Impacts of Acoustic-Phonetic Variability on Perceptual Development for Spoken Language: A Review

Abstract

This article reviews research on when acoustic-phonetic variability facilitates, inhibits, or does not impact perceptual development for spoken language, to illuminate mechanisms by which variability aids learning of language sound patterns. We first summarize structures and sources of variability. We next present proposed mechanisms to account for how and why variability impacts learning. Finally, we review effects of variability in the domains of speech-sound category and pattern learning; word-form recognition and word learning; and accent processing. Variability can be helpful, harmful, or neutral depending on the learner’s age and learning objective. Irrelevant variability can facilitate children’s learning, particularly for early learning of words and phonotactic rules. For speech-sound change detection and word-form recognition, children seem either unaffected or impaired by irrelevant variability. At the same time, inclusion of variability in training can aid generalization. Variability between accents may slow learning—but with the longer-term benefits of improved comprehension of multiple accents. By highlighting accent as a form of acoustic-phonetic variability, and considering impacts of dialect prestige on children’s learning, we hope to contribute to a better understanding of how exposure to multiple accents impacts language development and may have implications for literacy development.

Introduction

Acoustic-phonetic variability is rampant in infants’ early language experience, yet infants somehow learn words and categories and patterns of sounds. The question of how we experience constancy despite fluctuation in sensory input stems from at least 500 BCE in remarks by Heraclitus. Heraclitus’ ideas have been interpreted in two ways: Plutarch gleaned the idea that “it is not possible to step twice into the same river” (Graham, 2019), suggesting environmental variability is a challenge to learning and processing. More recent interpretations posit that constancy and change are fundamentally linked and that variability is a precondition for constancy (Graham, 2019; Marcovich, 1966). Tension between variability-as-nuisance and variability-as-fundamental pervades research on spoken-language development. For modern scientists examining learning by humans and machines, lack of invariance is a perpetual challenge. Nonetheless, some recent work suggests acoustic-phonetic variability lends stability to children’s representations of words and sounds.

This article reviews research on when acoustic-phonetic variability facilitates, inhibits, or does not impact perceptual development for spoken language, elucidating mechanisms by which variability helps form robust sound categories. We review different sources and structures of variability; proposed mechanisms for how variability impacts language learning; and effects of variability in the domains of speech-sound category and pattern learning, word-form recognition and word learning, and accent processing. We hope that highlighting accent as a form of acoustic-phonetic variability, and considering impacts of dialect prestige on accent processing, will help illuminate how exposure to multiple accents impacts language development and may impact literacy development.

In the context of recent calls to diversify STEM disciplines (e.g., #ShutDownSTEM; Chen, 2020), we aim to critically evaluate our field, consciously reduce citation bias, and call for increased diversity in this research area, in researchers, topics, and participants. Much of basic research on perceptual development for spoken language in the US has been conducted by and with predominantly White, educated people (Heinrich, Heine, & Norenzayan, 2010) speaking White American English (WAE, also sometimes referred to as Mainstream or Standard American English; Wolfram, 2007; Y. Holt & Rangarathnam, 2018). Research on other dialects often treats WAE as a standard from which other forms are derived, rather than one of multiple variants. This bias limits generalizability and does a disservice to the diverse population who should benefit from research (DeGraffe, 2018; see also DeGraffe, 2005). Yet the limited research on other dialects that does exist has generated crucial insights into basic-science questions about effects of input variability on language learning, which are intimately connected to important applied topics, such as reducing racial disparities in literacy outcomes (Washington & Craig, 2000; Thompson, Craig, & Washington, 2004) and speech-language evaluation (e.g., Moland, 2011). Accordingly, we highlight work by and with those from minoritized groups, including work by scholars identifying as Black, Indigenous, or People of Color (BIPOC), focusing in particular on African American English (AAE). We aim to draw attention to high-quality, relevant research, some of which has been underappreciated and under-cited in proportion to its value.

We also emphasize strongly that future work connecting basic research on perceptual learning for spoken language with topics such as racial disparities in literacy outcomes must center perspectives of children and communities, working in careful partnership with—and ideally led by—communities of color, given their historic exploitation in research (Ndlovu-Gatsheni, 2019; Patel, 2015). The community-based participatory research framework provides a roadmap for such partnerships (Israel, Schulz, Parker, & Becker, 2001; Holkup, Tripp-Reiner, Salois, & Weinert, 2004). Future research should also consciously apply a strengths-based rather than deficits-based lens (Dudley-Marling & Lucas, 2009; Michaels, 2013; Avineri et al., 2015; Baugh, 2017; Cole, 2013; Sperry, Sperry, & Miller, 2019).

Sources and Structures of Variability

Sources of variability.

Acoustic-phonetic variation comes from a variety of sources. One source is contrastive differences between words and sounds (i.e., those the language uses to differentiate meanings), and this type of variability is critical input to distributional learning (Maye, Werker, & Gerken, 2002; Vallabha et al., 2007). For example, sounds like b and p differ in voice-onset time (VOT) and pitch. Many other sources of variability are non-contrastive. Elman and McClelland (1984, 1986) distinguished two types of non-contrastive variability. The first, which they aimed to account for in the TRACE model of speech perception, was “lawful” sources of variability, from, e.g., phonological rules or segmental context, e.g., impacts of preceding consonants on acoustic-phonetic realizations of speech sounds.

Elman and McClelland (1986) distinguished lawful variability from such variability as “cross-speaker and within-speaker token-to-token variability.” The TRACE model did not attempt to account for variability that was not lawful, as it is more difficult to adapt to systematically (as it varies more randomly). Nevertheless, its prevalence means it must still be accommodated by learners. Realizations of sounds differ between talkers, in part due to biological features including age (Schötz & Müller, 2007), vocal-tract size, health, and articulatory strategies (Johnson, Ladefoged, & Landau, 1993). Between-talker differences influence talker-voice characteristics (e.g., absolute formant frequencies) and can impact word learning and recognition.

Within-talker variability can also lead to different realizations of a sound from instance to instance. It can be due to myriad factors including changes in health status, fatigue (Greeley et al., 2007), circadian rhythm (Heald & Nusbaum, 2015), cognitive load (Segbroeck et al., 2014; Yap, Epps, Choi, & Ambikairajah, 2010), affective state (Cowie & Cornelius, 2003; Barrett & Paus, 2002), and recent linguistic experience (Evans & Iverson, 2007). Another source of not-lawful variability is essentially random variability that pervades speech, such that the same speaker using the same accent, register, affect, and sentence context will not produce identical realizations of a sound twice.

Finally, individual speakers also differ systematically in socially conditioned features (e.g., accent, sociolect). Accent variability occurs due to regional and social variation in language sound patterns (for example, AAE vs. WAE, or Chicano English vs. California Anglo English; Guerrero, 2014). Speakers can also employ distinct speech registers (such as formal vs. informal). It is not clear whether Elman and McClelland would consider differences across accents, sociolects, and registers as lawful or not.

Structures of variability.

Variability can be structured in a number of ways. Four particular structures of variability appear in Figure 1, which uses VOT and second-formant frequency as example dimensions. Table 1 summarizes experimental and real-world examples of each variability structure. Variability can be patterned, i.e., organized by modes or clusters. VOT, which differentiates sounds like English b and p at syllable onset, patterns bimodally (e.g., Lisker & Abramson, 1964). Alternatively, variability can be (seemingly) random noise (unimodally or uniformly distributed).

Figure 1: Possible relationships between speech and nonspeech variability in laboratory studies and natural settings.

Colors refer to speech-sound category. Rugs indicate frequency of each value on each axis.

Table 1:

A small number of studies and real-world scenarios where each type of variability depicted in Figure 1 is observed.

Within-category noise variability on contrastive dimension	Non-contrastive variability is noise-like	Non-contrastive variability is patterned but not correlated	Non-contrastive variability is patterned + correlated
√ Galle, Apfelbaum, & McMurray, 2015	√ Rost & McMurray, 2009, 2010; √ Quam et al., 2017, Exp. 1	√ Barcroft & Sommers 2005, Exp. 2, moderate variability condition (3 speakers)*	∅ Quam et al. 2017, Exp. 2 √ Creel 2014a
Language input predominantly from one caregiver	Learning from a large number of talkers with similar speech cues (e.g., VOTs)	Learning from small number of caregivers with distinct voices but similar speech cues (e.g., VOTs)	Certain talkers use certain speech forms (e.g., Foulkes, Docherty, & Watt, 1999)

Note. Top row: Experiments where: √ variability helps; ∅ no effect of variability. Bottom row: possible situations where such variability might occur.

^*This study was conducted with adults learning second-language vocabulary. Three speakers facilitated learning relative to one speaker.

Variability can either contribute or not contribute to distinguishing words. Some researchers use the terms relevant vs. irrelevant somewhat synonymously with contrastive vs. non-contrastive, but this the use of the terms relevant and irrelevant involves at least two presuppositions: (1) that learners’ goal is to distinguish words, not to (say) use coarticulatory cues to speed up word recognition, or to differentiate parents’ voices to identify the talker (e.g., the top vs. bottom two clusters in Figure 1, Panel 3); (2) that speech-sound variability is uncorrelated with non-segmental variability. Even accepting the first presupposition, that learners are learning words, the second is sometimes wrong, as experimental or natural situations may result in correlations between speech-sound and (e.g.) talker variability.

Instead of conflating contrastive and relevant variability, we use ‘relevant’ and ‘irrelevant’ more synonymously with ‘lawful’ (Elman & McClelland, 1986) vs. ‘unlawful,’ though we note that relevance must be defined relative to the task. Contrastive variability is lawful, and typically relevant. For example, assuming learners are trying to differentiate words, VOT variability is relevant (Figure 1, x-axes). Non-contrastive variability can be either lawful (and typically relevant), as for variability due to coarticulation and allophonic variation, or unlawful (and typically irrelevant), as for token-by-token variability across and within talkers. For example, in settings without talker variability—language-learning classes with one instructor, households with one caregiver, or experimental contexts—unlawful, irrelevant variation still occurs between instances of a speech sound (Figure 1, Panel 1).

Two additional structures of variability were not considered by Elman and McClelland (1984, 1986), but an argument could potentially be made that they are lawful, as they are patterned. First, in settings with, e.g., a small number of talkers (Figure 1, Panel 3), talker variability can be patterned (as contrasted with settings with a larger number of talkers; Figure 1, Panel 2), even if still irrelevant to linguistic tasks like word recognition. Second, it is possible for typically “irrelevant” variability, such as talker variability, to be correlated with “relevant” variability (Figure 1, Panel 4). When talker characteristics are correlated with speech sounds, either in experiments or natural settings, voice characteristics are actually relevant to word learning and recognition.

When acoustic-phonetic variability is correlated with speech sounds, it is important to distinguish two learning cases. First, when cue correlations persist past initial learning (especially for strong correlations), there will be relatively fewer instances where learners must generalize to uncorrelated or anticorrelated cases (blank spaces in Figure 1, Panel 4). Thus, correlated cues should largely facilitate learning and recognition by helping separate input distributions (as in Figure 1, Panel 4). This might happen in an experiment where a particular novel word, say darg, is always spoken by Talker 1, while the otherwise-similar darge is always spoken by Talker 2, and talker-word correlations persist throughout the experiment (Creel, 2014a; see also Creel et al., 2008, Creel & Tumlin, 2011, in adults; but see Quam, Knight, & Gerken, 2017); or in a less-deterministic case where, e.g., men tend to preaspirate less than women (Foulkes, Docherty, & Watt, 1999).

Second, when cue correlations do not reliably persist past initial learning, learners must generalize past their learning data. This might happen, e.g., when children learn certain words only from one caregiver and others from another caregiver, but must eventually generalize to other speakers. In these cases, short-term facilitation from correlated cues may come at the cost of reduced generalization beyond particular contexts or talkers, reflected in reduced recognition speed or accuracy. Even worse, correlated cues that are salient but eventually irrelevant might pull attention away from relevant dimensions, impairing learning (relative to random variability; see Quam, Knight, & Gerken, 2017).

A particularly complex case of variability patterning (which is again ambiguous in terms of lawfulness) arises from accent differences. How does variability pattern between accents? For some sound contrasts, accent variability is inconsequential because two accents have very similar sound patterns (Figure 2, Panel 1; see Table 2 for experimental and real-world examples). Examples include /i/ and /i/ (peak and pick) in German-accented and US-accented English (Flege, Bohn, & Jang, 1997), or many corresponding vowel contrasts in Los Angeles Chicano English and the local WAE (Santa Ana & Bayley, 2008). Other cases, however, create the possibility of confusion. One accent may merge a distinction another distinguishes (Figure 2, Panel 2). For example, Chicano English and Californian WAE merge the vowels in cot /ɑ/ and caught /ɔ/ (Santa Ana & Bayley, 2008), but many other US accents do not. Or, two accents may distinguish sound categories in different ways, either due to shifted speech-sound boundaries (Figure 2, Panel 3), e.g., lax front vowels in the Californian Vowel Shift (Eckert, 2008), or different sets of cues or cue weightings (Figure 2, Panel 4), as for coda voicing in AAE vs. WAE (Y. Holt, Jacewicz, & Fox, 2016; see also Baran & Seymour, 1976; Farrington, 2018). Unless learners can track accent context, the variability depicted in Figure 2 Panels 2-4 may introduce confusion. If learners can track accent context, they can successfully interpret sounds using accent-specific categories (for a cross-linguistic example, see Quam & Creel, 2017).

Figure 2: Some of the possible forms of between-accent variability in two speech-sound categories.

Rugs indicate frequency of each value on each axis.

Table 2:

A small number of studies and real-world scenarios, experienced by child listeners, where each type of accent variability depicted in Figure 2 is observed.

Accents show identical patterning	One accent merges sounds	Accents have shifted cue boundaries	Accents separate patterns differently
German-accented and US-accented English /i/ vs. /i/ (Flege et al., 1997)	The typical depiction of AAE coda devoicing as merger (of, e.g., had and hat)	White & Aslin, 2011; Weatherhead & White, 2016	AAE vs. WAE coda voicing (Y. Holt et al., 2016)
Many similar vowel contrasts in Los Angeles Chicano English and WAE (Santa Ana & Bayley, 2008)	caught/cot merger in California WAE but not northeastern WAE; salary-celery /æ/-/ε/ merger before /l/ in East LA Chicano English (García, 1984)	Lax front vowels in Californian Vowel Shift (Eckert, 2008)	AAE > WAE child listeners at AAE and cross-dialect comprehension (Baran & Seymour, 1976)

Note. Unshaded cells contain experimental or quasi-experimental tests of effects of variability on recognition or comprehension; shaded cells contain real-language cases where such variability has been described or measured. Because there is relatively little evidence of facilitation vs. inhibition based on particular accent sound-pattern relationships, we do not mark facilitation/inhibition, but discuss it in the text.

Putative Mechanisms for Impacts of Variability on Learning

Models of the earliest stages of perceptual learning for spoken language tend to focus on how infants discover the constituent sounds of their language(s). Despite evidence of continued sound-category development well past infancy (Hazan & Barrett, 2000; McMurray, Danelz, Rigler, & Seedorff, 2018; Nittrouer, Lowenstein, & Packer, 2009; Ohde & Haley, 1997; see Creel & Quam, 2015, for discussion), language-specific tuning within the first year is a focus of substantial theoretical and modeling efforts. Exposure to the native language over the first year tunes infants’ early, possibly species-general (Kluender, Diehl, & Killeen, 1987; Kuhl, 1981; Kuhl and Padden, 1982, 1983) sensitivities to sound patterns (Eimas et al., 1971; Jusczyk et al., 1980; Lasky et al., 1975; Sundara et al., 2018, Trehub, 1976). For vowels, this reorganization happens by 6 months (Polka & Werker, 1994); for consonants, by 12 months (Werker & Tees, 1984), though there are exceptions (e.g., Kuhl et al., 2006; Mazuka, Hasegawa, & Tsuji, 2014; Narayan, Werker, & Beddor, 2010; Polka, Colantonio, & Sundara, 2001).

We discuss models in order of their claims about the internal structure of representations, moving from models assuming more- to less-abstract representations of input variation. Historically, models of spoken-word processing traditionally assumed acoustic-phonetic variability was normalized to match word tokens to abstract representations, or models simplified their input to mostly lawful variability. Exemplar approaches, on the other hand (Goldinger, 1996, 1998; Johnson, 1997), store all variability in individual, detailed experiences, out of which emerge clusters representing words and speech sounds (and talkers and accent-specific patterns). Variability on irrelevant dimensions during learning helps relevant dimensions emerge as stable (Pierrehumbert, 2003), while also helping learners detect newly relevant information, like talker-word correlations (Goldinger, 1996, 1998). Thus, exemplar-based models address impacts of input variability on learning more directly than other models, though all the models we highlight focus on learning from input distributions.

Several theoretical frameworks attempt to capture reduced discrimination of nonnative speech-sound contrasts across infancy. Kuhl et al.’s (2008) Native Language Magnet Theory Expanded (NLM-e) argues decreased discrimination results from magnetic pull of category prototypes, which abstract away from input variability. Kuhl et al. predict this learning mechanism would apply to monolingual and bilingual learners, but bilinguals may show slower and/or different learning trajectories in response to the demands of bilingual input. One might infer that children hearing multiple accents would also show slower learning trajectories within each accent due to greater input complexity. Best’s (1994) Perceptual Assimilation Model predicts relative discriminability of nonnative speech-sound contrasts based on their relationship to native sound categories. If each sound in a contrast falls into a different native category, infants are predicted to continue to discriminate them (akin to between-accent relationships in Figure 2, Panel 1). If the sounds both fall within one native category, discrimination is predicted to worsen (akin to Figure 2, Panel 2).

The PRIMIR framework (Werker & Curtin, 2005), accounts for different sensitivity to speech-sound distinctions across ages and tasks. Speech is accessible through general-perceptual, word-form, or phonemic planes, but children’s biases, developmental level, and task requirements constrain access. PRIMIR was developed to address a complex pattern of findings about infants’ word learning as assessed in the “Switch” habituation procedure (Werker, Cohen, Lloyd, Casasola, & Stager, 1998). In one version, infants habituate to two word-object associations, and then at test, word-object pairs are maintained or switched. Word learning is inferred from a novelty response: longer looking to switched word-object pairings. Across many studies, 14-month-olds fail to distinguish similar-sounding words like /bi/ vs. /di/—even words differentiated by two features (/pin/ vs. /din/; Pater, Stager, & Werker, 2004)--despite distinguishing the sounds (e.g., /b/ vs. /d/, Stager & Werker, 1997; Werker, Fennell, Corcoran, & Stager, 2002). According to PRIMIR, 14-month-olds fail because they are processing word-form information (and representing it with an exemplar-based structure), while 17-month-olds succeed (Werker et al., 2002) by processing phonemic information. This might predict irrelevant variability would facilitate learning most strongly at 14 months (Rost & McMurray, 2009), prior to phonemic processing of words.

Jusczyk’s (1993) Word Recognition and Phonetic Structure Acquisition (WRAPSA) model is distinct from the approaches described above in assuming infants store specific instances of syllables (not phonemes), comparing them to traces of prior exemplars (variants), not prototypes. As experience accumulates, the auditory-analysis system adjusts weights of different features to highlight contrastive dimensions. In WRAPSA, young infants treat non-contrastive and contrastive variability as equally relevant, but non-contrastive variability receives reduced attention with age, suggesting (in line with PRIMIR) that younger infants would be more sensitive to facilitation from irrelevant variability than older learners. Apfelbaum and McMurray’s (2011) associative-learning model aims to explain facilitative effects of variability on 14-month-olds’ Switch word learning (Rost & McMurray, 2009). The model is somewhat consistent with exemplar models, but incorporates dimensional weighting very similar to WRAPSA (Jusczyk, 1993). Variability facilitates learning in the model by affecting association strengths between acoustic dimensions and visual referents. Thus, it may predict variability would not facilitate learning in tasks that lack visual referents, such as typical laboratory speech-sound discrimination tasks.

Sumner, Kim, King, and McGowan’s (2014; Sumner, 2015) exemplar-based framework aims to account for better memory for less-frequent but socially prestigious² forms of a word. (Note that, while the framework was based on adult studies, it is potentially relevant to children.) Sumner and colleagues posit that more-prestigious exemplars of a word are weighted more strongly (Sumner et al., 2014; Sumner, 2015). On this account, children learning multiple dialects may successfully learn prestigious forms from limited input, though this is not consistent with data indicating that AAE-speaking children’s adjustment to WAE, as the dialect of school, is often challenging (Washington & Craig, 2000). An alternative social-weighting approach, in line with traditional exemplar approaches, is that social context (talker, register, accent) is encoded within each exemplar. This would allow perceivers to up-regulate activation of contextually appropriate forms, yielding more efficient recognition. On this account, evidence for amplification of prestige forms could actually reflect contextual expectations: college students might expect to hear “prestigious” speech in scientific experiments. For multiple-accent learners, this account predicts that hearing fewer instances of words in each accent than single-accent learners will lead to slower accrual of exemplars, resulting in slower cluster formation (see discussion of NLM-e above). Thus, children learning multiple accents might show both context sensitivity and more difficulty with prestige forms compared to children hearing primarily the prestige dialect.

Evidence for Effects of Variability

We summarize evidence from three domains in which variability impacts children’s perceptual development for spoken language: learning and processing of speech-sound categories and patterns; word-form recognition and word learning; and accent. For the first two domains, we first describe how children learn linguistic structure from distributional patterns in the input, and then address impacts of variability on dimensions irrelevant to the task (e.g., sound discrimination or word learning). For accent, the distinction between relevant and irrelevant variability is less clearcut.

We focus mainly on child learners. We sometimes include brief descriptions of learning in adults, in some cases because a research question has been more thoroughly investigated in adults, and more generally because adults represent a developmental target or endpoint. For example, we note that a complementary literature in adult word and speech-sound learning focuses on influences of word knowledge on speech-sound retuning (termed lexical feedback; Norris, McQueen, & Cutler, 2003), with some research on bottom-up perceptual learning (e.g., Goudbeek, Swingley, & Smits, 2009; Idemaru & L. Holt, 2011). Like children, adults must balance stability of representations with perceptual flexibility (Baese-Berk, 2018; Earle & Quam, 2022). While we focus on development, we refer interested readers to Baese-Berk (2018) and Barriuso and Hayes-Harb (2018), among others, for more information.

Learning and processing of speech-sound categories and patterns.

In this section, we first review evidence from learning and processing of speech-sound categories, and then infants’ learning of patterns of speech sounds. In each domain of learning, we consider impacts of irrelevant variability that is clearly non-phonological, e.g., talker, pitch, and affect, on learning.

Learning and processing of speech-sound categories.

Contrastive variability between speech sounds is essential for infants’ phonological learning. Infants appear to infer one sound category if the distribution of a cue is unimodal and two if it is bimodal (Maye, Werker, & Gerken, 2002; see Vallabha, McClelland, Pons, Werker, & Amano, 2007 for a computational model). Infants can also use bimodal visual information to learn two phonemes from a unimodal phonetic distribution (Teinonen et al., 2008). Thus, variability between speech sounds can facilitate learning of sound categories by serving as input to distributional-learning mechanisms.

Variability during phonetic learning can also occur on non-contrastive dimensions. Such variability can be lawful (Elman & McClelland, 1986), as for variability across different within-word contexts, or unlawful, as for within-talker or between-talker variability across tokens. Here, we consider impacts of unlawful variability on infants’ speech-sound processing, but in the next section we consider infants’ learning of lawful variability (patterns of speech sounds). Even in adult learners, research on monitoring for phonemes suggests that unlawful talker variation interferes with speech-sound identification (e.g., Mullennix & Pisoni, 1990). Compared with mature listeners, infants could be even more susceptible to interference, given that some adult studies suggest variability can overwhelm learners with weaker perceptual skills (Antoniou & Wong, 2016; Perrachione, Lee, Ha, & Wong, 2011; Sadakata & McQueen, 2014).

To our knowledge, no studies have found facilitative effects of irrelevant, unlawful acoustic-phonetic variability on infants’ speech-sound processing, with a few cases of inhibitory effects. Two infant studies have found null effects. Quam et al., 2020 found that talker variability (4 habituation talkers) did not impact 7.5-month-old infants’ discrimination of either a native, salient onset contrast /b/ vs. /p/ or a non-native, less-salient onset contrast /n/ vs. /ŋ/. Kuhl (1983) trained 6-month-old infants to discriminate a vowel pair produced by a single synthesized “talker.” Infants generalized discrimination to simulated men’s, women’s, and children’s voices, suggesting synthesized talker variability does not disrupt infants’ discrimination. Two other early studies incorporating variability into training stimuli found mixed null and inhibitory effects. In the high-amplitude-sucking procedure, 4- to 16-week-old infants’ pitch discrimination was impaired by vowel variation (which we consider to be “irrelevant” for pitch discrimination), but not vice-versa (Kuhl & Miller, 1982; see also Lebedeva & Kuhl, 2010). The asymmetry likely reflected differences in acoustic salience (vowel>pitch) rather than phonological knowledge. When testing immediately followed familiarization, Jusczyk, Pisoni, and Mullennix (1992) found 2-month-olds detected changes to syllables (/bʌg/ vs. /dʌg/) whether they heard a single talker or multiple talkers. When a 2-minute delay was incorporated before the test phase, only infants in the single-talker condition detected changes. Thus, while infants track variability between speech sounds themselves to learn speech-sound categories, unlawful, irrelevant variability appears to either not impact or inhibit infants’ detection of changes in speech sounds.

Learning of lawful variability: patterns of speech sounds.

Some variation in speech sounds is non-contrastive, but lawful. One common type is allophonic variation, where a phoneme has different, but related, sound forms, lawfully conditioned by different phonetic contexts. For example, r in “braid” is voiced, while r in “prayed” is voiceless due to the preceding voiceless consonant. Infants can learn a novel phonological alternation and group allophones (surface realizations) into a single “functional category” by 12 months (White, Peperkamp, Kirk, & Morgan, 2008; see also Chong & Sundara, 2015). Infants’ ability to learn this patterning depends on phonetic similarity between sounds, consistent with the tendency for allophones to be phonetically similar (White & Sundara, 2014). Whether sounds are allophonic or contrastive also impacts infants’ learning of a novel phonotactic pattern (sound combination) at 11 months. Infants can only learn patterns relying on sound distinctions that are contrastive—not allophonic—in their language (Seidl et al., 2009).

Unlawful variability can sometimes facilitate infants’ learning of lawful variability. Seidl, Onishi, and Cristia (2014) reported facilitative effects of talker variability, for infants at 4 and 11 months, on phonotactic-pattern learning: discrimination of nonce words that were phonotactically compatible vs. incompatible with familiarization. Infants use talker differences to track grammatical-like rules (Gonzales, Gerken, & Gómez, 2018), but apparently not to track phonological statistics (Benitez, Bulgarelli, Byers-Heinlein, Saffran, & Weiss, 2020). Richtsmeier, Gerken, & Ohala (2011) found facilitation from talker and word-type frequency on preschoolers’ production of phonotactic sequences (see also Richtsmeier, Gerken, Goffman, & Hogan, 2009). In summary, infants and toddlers appear to correctly attribute variability in segments to allophony and other types of phonological alternations, differentiating allophonic variation from constrastive variation. Learning of such lawful variation in speech sounds can sometimes be facilitated by variability on irrelevant dimensions.

Word-form recognition and word learning.

In this section, we first review impacts of unlawful variability on early word-form recognition, the component of word learning that involves forming a representation of the sound pattern of the word, as opposed to learning the meaning or referent of the word or the association between sound and meaning. We then review impacts of unlawful variability on word learning—the mapping of word forms to meanings. As will become apparent here, and even more apparent in discussing accent variation, the boundaries between relevant and irrelevant variation are somewhat murky. Thus, the last topic we address in this section is the fact that learners must contend with ambiguity about whether segmental distinctions are relevant or irrelevant.

Impacts of unlawful, irrelevant variability on word-form recognition and word learning.

As with speech-sound learning, early word-form recognition and word learning require recognizing acoustic-phonetic forms by tracking variation on relevant dimensions, while ignoring irrelevant variation. A fundamental question is when children discover which variation is contrastive. This question has been probed somewhat differently for word-form recognition vs. word learning. For word-form recognition, researchers have tested how readily children generalize a word form (typically presented in isolation during familiarization) across a change in irrelevant properties to recognize it in fluent speech. By 9-10.5 months, infants recognize familiarized word forms despite unlawful variability in talker gender (Houston & Jusczyk, 2000), affect (Singh, Morgan, & White, 2004), or pitch (Singh, White, & Morgan, 2008) between familiarization and test.

Earlier, at 7.5 months, word recognition is disrupted by unlawful variability, unless familiarization words are embedded in fluent, natural speech (van Heugten and Johnson, 2012). While 7.5-month-olds generalized familiarized word forms to a new same-gender talker (Houston & Jusczyk, 2000), recognition was disrupted by a 1-day delay between familiarization and test (though reminding infants of the original talker in the test session helped them to generalize; Houston & Jusczyk, 2003). Relatedly, in some cases, 7.5-month-old infants only recognize words spoken in a familiar voice (their mother’s; Barker & Newman, 2004). Thus, unlawful acoustic-phonetic variability challenges early recognition of word forms. While 7.5-month-old infants’ word-form recognition is easily disrupted by irrelevant, unlawful variability, strategic introduction of training variability (e.g., multiple talkers or multiple vocal affects) can aid generalization across acoustic differences. While 7.5-month-olds do not generalize familiarized word forms across changes in vocal affect (Singh et al., 2004), incorporating affect variability into familiarization facilitates generalization (Singh, 2008). Thus, irrelevant, unlawful variability is challenging for infants’ early word-form recognition. However, it is also critically important for the formation of robust, generalizable word-form representations.

For word learning, as for word-form recognition, unlawful, irrelevant variability during learning may introduce additional complexity, but several studies suggest it facilitates learning. Talker variability facilitates infants’ associative word learning in the Switch procedure: when habituation words /buk/ and /puk/ were spoken by 18 male and female talkers, 14-month-olds successfully learned words (Rost & McMurray, 2009, 2010; see replication in Quam, Knight, & Gerken, 2017). Galle, Apfelbaum, and McMurray (2015) demonstrated that even single-talker variability (noise variability around voicing-category means) facilitates learning if it presents a wide enough scope of acoustic values. Facilitation at 14 months, specifically, is consistent with PRIMIR (which argues 14-month-olds are processing word forms holistically and storing exemplars; Werker & Curtin, 2005); as well as with WRAPSA (which argues irrelevant dimensions are weighted more heavily earlier in infancy; Jusczyk, 1993). Facilitation in word-learning tasks but not sound-discrimination tasks (summarized earlier) is also consistent with Apfelbaum & McMurray’s putative mechanism, that variability operates on cue weights between acoustic dimensions and visual referents (the latter of which are not included in common laboratory discrimination tasks).

Yet not all irrelevant variability facilitates infants’ word learning. While Höhle et al. (2020) replicated the talker-variability facilitation effect, variability in the visual appearance of objects did not facilitate word learning (see Sommers & Barcroft, 2013, for a related finding in adults). Variability could need to be in the speech signal to help infants rule out acoustic-phonetic dimensions as potentially relevant for differentiating words. However, Quam and Swingley (in prep.) found a null effect of acoustic-phonetic variability: 18-month-olds learned words in the Switch habituation procedure irrespective of whether variability was introduced on an irrelevant dimension (pitch contour for vowel-contrasted words, or vowel for pitch-contour-contrasted words).

While in many cases learners use variability on irrelevant dimensions to downweight those dimensions, sometimes they can use non-contrastive variability as a correlated cue to help them track differences in words (or grammatical patterns). Three- to five-year-olds (Creel, 2014a; and adults, Creel, Aslin, & Tanenhaus, 2008; Creel & Tumlin, 2011) use talker differences, such as a male voice always saying “marv” and a female voice always saying “mard,” to recognize otherwise-similar-sounding words more quickly in real time. These results are consistent with an exemplar approach in which learners store both speech-sound and talker information. However, in one study, 14-month-olds did not learn words when voice gender covaried with word, despite learning words with random talker variation (Quam, Knight, & Gerken, 2017). This set of findings suggests an age trajectory for word learning in which irrelevant variability, rather than being tuned out over development, may increase in usefulness over development. What is not clear is whether early presence of covarying cues may be detrimental to post-learning generalization, though as discussed in the section on structures of variability, there is less pressure to generalize if cue correlations are stable in learners’ environments.

In summary, irrelevant variability impairs early word-form recognition, but inclusion of irrelevant variability during familiarization appears to aid generalization across irrelevant changes in word forms. While some studies have shown facilitation from noise-like variability for word learning, others have found null effects. The structure and source of variability may interact with age to determine whether variability will facilitate learning: random acoustic (but not visual) variability facilitates word learning at 14 but not 18 months. Correlated talker-gender cues seem to facilitate certain types of learning, perhaps only at certain ages. They do not appear to facilitate 14-month-olds’ word learning, though they facilitate word recognition in preschool and beyond.

Coping with ambiguity in whether variation is contrastive.

For early word learning, differentiating contrastive vs. non-contrastive changes in words is complicated by the fact that what might appear to be contrastive changes in segments (not attributable to allophony, accent, etc.) could actually be inadvertent misspeakings. Thus, there is a tension for listeners: how does one know if a variant is inadvertent or deliberate? If inadvertent, one should disregard the variation and activate the closest-matching real word. If intentional, one should treat the change as phonologically relevant (Swingley & Aslin, 2000, 2002) and infer a new word with a distinct referent (Quam & Swingley, 2010; though see Swingley, 2016). One factor listeners consider is the degree or severity of the mispronunciation, e.g., how many phonetic features are changed. We illustrate with a series of word-learning studies by Quam and Swingley (2007, 2010, 2022), which focused primarily on toddlers, but included adults for comparison. Each 2.5-year-old or adult was repeatedly exposed to a novel object labeled “deebo” (/diboʊ/). A distracter object was visually familiarized, but not named. In test, they saw the two objects and heard either “deebo” or a mispronunciation. Visual fixations to objects assessed learners’ interpretations of words.

Results (Figure 3) differed both by age and degree of mispronunciation. Toddlers were sensitive to the more substantial two-feature vowel mispronunciation “dahbo” /daboʊ/, exhibiting more looks to the target for “deebo” than for the mispronunciation (Figure 3). But only 46% of children fixated the distracter more than the target (Quam & Swingley, 2010), suggesting most children did not fully reject the target as the referent of the mispronounced word. For the milder mispronunciation, toddlers did not show a decrement in target looking, implying acceptance or even obliviousness to the mispronunciation (Quam & Swingley, 2022; see White & Morgan, 2008, for similar findings with familiar words in 19-month-olds). Adult listeners were sensitive to both mispronunciations, but even adults seemed to interpret mispronunciations differently depending on their magnitude. For the two-feature vowel mispronunciation “dahbo” /daboʊ/, 75% fixated the distracter more than the target, suggesting they interpreted “dahbo” as a novel word. By contrast, for the single-feature consonant-voicing mispronunciation “teebo” /tiboʊ/, only 33% fixated the distracter more (Quam & Swingley, 2007).

Figure 3: Response to mispronunciations of a newly learned word (“deebo”) is related to degree of mispronunciation.

Adults and 30-month-olds showed a larger target-fixation decrement to a two-feature change in vowel height and front/backness (/i/ to /a/ in “dahbo”) than to a single-feature change in consonant voicing (/d/ to /t/ in “teebo”).

One might ask how well looking measures match children’s final interpretations. Does a high proportion of looks to a novel object indicate the child thinks it was the object named, or could it instead indicate uncertainty or interest in the novel object? Studies using both manual responses and looks suggest strong agreement between the two measures. Swingley (2016) presented 27-month-olds with familiar-word mispronunciations (e.g., “guck” for duck) in the presence of one familiar (duck) and one novel object. Children’s manual (touchscreen) responses and looking patterns both suggested children chose the duck the majority of the time for “guck,” while they mostly looked at and touched the novel object when hearing an unrelated form (e.g., “meb”). In older children, aged 3-5 years, Creel (2012) found something similar: children looked and pointed less at a picture (e.g., fish) when its name was mispronounced (e.g. “fesh”, and even less for more-distant “fosh”), but the familiar object (fish) was still the predominant response.

These studies taken together suggest children do not automatically assume a mildly mispronounced word is an unfamiliar one. Thus, interpretation of potentially relevant variation is gradient: in many cases—particularly for toddlers—variant forms are treated as close enough, with familiar(ized)-word activation decreasing as mispronunciation severity increases. Still, this gradient sensitivity must be reconciled with children’s demonstrated ability to distinguish similar-sounding familiar-word pairs. For example, children in Swingley (2016) treated “puhk” (a non-word rhyming with book) like “book,” but same-age children in Barton (1976) readily distinguished “bear” and “pear.” Why are b and p distinguishable in familiar words but not in novel words? This pattern of results is consistent with an exemplar account. If children learn entire word forms rather than individual sounds, knowing “bear” and “pear” does not imply differentiation of “book” and “puhk.” Differential sensitivity across acoustic contexts might also allow learners to use variability in one accent (e.g., where a contrast is meaningful) but not another (e.g., where it is merged).

In summary, toddlers’ detection of variation in segments in newly learned words is somewhat less robust than adults’. It is also gradient, with stronger detection for multiple-feature than single-feature mispronunciations. Sensitivity to partial similarity provides a mechanism for recognizing word forms not only when speakers make errors, but also when speakers systematically produce less-familiar forms, as in cases of unfamiliar accents. Thus studies of mispronunciation sensitivity provide an interesting counterpoint to the literature on accented-speech comprehension and adaptation (discussed below).

Accent variability.

We now address accent variability, which is highly complex, for several reasons. First, it is not clear whether or not cross-accent differences should be considered lawful—they were not directly addressed by Elman and McClelland (1984, 1986), who described how segments pattern systematically within an accent (though they did describe how bilinguals interpret sounds with respect to the language context). Second, while differences across accents do not signal lexical contrast, as described above, there can be ambiguity between accent differences vs. contrastive changes. Third, a difference between accents typically encompasses multiple sound distinctions. In many experimental studies, young learners’ detection of mispronunciations of words has been interpreted as a response to “accent” variability, but naturalistic accent differences are rarely as clear-cut as experimental mispronunciations of words. Rather than always shifting from one speech-sound category familiar to a child (e.g., /d/) to another familiar category (/t/), an accent might (partially) merge a category or change cues, as illustrated in Figure 2.

Fourth, naturalistic accent covaries with non-phonological features of dialect. In the sections below, we mainly use the term ‘accent,’ but we will also sometimes use the term ‘dialect,’ which has a somewhat distinct meaning. Accent refers specifically to pronunciation: either to differences between first- and second-language (L2) speakers’ pronunciations, which may be impacted by the speech sounds and prosody of their first language (L1); or to variation in pronunciation across different dialect groups. Dialect refers more broadly, not just to pronunciation, but also to vocabulary and grammar. As one aspect of dialect, accent often activates stereotypes, that is, overgeneralized language-external notions about speakers based on their perceived identity. This means that, in some cases, it is unclear whether recognition difficulty results directly from accent properties in speech or from stereotype activation. Listeners may experience subjective difficulty understanding speech because of negative associations with speakers’ cultural or linguistic backgrounds (Kang & Rubin, 2009; Kang, Rubin, & Lindemann, 2015) and then blame speakers for their difficulty understanding (Lippi-Green, 2012). In our review, we assume that recognition difficulty results from unfamiliarity of sound forms, but alternative social-bias explanations should be kept in mind.

Finally, given the high perceived social relevance of accent (e.g., as a cue to group membership), accent variability blurs distinctions between relevant (or lawful) and irrelevant (or unlawful) variability. Thus, instead of organizing the subsections below by relevant vs. irrelevant variability, we follow a different structure. We first discuss recognition and learning of words in unfamiliar accents. Then, we discuss learning in multiple accents, addressing impacts of dialect prestige.

Recognition and learning of words in unfamiliar accents.

Infants’, toddlers’, and preschoolers’ word learning in the face of accent variability is a research topic of increasing interest. As is typically the case in the experimental literature, there is a longer record of research on accent comprehension and adaptation in adults, which provides an endpoint measure of sorts for child accent processing. Briefly, adults are less adept at comprehending unfamiliar-accented than familiar-accented speech, but with exposure to an accent they rapidly improve (e.g., Bradlow & Bent, 2008; Clarke & Garrett, 2004; McQueen, Norris, & Cutler, 2003). Thus there is a cost to accent variation, but experience can reduce it.

In infants, a possible precursor to comprehension or adaptation to accented speech may be the ability to detect that two accents differ. Some research suggests accent-difference detection by 12 months (Butler, Floccia, Goslin, & Panneton, 2011; Cristia et al., 2014; Kinzler, Dupoux, & Spelke, 2007; Mulak, Bonn, Chládková, Aslin, & Escudero, 2017; Nazzi, Jusczyk, & Johnson, 2000). However, other studies indicate children as old as 5 years or more have difficulty differentiating their accent from another (Creel, 2018; Floccia, Butler, Girard, & Goslin, 2009; Girard, Floccia, & Goslin, 2008; Jones, Yan, Wagner, & Clopper, 2017), with slow developmental improvement. The apparent discrepancy between these research areas likely reflects methodological differences (Creel & Quam, 2015), so it is not clear how findings at earlier vs. later ages relate to each other.

There is also a lengthy timeline for recognition of words in unfamiliar accents, but it is not clear whether this skill is strongly linked to accent-difference detection. Children must develop “phonological constancy,” recognizing an accented form as a familiar word. With natural exposure to an accent, infants can demonstrate phonological constancy as early as 6 months (Kitamura, Panneton, & Best, 2013). For unfamiliar accents, infants can recognize familiarized words despite changes to the talker and the accent around 13 months (Schmale & Seidl, 2009; see also Schmale, Cristià, Seidl, & Johnson, 2010). However, Best et al. (2009) reported a later time-course: 15-month-olds only recognized a word as familiar in their native accent (see also Mulak et al., 2013), while 19-month-olds did so in a different accent. Infants under 2 can only recognize familiar words in an unfamiliar accent if they are first familiarized with the accent in a highly familiar story, but by age 2 they do not need accent familiarization, suggesting good generalization to unfamiliar accents (van Heugten & Johnson, 2014; van Heugten, Krieger, & Johnson, 2015). For novel-word learning, by 24 months, toddlers can learn a word in an unfamiliar accent after under 2 minutes of exposure to the accent (Schmale, Cristià, & Seidl, 2012). However, when taught a word in their native accent, only 30-month-olds—not 24-month-olds—recognize it in another accent (Schmale, Hollich, & Seidl, 2011).

Studies in older children suggest that difficulty understanding accented words does not end in toddlerhood, but extends at least into mid-childhood. Studies using sentence-repetition measures have found that children have greater difficulty comprehending words (Bent, 2014) and sentences (Bent & Atagi, 2015, 2017) in unfamiliar-accented speech (including unfamiliar native accents: Bent & R. Holt, 2018; see also Nathan, Wells, & Donlan, 1998). While performance improves with age and phonological skill (Bent & Atagi, 2015) and when semantic cues are provided (R. Holt & Bent, 2017), children as old as 8 still have more difficulty than adults (typically evaluated via transcription measures) in comprehending unfamiliar accents (Bent & Atagi, 2015).

Why do studies of older children show extended difficulty when studies with toddlers suggest good accent comprehension by 24-30 months? One reason may be the difficulty of the task. Toddlers are often tested with picture-fixation tasks that provide a small set of response alternatives (hearing “feesh” while seeing fish and banana pictures), while older children are tested with a repetition task, which provides less context (no pictures) and has a free response set. Creel, Rojo and Paullada (2016) found that this task difference is substantial: monolingual English-learning 3- to 5-year-olds showed good recognition of Spanish-accented words and sentences in a picture-recognition test—nearly comparable to their performance on WAE-accented words—but a much larger gap on the unfamiliar (Spanish) accent when children had to repeat heard words. In addition to emphasizing the importance of methodological factors (Creel & Quam, 2015), this collection of studies, taken together, suggests children may need more contextual support to recognize unfamiliar-accented variants at younger ages, with less support needed as development progresses.

Learning in multiple accents.

The question we address here is whether dual-accent input slows learning in each accent. Multiple studies make clear that children can learn to recognize and produce at least some types of cross-accent variability. Infants and toddlers can learn to interpret word forms differently depending on the accent of the speaker, both via lab training (Weatherhead & White, 2016; see also White & Aslin, 2011) and real-world exposure (van der Feest & Johnson, 2016; and see Smith et al., 2007, 2013 for children as young as 3 producing multiple phonological variants depending on the formality of a situation). However, several lab and observational studies suggest accent variation or multiple phonological variants may slow learning, relative to learning primarily in a single accent. To our knowledge, only one set of studies has simulated multi-accent input in the lab. Creel and colleagues have found that 3- to 5-year-old children (Creel, 2014b; Frye & Creel, under review) have more difficulty learning novel words when they are pronounced with segment variability (e.g., an object is called both “beesh” and “peesh”) vs. segment consistency (an object is always called “beesh”). This effect is gradient, with greater difficulty (lower pointing accuracy) when two labels for an object are phonologically dissimilar (an object is called “dedge” and “vush”; see Muench & Creel, 2013, for similar findings with adults).

Related evidence from infant studies suggests natural exposure to multiple accents may lead to slower comprehension in a given accent (Durrant, Delle Luche, Cattani, & Floccia, 2014; Floccia, Delle Luche, Durrant, Butler, & Goslin, 2012; Buckler, Oczak-Arsic, Siddiqui, & Johnson, 2017). Children exposed to multiple accents sometimes show more difficulty recognizing words in one of the accents than do children exposed to a single accent. However, this slower learning may occur because there is simply more for bi-accentual children to learn. A tradeoff often underemphasized in deficits-based accounts of language development in bilingual and bi-accentual children is that, while a monolingual or mono-accentual child may learn one accent faster, a bilingual or bi-accentual child can communicate effectively with more people and maintain stronger family and community connections (Yu, 2013).

To summarize, several laboratory studies show children can cope with accent variability, but it increases learning complexity and thus introduces challenges for learners, both in and out of the laboratory. At the same time, natural dialect exposure can be viewed as cultural and linguistic enrichment. The complexity-of-learning account predicts that learning difficulty will be proportional to the degree and nature of dissimilarity between a child’s two accents. This makes it critical to accurately characterize cross-accent variability, as we illustrate in the next section.

Impacts of dialect prestige on learning in multiple accents.

Another important consideration in learning and processing of multiple accents is impacts of dialect prestige (as imposed by society). Because of the ubiquity of prestige-dialect (WAE) speakers in media and society (their currently greater US population share, overrepresentation in political office and media), children who speak non-prestige dialects become familiar more quickly with prestige dialects (Perrachione, Chiao, & Wong, 2009) and have better comprehension of them (see Baran & Seymour, 1976; Nober & Seymour, 1974, 1979) than prestige-dialect learners have for non-prestige dialects. And because a prestige or majority dialect is likely to be the dominant dialect of teachers and educational materials (Craig & Washington, 2004), children who speak a non-prestige dialect at home are much more likely to learn to dialect switch (i.e., code-switch between dialects) than speakers of mainstream dialects (Thompson et al., 2004). Dialect switching requires adapting not only to phonological features (e.g., Thomas, 2007) but also to syntactic, morphological, and lexical patterns (Washington & Craig, 1994). Given that natural dialects vary on many dimensions simultaneously, learning a dialect other than WAE outside of school and then learning to dialect switch within an academic setting (Craig & Washington, 2004) might benefit children cognitively, akin to cognitive benefits attributed to bilingualism (Bialystok, 2007).

However, other work suggests AAE speakers experience substantial difficulty in WAE-dominated learning contexts (Washington & Craig, 2000), due to added complexity of adjusting to another dialect while simultaneously processing information needed for academic success (Craig & Washington, 2004; Thompson et al., 2004). This suggests that even if there is a boost to prestigious forms (Sumner, 2015; Sumner et al., 2014), it is insufficient for many learners. Recent research has noted that these challenges are heightened by the fact that content in standardized assessments (Moland, 2011) and curricular materials is almost always in WAE (Thompson et al., 2004). Children with more prior exposure to WAE in more integrated communities and/or stronger reading and vocabulary scores are better able to dialect switch (Craig & Washington, 2004), indicating they are mastering the complexity of bi-dialectal input to successfully acquire two dialect systems.

We now illustrate a case from AAE where mainstream characterizations of so-called “non-mainstream” dialects and neglect of relevant literature has impeded research progress. Among other features, AAE is described as having a merger of /i/ and /ɛ/ before nasal consonants (as in pin and pen) and final-consonant devoicing (“had” sounds like “hat”; for an overview see, e.g., Thomas, 2007). Thus, for these sounds, the AAE/WAE relationship is characterized as shown in Figure 2, Panel 2. However, it is possible that in some cases, accents like AAE with apparent mergers actually preserve contrasts, but along different dimensions (that WAE-speaking researchers do not hear), as illustrated in Figure 2, Panel 4.

Y. Holt et al. (2016) acoustically document a pattern in AAE speakers suggesting possible maintenance of coda-voicing differences. Multiple cues have been linked to coda-voicing differences, including preceding vowel length (the vowel in “had” is longer than “hat”), consonant-closure duration, and closure voicing (see, e.g., Hayes-Harb, Smith, Bent, & Bradlow, 2008). Y. Holt and colleagues measured vowel durations in 28 word pairs ending in d vs. t, such as had and hat, in AAE and WAE speakers from the same region in western North Carolina. AAE speakers produced a larger vowel-duration difference in these pairs than WAE speakers did (Y. Holt et al., 2016). This suggests AAE speakers in this region maintain coda-voicing contrasts but in a different way than WAE speakers—via stronger weighting on preceding vowel duration. This is consequential in that some research programs start with the assumption that AAE-speaking children have difficulty learning to read because merged distinctions lead to decreased spelling-sound correspondence. If instead AAE speakers do distinguish these sounds but rely on vowel-length distinctions over other cues, these could be highlighted by teachers and mapped onto spelling differences. Future research should focus both on developing best-practices for dialect-specific literacy instruction and on other sources of cross-dialect learning difficulty, such as reduced comprehension of instructions from differently accented instructors.

A study by Baran and Seymour (1976) suggests the acoustic distinctions AAE speakers produce are more intelligible to AAE than WAE listeners. AAE-speaking children surpassed WAE-speaking children in identifying AAE speech contrasts (predominantly coda-voicing pairs). In addition, AAE-speaking children were better at understanding WAE speech than vice-versa. Further, Nober and Seymour (1974, 1979) showed a similar pattern in pre-service teachers who spoke either AAE or WAE and tried to identify AAE and WAE child speech. AAE listeners were equally good at comprehending AAE and WAE, comparable to WAE speakers on WAE speech, and better than WAE speakers on AAE speech (Nober & Seymour, 1974, 1979). Neglect of Seymour’s prior research (e.g., Baran and Seymour, 1976, has been cited only 24 times) has obscured evidence for the presence of potentially contrastive patterning of coda-voicing distinctions in some AAE speakers, counter to the common narrative that AAE obscures such distinctions. Happily, some newer work (Y. Holt et al., 2015, 2016; Y. Holt, 2018) is beginning to fill this gap. Further research is needed, along with greater acknowledgment from reviewers, editors, and funding agencies that dialect variation is an important topic of inquiry and that non-“mainstream” dialects are not special, rare cases, but widespread in learners’ experiences. Better understanding of the phonological characteristics of AAE could inform more accurately targeted, phonologically based literacy instruction.

In summary, the question of how children learn in multiple accents is inextricably linked with sociological issues of dialect prestige. Dialect prestige introduces many asymmetries between the experiences of different groups of children as they cope with accent variability. Children dominant in a dialect of relatively lower prestige (as imposed by society) show some advantages over children dominant in a prestige dialect, such as greater familiarity with their non-dominant dialect and greater ability to dialect switch. At the same time, the tendency for academic settings to teach in the prestige dialect can introduce barriers for learning. Efforts to reduce so-called ‘achievement gaps’ for BIPOC children should carefully consider impacts of pressures to dialect (and language) switch on children’s academic access. More attention and resources must be paid to important research suggesting that speakers of AAE may maintain contrasts on different acoustic dimensions than speakers of WAE.

Conclusion

In summary, our review indicates that children learn language sound patterns in the face of rampant variability in the acoustic-phonetic realizations of sounds and words. Variability can be helpful, harmful, or neutral depending on the learner’s age and learning objective. Irrelevant, unlawful variability can facilitate children’s learning, particularly for early learning of words and phonotactic rules, but facilitation may depend on children’s age and the nature of the variability. In other domains, such as speech-sound change detection and word-form recognition, children seem either unaffected or impaired by unlawful variability, with more evidence of inhibitory effects of variability at or before 7.5 months of age. At the same time, inclusion of variability in training can aid generalization for young learners. While in some ways a source of “irrelevant” variation for sound or word learning, accent differences often interact with contrastive dimensions, so they can slow word learning temporarily and introduce academic barriers. However, exposure to multiple accents has benefits, including dialect-switching flexibility and likely better maintenance of family and community ties. More research is needed to better characterize how pressure to dialect switch contributes to so-called ‘achievement gaps’ for BIPOC children. Better characterizing the phonological properties of dialects like AAE could also contribute to more targeted and effective literacy instruction.

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.