averaged over two successive trials) decreased to 65% or less of their longest looking time (averaged over two successive trials), the session continued into the test phase. The habituation phase comprised at least 4 trials (i.e., in case the average looking time across the 3rd and 4th trial already dropped below 65% of the average looking time across the first two trials), and at most 14 habituation trials (pilot testing indicated that 14 trials is the ideal maximum, considering the large span of infant ages tested, and is comparable to the maximum in, e.g., Ref. 11). Infants who did not habituate until the 14th trial were excluded from the analyses. The test phase comprised two rounds of three test trials, one of each type, namely a length change, a quality change, and a no-change control trial. The three trials within a test round were presented in pseudo randomized order, which was identical for the infant's second test round. Only infants who completed the first test round (i.e., at least three test trials) were included in the analyses.

An experimental trial was initiated after the infant looked to an immediately preceding attention-getter for 2 consecutive seconds. The duration of the experimental trials themselves was not contingent on infant looking, each trial was played in its full, 12.3-s, length. The experiment was administered in PyHab.²⁶ The experimenter was outside the testing booth, blind to the auditory stimulation. She had a frontal view of the infant's face via video and indicated by button presses whether the infant was looking at the screen. This information was used to automatically determine the initiation of each experimental trial as well as the habituation criterion. The looking-time data obtained from online coding were then used in the analyses.

Total looking times in seconds, per participant and per trial, were analyzed with a linear mixed-effects model (lme4 and lmerTest packages^27,28) in R.²⁹ The predictors were test trial type (two contrasts, comparing the duration change trial and the spectral change trial to the control trial), infant age in months (three contrasts, comparing every younger to the immediately older group), test round (first vs second), and all their two- and three-way interactions. The model included random intercepts for participant and condition (i.e., one of six pseudorandom presentation orders) adjusted for trial type and test round¹. The data and analyses are openly available at https://osf.io/634eh/?view_only=d584be0474004219ab30a6800d65f05b.

Figure 1 shows the distributions of raw response time data for each test trial type in the first and second test round across age groups. Figure 2 shows the model predictions, including a control for the random effects. Inspection of the figures indicates that especially in the second test round the length change elicited longer looking times than the control trial, visibly in the 4-, 6-, and 8-month-olds, while a slightly different pattern can be seen in the 10-month-olds. Table 1 shows the fixed-effects model summary.

The significant intercept shows that the average looking time across all age groups and test trial types was 7.4 s. There were no main effects of trial type, test round, or age, but the analysis yielded a significant interaction of trial type and test round (estimated mean = 0.259 s, SE = 0.126, p = 0.042). To unpack the interaction, we ran a simpler model for each test round separately, with the same two contrasts for trial type and three contrasts for age as predictors, and random intercepts for participant and condition. The analyses did not detect any significant main or interaction effects in the first test round but yielded a main effect of trial type—the length vs control contrast—in the second round, where the length change trial yielded longer looking times than the control trial by an average of 0.435 s (SE = 0.212, t[152.5] = 2.055, p = 0.0416).

Note that the initial full model indicated a triple interaction of trial type 2, test round, and age contrast 3, which was just above the 5% significance level with p = 0.054. However, in neither of the two simpler models per test round did we observe a (nearly) significant interaction of trial type 2 and age contrast 3 (p's > 0.1). Inspection of Figs. 1 and 2 suggests that the initial near-significant effect might be driven by a between-test round decrease in the 10-month-olds' looking times to the spectral change trial. This could be due to a general loss of interest in the task in our oldest group, which is a behavior typically observed in older infants and is suggested by the non-significant but relatively large effect of age contrast 3 (see Table 1 where the estimated mean difference between looking times in the 10-month-olds and 8-month-olds is 0.704 ± 0.385 s, with p = 0.072).

We tested how infants acquiring a language that contrasts vowels both in terms of spectral properties and duration react to native-language phoneme changes in vowel quality and vowel length, and whether their perceptual sensitivity for each type of contrast differs across different ages. The results suggest that infants, in general, across the tested age range of 4 to 10 months, dishabituated to a change in vowel length, while no evidence was found for dishabituation to vowel quality. This indicates that infants acquiring Czech process and respond to changes in vowel length, and that their sensitivity to a change in vowel quality is probably more subtle (if any). A lack of developmental patterns² speaks in favor of an overall greater sensitivity to native vowel-length than to native vowel-quality contrasts in Czech-exposed infants within most of their first year of life. Note, however, that the pattern seen in our oldest group, the 10-month-olds, did not match exactly that of the younger groups (although no significant difference was revealed), which could have been either due to their lower interest in the experiment itself or due to a developmental change that our analysis failed to reliably detect. Perhaps somewhat unexpectedly, the effects of dishabituation found here were observed only in the second round of test trials. This delayed dishabituation effect might be an inherent property of the multiple-change paradigm where the infants' speech processing faculty might need some lag time to reconstruct the various types of phonological categories they have been presented with before they can respond accordingly.

The failure to detect dishabituation to a change in native vowel quality in infants between 4 and 10 months of age is rather surprising given that previous literature reports 6 months as the age of acquisition of native vowel quality contrasts. In other words, the apparent absence of dishabituation to changes in vowel quality even in our oldest groups of infants aged 10 months seems to push the age of acquisition of native vowel quality contrasts well beyond the 6th month as the age of vowel acquisition reported previously.⁶ Such apparent lack of robust vowel quality discrimination even at 10 months has two potential explanations. One is that infants exposed to a quantity language may first acquire vowel length and only then establish adult-like categories for the individual vowel qualities (which is delayed compared to infants learning a language without phonemic vowel length). If that is the case, native vowel contrasts are not universally acquired by about 6 months of age (as has been argued so far) but rather the age of acquisition varies depending on the type of vowel contrast to which the infant is exposed in their language. This account seems to be in line with the neural discrimination patterns in infants learning Finnish, which has phonemic vowel length. Cheour et al.²³ demonstrated that Finnish infants have reliable neural mismatch responses to both a smaller native and a large non-native vowel quality contrast already at 6 months, but only at 12 months do their responses reflect the linguistic status of the vowels whereby the acoustically smaller native contrast elicits a larger response than the acoustically larger non-native contrast. The possibility that 6 months is the age of vowel acquisition only for infants whose language lacks phonemic vowel length needs to be tested in future work, ideally comparing our findings with the results for infants acquiring a language with similar vowel qualities as Czech but no vowel length contrasts (such as Greek or Spanish). If an early language-specific environment affects the order of acquisition for vowel contrasts, as suggested here, it further emphasizes the role of early input and early experience for language acquisition. It will be important to study whether young infants' discrimination of different types of phonological contrasts is affected by other conditions, such as hearing impairment or risk for language disorders.

An alternative explanation for not finding evidence of vowel quality discrimination even for the oldest group in the present study is that, although the infants tested here had in fact acquired the Czech /ɛ/–/a/ contrast, they failed to manifest sensitivity to it in the present experimental paradigm. The fact that the infants were presented with both the length and the quality change within a single session could have perhaps caused them to selectively focus only on that contrast for which they have more robust representations. It is possible that, had the vowel-quality change been presented alone, the infants would have shown reliable dishabituation. Note that although in most studies to date, infants were tested with a single type of contrast within an experimental session, some previous studies have used the multiple-change paradigm.^30,31 The question to be pursued in the future is whether in a single-vowel-type behavioral paradigm, Czech infants would exhibit sensitivity to vowel quality changes, perhaps as early as at 6 months of age, and to what extent such findings would be informative about the degree of their phonological attainment.

An interesting direction for future research would be to examine Czech (or other quantity-language) infants' processing of vowel length and quality changes at the neural level since behavioral measures may not always uncover developmental changes in infants' underlying linguistic competencies.³² The event-related potential paradigm extending the present behavioral study could implement a multiple-oddball presentation and assess the mismatch response to changes in vowel length and vowel quality as an index of phonetic discrimination at the pre-attentive level.

This research was supported by the Czech Science Foundation Grant No. 18-01799S and by Charles University Grant No. Primus/17/HUM/19. We are grateful to Radka Klimičková, Kristýna Hrdličková, and Veronika Ungrová for assistance with experiment administration.