Lexically conditioned phonetic variation in motherese: age-of-acquisition and other word-specific factors in infant- and adult-directed speech

1.1 Phonetic variation in infant-directed speech

Speech directed toward infants is characterized by differences in pronunciation relative to speech to adults. Overall, mothers talking to their infants tend to produce slower rates of speech (Fernald 1992), higher and broader pitch ranges (Fernald 1984; Smith and Trainor 2008), hyperaticulated vowels (Kuhl et al. 1997, Kuhl et al. 2008; Rattanasone et al. 2013), and more canonical allophonic consonant variants (Dilley et al. 2014), compared to speech directed toward adults. Such infant-directed effects have been found in a number of languages (Ferguson 1964; Fernald et al. 1989), though the details of the effects are not always identical (e.g., Benders 2013; Igarashi et al. 2013).

The speech adjustments observed in infant-directed speech are often assumed to be made for the ‘benefit’ of a specific kind of interlocutor – an infant language learner – though the exact nature of this benefit is not entirely agreed upon. IDS modifications, particularly prosodic effects such as higher F₀ and expanded pitch range (Fernald 1984), seem to serve to regulate the attention of the infant (Fernald 1991, Fernald 1993; Trainor and Desjardins 2002), as well as to communicate speaker affect (Bryant and Barrett 2007). Indeed, it has been demonstrated that speech exhibiting these infant-directed speech properties is preferred by infants over adult-directed speech (Fernald 1985; Fernald and Kuhl 1987). (See, e.g., Cristia 2013; Soderstrom 2007 for reviews of IDS properties.)

Some researchers have suggested that IDS modifications occur to promote language acquisition (Werker et al. 1994), for example by highlighting phonetic parameters that carry phonemic distinctions in a language (Kuhl et al. 1997, Kuhl et al. 2008) or by facilitating word recognition and word segmentation by infants (Thiessen et al. 2005; Singh et al. 2009; Song et al. 2010). In any case, the stronger attention that infants pay to IDS could increase the opportunity for the features of IDS to influence language learning, even beyond what might be expected from the relative frequency with which infants hear this type of speech (de Boer 2005). And it has been demonstrated computationally, in some studies at least, that IDS is more learnable than ADS, at least as far as the identification of vowel qualities is concerned (de Boer and Kuhl 2003; de Boer 2005).

Other researchers have suggested that the phonetic enhancement features of IDS in particular are not goal-directed at all, but are rather a by-product of the prosody and rate features of IDS (e.g., McMurray et al. 2013; Martin et al. 2014). For instance, because IDS is typically slower with more stressed monosyllabic words, it may simply result in less formant undershoot and greater prosodically-conditioned hyperarticulation (McMurray et al. 2013). So, while some features of IDS may be produced for the benefit of the infant, others may not be a specifically encoded part of the signal at all. In fact, other recent research even finds evidence for non-hyperarticulatory patterns in IDS. For example, VOT increased for both voiced and voiceless stops in infant-directed speech, such that the voicing contrast is not actually enhanced (McMurray et al. 2013). Furthermore, although the vowels /i/, /a/, and /u/ may be more peripheral in IDS (Kuhl et al. 1997), other vowels are not (McMurray et al. 2013; Cristia and Seidl 2014); and even the vowels reported to be hyperarticulated show broader distributions and greater overlap in IDS (Kirchhoff and Schimmel 2005; Soderstrom 2007; Cristia and Seidl 2014; Martin et al. 2014). This variability results in a tendency for contrasts in IDS to be less clear and to yield worse performance for automatic speech recognition when algorithms are trained on IDS input (Kirchhoff and Schimmel 2005; Martin et al. 2015).

However, whether infant-directed speech effects promote, intentionally or not, language acquisition, or whether they are simply a coincidence, they are part of a listener-specific speech style. Speech to all kinds of listeners (e.g., listeners in a noisy environment, hearing-impaired adults, non-native listeners) includes specialized features, many of which overlap with the features of IDS. Such speech is known to enhance clarity or intelligibility, without any didactic intent. (See a brief review of such ‘clear speech’ effects below.) It would be somewhat surprising, in fact, if speech to infants did not attend to clarity in some manner as well.

1.2 Clear speech

As just noted, in laboratory speech more generally and in speech directed toward (particular groups of) adult interlocutors, there is also systematic variation in acoustic features of pronunciation, which is generally assumed to occur for the benefit of a listener. Communicative context leads speakers to adjust properties of their speech. For example, speakers talk louder and slower in noisy environments (Lombard 1911; Lane and Tranel 1971). Speech directed toward listeners with a hearing impairment is louder and slower (Picheny et al. 1986; Bradlow et al. 2003). In fact, explicitly instructing talkers to speak ‘clearly’, relative to ‘conversational’ or ‘citation’ speech, produces a range of acoustic-phonetic adjustments, including increased intensity (Picheny et al. 1986), longer segment durations and slower speech rate (e.g., Picheny et al. 1986; Bradlow et al. 2003), larger F₀ range (Bradlow et al. 2003), and enhancement of higher frequency (>1000 Hz) components in the long-term spectrum (Krause and Braida 2003). Prior studies have also shown more segmentally-focused spectral effects in clear speech in the form of vowel hyperarticulation or vowel space expansion (Picheny et al. 1986; Moon and Lindblom 1989, Moon and Lindblom 1994; Bradlow et al. 2003; Krause and Braida 2003; Scarborough and Zellou 2013), as well as maintained or increased consonant-to-vowel coarticulation (Mathies et al. 2001; Bradlow 2002; Scarborough and Zellou 2013). Indeed, these clear speech modifications yield speech that is more intelligible than ‘conversational’ speech (Picheny et al. 1986; Bradlow et al. 1996).

Context also conditions speech adjustments at the level of individual words. Less predictable words are produced more clearly than highly predictable words (Lieberman 1963; Scarborough 2010). Similarly, the first mention of a word in a narrative is produced more clearly than the second mention of that word (Fowler and Housum 1987). Thus, factors that influence the predictability of a word within a communicative context condition speech clarity at the word level. In other words, clarity is increased where listeners might be less able to rely on inferences from context to provide top-down cues to aid them in lexical perception.

1.3 Lexical factors and clarity

Other word-level factors, in particular lexical frequency and phonological neighborhood density, have been shown to have systematic effects on phonetic realization in adult-directed speech as well. With respect to lexical frequency, it has been shown repeatedly that high-frequency words are reduced relative to low-frequency words (e.g., Zipf 1935): they are produced with temporal reduction (Fidelholz 1975; Jurafsky et al. 2000; Bell et al. 2009), more contracted vowel spaces (Munson and Solomon 2004), and greater final segment deletion (Raymond et al. 2006). Phonological neighborhood density is a measure of the number of phonologically similar words (‘neighbors’) there are for a given word (Luce et al. 1990). Words with more neighbors, i.e., those from dense phonological neighborhoods, exhibit more hyperarticulated vowel spaces (Wright 2003; Munson and Solomon 2004), as well as a greater degree of nasal and vowel-to-vowel coarticulation (Scarborough 2013), than words from sparser neighborhoods.

These word-level properties have systematic influences on lexical perception as well. Specifically, high-frequency and low-ND words are perceived more quickly and accurately than low-frequency and high-ND words across a range of tasks (Howes 1957; Luce 1986; Goldinger et al. 1989; Vitevitch 1997, Vitevitch 2002; Vitevitch and Luce 1998; Luce and Pisoni 1998), suggesting a processing advantage for more frequently occurring words and words with more unique phonological structures. Such effects can be understood in terms of the process of spoken word recognition, which involves picking out a target word from among its competitors in the lexicon, all of whose lexical representations are simultaneously activated by an acoustic-phonetic input (Luce 1986; Marslen-Wilson and Zwitserlood 1989; Luce and Pisoni 1998; Norris et al. 2000). More frequent words have stronger representations, due to previous repeated activation, and are therefore more likely to be correctly picked than their less frequent competitors. And words from sparser neighborhoods are simply subject to less competition, and are therefore more likely to be correctly picked.

An explicit connection between lexically conditioned production patterns and perceptual difficulty in the low-frequency and high-ND words has been made by some researchers (though cf. Gahl et al. 2012). For instance, the hyperarticulation in high-ND words has been interpreted as serving to mitigate lexical difficulty, since the hyperarticulated vowels would render the high-ND words more distinguishable from their competitors (Wright 2003; Scarborough 2013; Scarborough and Zellou 2013). The increase in coarticulation that co-occurs with increased hyperarticulation in high ND words suggests that the coarticulation too may serve to improve the perceptibility of these words (Scarborough 2013). Since coarticulation is the overlapping production of discrete sounds, it is a source of robust, if redundant, information about the phonemes present in a word (Beddor 2009). For example, the nasalized vowel [ʌ̃] in ‘bun’ provides information about the upcoming nasal consonant, even before that segment occurs. Such information can improve word perception by delivering cues to two phonemes at once and thus allowing early prediction of the nasal (Ali et al. 1971; Lahiri and Marslen-Wilson 1991).

Indeed, recent empirical evidence indicates that both hyperarticulation and increased nasal coarticulation facilitate lexical perception. Speech exhibiting hyperarticulated vowels has been shown to be more intelligible than speech with less hyperarticulated vowels (Bradlow et al. 1996). Furthermore, words produced with hyperarticulation and the greatest increased nasal coarticulation, as found in listener-directed clear speech, are perceived more quickly than words produced with hyperarticulation but decreased nasal coarticulation (Scarborough and Zellou 2013). Also, earlier onset of anticipatory nasality in a vowel preceding a nasal consonant speeds lexical identification (Beddor et al. 2013). These findings suggest that that modifications made by speakers in clear speech, including increasing degree of nasal coarticulation and/or increasing vowel hyperarticulation, enhance the intelligibility of the speech signal (intentionally or not).

1.4 Lexical factors in infant-directed speech

If lexical factors like neighborhood density and word frequency are factors that condition accommodative adjustments in lab speech or speech directed to adults, we might predict that they should have similar effects in infant-directed speech, which is fairly explicitly accommodative. However, an infant lexicon is quite different from an adult lexicon – both in size and in organization. Not only do infants know fewer words, but they also have less experience with each word, and they know less about relations between words. Thus, neither lexical frequency nor neighborhood density, as calculated for adults, is a good representation of lexical difficulty for infants. Thus, if frequency and neighborhood density condition effects in speech production because they represent lexical difficulty for a listener that speakers try to compensate for, they would seem not to be relevant factors in IDS after all. Rather, a more age-appropriate lexical measure should be a better basis for lexical adjustments in infant-directed speech.

One such alternate approach might be to consider frequency and neighborhood density calculated over child speech. In fact, such lexical calculations have been made in work by Storkel and colleagues (e.g., Storkel and Hoover 2010, Storkel and Hoover 2011; Storkel et al. 2013), based on corpora of speech from kindergarteners and first-graders. However, even such norms are not fully appropriate for the current study, which investigates infant interlocutors (from whom speech corpora are obviously not available). Furthermore, although child-derived lexical statistics might be very useful in explaining child productions, word-learning in children, or lexical perception by children, they are possibly not relevant in the domain of speech produced by adults, even if produced to infants, since it is not clear how adults would have access to statistics derived from a real child lexicon in any case.

Lexical age-of-acquisition (AoA), or the adult-reported age at which a word was learned (Caroll and White 1973; Gilhooly and Logie 1981), is a better candidate for an appropriate measure of lexical difficulty for infants. First, such ratings make explicit reference to the developing lexicon, and insofar as children acquire easier words earlier, they refer directly to lexical difficulty for children. Further, the fact that AoA is based on subjective ratings from adults (Kuperman et al. 2012) means that AoA represents an adult’s assessment of how hard a given word might be for a child – precisely the sort of information that is accessible to an adult and that could influence the adult’s production to a child word-by-word.

Even though AoA measures are subjective ratings by adults, numerous studies report evidence that these measures do indeed reflect lexical difficulty for young children, both in producing and perceiving words. For example, elementary school-age children are faster at producing words with a lower AoA, both in picture-naming (Gilhooly and Gilhooly 1979; Gilhooly and Watson 1981) and word repetition (Garlock et al. 2001). Likewise, earlier AoA accurately predicts faster spoken word recognition (auditory gating task: Garlock et al. 2001; and auditory lexical decision: Cirrin 1984) and more accurate mispronunciation detection (Walley and Metsala 1992) in elementary school-age children. (Note that similar perceptual effects of AoA from word naming and lexical decision have been demonstrated for adults, as well; Brown and Watson 1987; Izura et al. 2011; Kuperman et al. 2012; inter alia). These effects have been explained as effects of lifetime cumulative frequency or a result of order of acquisition on representational activation (Kuperman et al. 2012). But since AoA measures are inherently subjective, they might also reflect frequency, ND, or any number of other linguistic, cognitive, or social factors (e.g., word length, semantic content, concept complexity, observational experience, etc.) that contribute to an adult’s evaluation of the difficulty of a given word in acquisition.

1.5 Current study

In this study, then, we investigate lexically conditioned variation in infant-directed speech, comparing it to lexically conditioned variation in adult-directed speech. The goal of the language-learning infant is not just to acquire the phonological inventory of the ambient language but also its lexicon, just as the goal of linguistic communication is to express not just sounds, but meaningful words (and phrases). Thus, we predict that infant-directed speech, like adult-directed speech, should contain lexically conditioned phonetic variation.

To that end, we investigate phonetic variation as a function of lexical factors in spontaneous infant-directed speech. To address the questions of (1) whether the lexically conditioned patterns of production in infant-directed speech, if they occur, are similar to those found in adult-directed speech, and (2) whether the factors that condition any such modifications are special to infant lexicons (or to assumed or deduced infant lexicons), we compare the patterns in infant-directed speech to those in spontaneous adult-directed speech. We consider these patterns in light of their hypothesized goal of enhancing the perceptibility of the speech signal for a given interlocutor. Thus, we predict specifically that lexical age-of-acquisition will condition phonetic variation in infant-directed speech, above and beyond any contribution of ND or lexical frequency, which condition variation in adult-directed speech.

2.1 Infant-directed speech

The infant-directed speech data for this study come from the Brent corpus (Brent and Siskind 2001). This corpus is a collection of spontaneous speech recordings of 16 English-speaking mothers talking to their preverbal infants, ranging in age from 9 to 15 months. The speakers were recruited and recorded in and around Baltimore, Maryland. Recording was done in the speakers’ homes, without an experimenter present, over more than a dozen sessions of 1.5–2 hours each. The middle 75 minutes of each session were extracted and transcribed by Brent and colleagues. (See Brent and Siskind 2001 for more detailed information about the data-collection methods for this corpus.) We automatically force-aligned the sound files to the transcripts at both the word and phoneme level using the FAVE-align component of the Forced Alignment and Vowel Extraction (FAVE) suite (Rosenfelder et al. 2011), based on the Penn Forced Aligner (Yuan and Liberman 2008).

2.2 Target words

The current study focuses on monosyllabic, monomorphemic content words containing exactly one nasal segment (either a pre-vocalic nasal segment (NV), e.g., nap, or a post-vocalic nasal segment (VN), e.g., hand) extracted from the recordings of these 16 mothers talking to their infants. As is discussed in Section 2.3.1, high vowels elude accurate acoustic nasality measurement, so words containing high vowels were excluded. The full data set consists of 109 unique word types containing a nasal segment, represented by 8,127 tokens. The average number of tokens per mother was 507 tokens, with a range of 179–1,135.

2.3 Acoustic measurements

2.4 Lexical variables

2.5 Model design

3.1 Hyperarticulation

The estimated coefficients and associated p-values from the model on hyperarticulation in IDS are presented in Table 2.

The model of hyperarticulation in IDS includes a significant main effect of AoA: with a positive coefficient, this effect indicates that words learned later have a greater degree of hyperarticulation. Word frequency was a marginally significant predictor of hyperarticulation: higher frequency words tended to be less hyperarticulated than lower frequency words. Neighborhood density was not a significant predictor of hyperarticulation in this IDS data set. There were no significant interactions among the three lexical predictors in the IDS hyperarticulation dataset.

The isolated univariate effects of (non-residualized) AoA and Frequency on hyperarticulation in IDS are illustrated in Figure 2: vowels from words that are reported as later-acquired and from words that are less frequent are more hyperarticulated than vowels from words that are reported to be learned earlier and words that are more frequent.

Recall that prior studies that have reported hyperarticulation in IDS as a register have focused on point vowels (/i/, /a/, /u/) (e.g., Kuhl et al. 1997, Kuhl et al. 2008; Cristia and Seidl 2014), and in fact, hyperarticulation may even be limited to these point vowels (McMurray et al. 2013). Although the current study does not address the presence or absence of hyperarticulation in IDS in general, we wanted to ask whether the AoA effects on hyperarticulation in IDS might also be limited to this subset of vowels. Since the current dataset examines only non-high vowels, /a/ is the only point vowel. Therefore, we ran a post-hoc model on a subset of the data excluding /a/ (i.e., on the non-point vowels). This non-point vowel model also revealed a significant effect of AoA (est=0.24, SE=0.14, p=0.008), a non-significant effect of ND (p=0.5), as well as a significant effect of frequency (est=−0.21, SE=0.17 p=0.03). Thus, it appears that the AoA-conditioned hyperarticulation generalizes across vowels in our data.

3.2 Nasal coarticulation

The estimated coefficients and associated p-values from the model on nasal coarticulation in IDS are presented in Table 3.

The model of nasal coarticulation in IDS includes a significant negative coefficient for AOA: later-acquired words have a lower A1-P0, which means they have greater nasalization. Word frequency was also a significant predictor of degree of nasality: higher-frequency words had less nasality than lower-frequency words. The model on nasality in IDS did not yield a significant main effect of neighborhood density. ^[2]

The isolated univariate effects of (non-residualized) AoA and Frequency on degree of nasal coarticulation in IDS are illustrated in Figure 3: words that are reported as later acquired and words that are less frequent are produced with a greater degree of vowel nasality (lower A1-P0 dB) than earlier-learned and more-frequent words. ^[3]

3.3 Interim discussion

The two models on vowel hyperarticulation and vowel nasality yielded systematic lexical AoA effects. In spontaneous speech to infants, mothers produced greater degrees of vowel hyperarticulation and nasal coarticulation in words that were subjectively rated as later-acquired, above and beyond other lexical effects. Meanwhile, ND had no demonstrable effect on these phonetic variables. Recall that prior studies have found evidence of consistent ND effects on hyperarticulation and coarticulation in laboratory speech, elicited as standard citation form speech or as explicitly (adult) listener-directed (e.g., Wright 2003; Munson and Solomon 2004; Scarborough 2013; Scarborough and Zellou 2013).

In order to understand the relationship between the current findings and those from prior studies, two possible explanations are considered: (1) Context matters in conditioning lexical variation; that is, AoA influences phonetic variation in IDS since that is the best metric for lexical difficulty for language learners; meanwhile, ND conditions production patterns in adult-directed speech because that is the best metric for lexical difficulty for interlocutors with fully-developed lexicons. Or, (2) AoA might condition lexical variation in all speech styles, for reasons related, for instance, to representational strength or cumulative frequency that this measure might represent, and prior studies of ADS did not uncover its influence since it had not previously been included as a predictor of speech production. These two possibilities can be addressed empirically via investigation of which lexical factors are significant in a model of adult-directed speech, with the same predictors as in the present IDS study. If the first explanation is correct, we would expect a non-significant effect of AoA and a significant effect of ND on conditioning phonetic variation in adult-directed speech. If the second explanation is correct, we would expect a significant AoA effect in the adult-directed speech, just as in the infant-directed speech. In the next section, we test these two hypotheses on a set of data from a spontaneous adult-directed speech corpus.

4.1 Buckeye corpus of adult-directed speech

The spontaneous adult-directed speech data came from the Buckeye corpus (Pitt et al. 2005, Pitt et al. 2007). The entire corpus consists of spontaneous speech directed toward an adult interviewer from 40 speakers from Columbus, Ohio, each lasting approximately one hour (see Pitt et al. 2005 for more information about this corpus). In order to construct a sample of speakers comparable to our sample of IDS speakers, we analyzed a subset of this corpus – all female speakers who were under the age of 30 – resulting in 10 individuals’ recordings.

The sound files were automatically forced-aligned to the transcripts at both the word and phoneme level using FAVE-align. As in the IDS study, monosyllabic, monomorphemic content words containing exactly one nasal segment were extracted from these recordings. The ADS data set consists of 88 unique word types containing a nasal segment, represented by 897 tokens (with 77 additional tokens of nasal words with /i/ in order to calculate vowel space centers for each subject using /a/ and /i/ means; these 77 tokens were not independently included in the analysis). Hyperarticulation and degree of nasality (as A1-P0) were measured and calculated for each token using the same methods described in Section 2.3.

Means, standard deviations, and ranges for Frequency, ND, and AoA (all non-centered) from the 88 word types containing exactly one nasal segment extracted from the subset of 10 young female speakers selected from Buckeye corpus are provided in Table 4. Note that the words are strikingly similar to the words from the IDS part of the study with respect to their lexical statistics.

Two lmer models were run on the hyperarticulation and nasal coarticulation data from the Buckeye ADS speech in the same manner as for the Brent data: fixed effects of centered Frequency, centered ND (residualized by frequency) and centered AoA (partialized by Frequency and residuals of ND), along with two- and three-way interaction terms. ^[4] Random intercepts of Speaker and Item were included as well.

4.2 Results

Two linear mixed models analyzed hyperarticulation and nasal coarticulation measured from 897 tokens (88 unique word types containing a non-high vowel and a nasal consonant) from 10 young women talking to an adult interlocutor.

The estimated coefficients and associated p-values from the models on hyperarticulation and nasality in ADS are presented in Tables 5 and 6, respectively.

In both models, ND is a significant predictor of phonetic variation: high ND words are more hyperarticulated and have a greater degree of nasality (=lower A1-P0). AoA is not a significant main predictor of either of these phonetic variables in the ADS data.

We note that there are several significant interactions present in the two models of ADS. First, in both the model on hyperarticulation and the model on nasal coarticulation, there is a significant interaction of ND and Frequency. In order to interpret this interaction, we binned the word types extracted for the adult-directed speech data set into categories split by the mean of each sample (i.e., high/low frequency, high/low ND, high/low AoA). Figure 4 illustrates the significant ND by frequency interaction in each model, with means (and standard errors) of nasal coarticulation and hyperarticulation in each of the sub-categories. It appears that the interaction is driven by differentiation of high- and low-frequency words in high-ND words, but not in low-ND words. In other words, the effect of phonetic modification was most extreme in the most difficult adult word types – high-ND, low-frequency words.

These patterns were confirmed in a series of post-hoc linear mixed models on subsets of the data. First, for hyperarticulation, two post-hoc lmer regressions were run on the ND-binned data – one for higher-ND words and one for lower-ND words – with Frequency set as a continuous fixed predictor and random intercepts for Speaker and Item. The post-hoc regression for higher-ND words showed a significant correlation with Frequency (est.=−0.19, SE=0.09, t=−2), while the post-hoc regression for lower-ND words did not show a significant frequency effect (t=0.97). For nasal coarticulation, frequency again correlated with acoustic nasality in higher-ND words (est.=1.6, SE=0.94, t=1.8), but not in lower-ND words (t=1.1).

While AoA was not a significant main effect in either of the ADS models, AoA did participate in a significant two-way interaction with ND and a three-way interaction with ND and Frequency in the hyperarticulation model (but not in the nasal coarticulation model). The two-way interaction, illustrated with word types binned by high/low ND and high/low AoA, is shown in Figure 5.

As can be seen in Figure 5, the significant interaction appears to be driven by the (small) difference between early and late age of acquisition in high ND words, while there is no AoA-differentiation in hyperarticulation for low-ND words. Like with the ND by Frequency interaction, the effect of phonetic adjustment for AoA only emerges in the more difficult high-ND words. To confirm this effect, two post-hoc lmer regressions were run on the ND-binned data – one for higher-ND words and one for lower-ND words – with AoA set as a continuous fixed predictor and random intercepts for Speaker and Item. The post-hoc regression for higher-ND words revealed a significant effect of AoA (est.=0.22, SE=0.06, t=3.5), while the post-hoc regression for lower-ND words did not have a significant AoA effect (est.=0.006, SE=0.35, t=0.12).