Abstract

Introduction

There is a large literature on how infants become more sensitive to differences between phones that map onto different native phonemes (sounds that are both present and meaningful in the input) and less sensitive to differences between phones that map onto different non-native phonemes (sounds that are neither present nor meaningful in the input) as they mature. This literature shows that infants begin to zero-in on the phonemes present in their native language sometime between 4 and 12months of age (Werker and Tees, 1984; Polka and Werker, 1994). However, categorizing sounds as either present in, as opposed to absent from, the input is only one step in language acquisition. Certainly this helps the infant to focus on her specific language’s properties and ignore other language’s properties, and this ability to build language-specific phonetics may even be fundamental in building a lexicon (Kuhl et al., 2008). However, the child must also learn to categorize sounds which are present, but not meaningful in the input language. This task is likely to recruit the same mechanisms as the native/non-native task. Specifically, in every language, there are sounds that are present in the input but do not map onto different native phonemes, since their different forms do not cue meaning distinctions and the child must learn to map these to a unified phonemic representation. For example, the diphthong I in I’m should map to the same underlying category as the diphthong in I’d, despite the fact that one vowel is nasal and the other oral. For ease of expression and reading, we will use the shorthand of “allophones” for phones that map onto the same phonemic category, and “phonemes” for phones that map onto different phonemic categories¹. In this paper we summarize evidence on the acquisition of allophones to answer two key questions: When and how does the learner determine whether two sounds are allophones or phonemes in the target language?

Before turning to the evidence on the acquisition of allophones and the mechanisms underlying their acquisition, it is important to discuss both how allophones and phonemes are defined within the linguistic, descriptive literature (Section “What are allophones”); and how they are processed by individuals with a fully developed grammar according to the psycholinguistic literature (Section “The end state”). We then review an emergent strand of literature that documents when infants begin to apply a differential processing of allophones and phonemes (Section “Infants’ processing of allophones”) and how they might have learned to make such a distinction between allophones and phonemes (Section “Mechanisms for learning allophones”). The final section (“Implications”) discusses how research on infants’ learning of phonological status can inform, and be informed by, other areas of investigation. Throughout this article, we identify areas where answers are still lacking. We hope that this review serves as a springboard for such work and helps to point to clear areas out of which this future work can grow.

What are Allophones?

There are two classical cases of allophony, which are commonly discussed in introductory linguistics courses (Trubetzkoy, 1939/1969; Kenstowicz, 1994). The first involves “sounds in complementary distribution.” Two sounds are in complementary distribution if the sound which should be used is completely predictable from the context; put differently, the contexts in which each sound can occur are completely non-overlapping. For example, in most varieties of American English, dark /l/ occurs syllable-finally (“ball”), whereas light /l/ occurs in all other positions (“lab”). Notice that no two words in American English differ only on whether they have a light or dark /l/. In other words, sounds in complementary distribution do not cue meaning distinctions. Finally, a third criterion for allophony in this case is that the two sounds must be somehow acoustically related, such that they may be interpreted as the “same” sound, on some abstract level. For instance, although // and /h/ are in complementary distribution in English (the former occurs only in syllable codas, the latter only in syllable onsets), phonologists would not want to posit that they are allophones since they are highly acoustically distinct (Bazell, 1954).

The second classical case of allophony relates to sounds in “free variation.” In this case, speakers can produce two or more different sounds in the exact same environment (e.g., ri[]er versus ri[d]er in American English); however, these differences are not lexically relevant. Much work debates the name “free,” since in many such cases the variant which is selected appears to be explained, to a considerable extent, by a number of structural, sociolinguistic, and idiolectal variables (e.g., Fischer, 1958). Nonetheless, it remains the case that two sounds which can be thus exchanged without semantic changes can be viewed as allophones. The traditional way of establishing whether two sounds are in free variation is by carrying out a minimal pair test. Minimal pairs are two word forms that differ in only one sound; if this sound swap results in meaning change or loss, then the two sounds are phonemes, but if it does not, they are allophones in free variation.

In phonology, as in life, things can sometimes get more complicated, and for the definition of allophony this is true in a number of ways. To begin with, there are cases of complementary distribution and free variation that are true in certain phonological and lexical contexts, but not others. For example, one could state that voiceless unaspirated and voiceless aspirated stops are in complementary distribution in American English, with the former occurring e.g., after /s/ (as in “stop”) and the latter e.g., in the onset of monosyllables (as in “top”). However, voiceless unaspirated stops are minimally different from the surface realizations of voiced stops in syllable-initial position when following a word ending in /s/, to such an extent that one-year-olds fail to discriminate them (Pegg and Werker, 1997)².

Moreover, sometimes two pronunciations are possible without meaning change in some structural positions (e.g., [i]conomy vs. []conomy) but not in others (e.g., wom[ɪ]n vs. wom[]n; though perhaps a clearer example is tense and lax vowels, both of which can occur in closed syllables, but only tense vowels occur in open syllables). Additionally, some sounds would fit the definition of phonemes, but may be present in only a handful of loanwords (such as a pronunciation of the composer Bach as Ba[x] versus Ba[k]; Halle, 1964); whereas for others there may be no minimal pairs, even though the linguist’s intuition indicates that two sounds are contrastive because they are both active (used in a phonological constraint/rule) and prominent (involved in some type of phonological, morphological, or even long distance effect; Clements, 2001)³. Scobbie et al. (1999) and Scobbie and Stewart-Smith (2008), among others, have discussed extensively another ambiguous case from Scottish Standard English, where some vowels have long and short variants that are contextually determined, yet for which some minimal pairs, with specific morpholexical characteristics, can nevertheless be found. This is the case for long and short variants of /ai/, which contrast minimally in “side” and “sighed,” with the long version being found in morphologically complex items. In spite of the existence of such minimal pairs, the two sounds are in free variation across speakers in some lexical items, such as “crisis.”

In view of such cases both within and across languages, Pierrehumbert (2003) proposes to do away with the distinction between phonemes and allophones and instead attempt an explanation of learners’ acquisition of positional allophones, defined as clusters of tokens in acoustic space. A comparable proposal was made in Ladd (2006), who goes further by pointing out that allophones are sometimes very meaningful sociolinguistically, and are thus highly perceptually salient to native speakers. Scobbie and Stewart-Smith (2008) argue, instead, that while the concepts of allophones and phonemes may be useful end points, a continuum could exist between allophones and phonemes, and propose that speakers/listeners’ grammars could well be fuzzy. More recently, Hall (2009) makes specific proposals as to how to predict perceptibility from gradient versions of an allophony/phonemicness scale. Clearly the limitations of the classical definitions of allophones and phonemes are not new (see e.g., Pike, 1947), but they are just now beginning to gain a unique combination of linguistic and psycholinguistic attention as it becomes increasingly clear that such phenomena are not marginal, and that such gradience is relevant to both language learning and processing. Indeed, a look through Table ​Table11 reveals a window into the scope of this gradience. While it is not the aim of this paper to debate phonological theory, nor to enumerate cases along this continuum, we keep the question of gradience in mind when considering how infant learners may approach the phonological system, and what types of allophones versus phonemes (i.e., at what point of the continuum) have been studied in previous experimental work. With this enriched view of allophony, we now turn to adults’ perception of these two (or more) “classes” of sounds.

Table 1

Perceptible: native speakers report hearing the difference; Unpredictable: the phone cannot be predicted by its phonological context; Lexical: there are many examples of minimal pairs sustaining the contrast.

Type	Perceptible	Unpredictable	Lexical	Example	First author of relevant study
Phonemic	Yes	Yes	Yes	AmE [b-d]	Whalen; Hacquard
	Yes	Usually	Yes	AmE [I-i]a
	Yes	Mildly	Marginal	ScE [x]b
	Yes	No	Several	Sc [ai-ai:]c
	Yes	No	No	German ich-ach	Hacquard
	Yes	Rarely	Yes	AmE [-d]d	Boomershine
	No	Rarely	No	AmE [$\ae -\tilde{\ae }$]e
Allophonic	No	No	No	[p-ph]	Whalen

Open in a separate window

(Many examples below are from Scobbie and Stewart-Smith, 2008.)

^aTypically contrastive, but neutralized in some positions, e.g., seat vs. sit, but see vs. */sI/.

^bOnly present in a handful of lexical items, e.g., lock vs. loch.

^cOnly present in a handful of items, and predicted by syllable structure e.g., side vs.sighed.

^d[d] is typically realized as [] in specific contexts, so they are usually predictable from the context, but there are some cases where they contrast, e.g., rider-writer.

^eSome talkers use more heavily while others do not. E.g., some nasalize in non-nasal environments, or fail to nasalize in nasal environments.

The End State: Adults’ Processing of Allophones

It should be noted from the outset that the study of whether listeners’ sensitivity for contrasts that are allophonic is lower than those that are phonemic faces certain methodological roadblocks, which are worthy of discussion here. One way to interpret this hypothesis is the following: Holding the listener and language constant, one would compare a contrast A that is phonemic against a contrast B that is allophonic, to find that A is processed better (discriminated more speedily and accurately; used for tracking phonological patterns; recruited for coding lexical contrasts). Much of the initial literature we review uses this design (e.g., Whalen et al., 1997; McLennan et al., 2003). When using this design, there is an obvious interpretation alternative to phonological status affecting perception: perhaps there is an intrinsic discriminability difference between A and B. A safeguard against this state of affairs is to test two sets of participants, who have different native languages, and hold the contrast constant, an approach that is also common in the literature (e.g., Johnson and Babel, 2010). In this scenario, intrinsic differences in discriminability of contrasts are irrelevant, since only one contrast is used. However, another problem arises, namely that the stimuli must be recorded in some language. If they are recorded in the language where the contrast is allophonic, they may be pronounced less clearly (provided that speakers tend to neutralize such contrasts), which is not desirable. But if they are recorded in the language where they are phonemic, then the test may facilitate the performance of listeners of that language, who will find the stimuli native. The solution that researchers are increasingly adopting is to use a third language where the contrast is phonemic, such that the stimuli are equally foreign to both sets of participants. Results from the latter approach actually fit in perfectly with conclusions derived from the two other (e.g., Boomershine et al., 2008), lending further credence to this body of literature.

In brief summary, previous work suggests that adults do not discriminate allophonic alternates as well as phonemes both behaviorally (Whalen et al., 1997; Peperkamp et al., 2003; Boomershine et al., 2008; Shea and Curtin, 2011) and electrophysiologically (Kazanina et al., 2006; Hacquard et al., 2007). Furthermore, adults rate allophones as more similar to each other than phonemes (Johnson and Babel, 2010). Additionally, words differing in sounds that are allophonic prime each other, but words differing in sounds that are phonemic do not (McLennan et al., 2003). These differences in processing come about as the result of native language exposure and thus second language learners have a hard time gaining sensitivity to sounds that are phonemic in the target language even when those same sounds are present allophonically in the learners’ native language (Kondaurova and Francis, 2008).

Thus, overall, perceptual evidence in adults confirms that allophonic and phonemic sounds are not treated similarly. Given that there may be a continuum of allophones to phonemes, as mentioned above, it is worthwhile to evaluate whether this differential behavior arises only for the categorical phonemic/allophonic stages, or also for intermediate cases. This is especially true given recent findings that listeners treat sounds differently based on the reliability of their distributions (Dahan et al., 2008). Specifically, in this study Dahan et al. (2008) examined adults’ perception of tensed and laxed variants of /æ/ in the environment of /k/ and /g/. When /æ/ was consistently tensed before /g/, but not /k/ they found a training effect in the experiment such that listeners, upon hearing e.g., the non-tensed /æ/ came to anticipate the following segment as /k/. Thus, when segments vary allophonically in a reliable way this can lead to differential processing very quickly. This is not the case when the variation is not predictable.

In Table ​Table1,1, we reclassify adult perceptual studies in terms of the type of contrast that has been examined. There are several studies which explore one of the endpoints (e.g., Whalen et al., 1997 examines a case of clear complementary distribution) and only a few studies which explore points in between. For example, in English [] can never map onto /d/, and therefore they form a phonemic contrast. Whereas [] is a possible realization of /d/ in word-medial context, there are also quite a few lexical exceptions where they occur in near overlapping distributions (e.g., rider vs. writer). Thus, the comparison of English adults’ perception of the [] and [d], on the one hand, against [] and [d], on the other, represents the study of an intermediate case of allophony. This was undertaken in Boomershine et al. (2008), who found poorer discrimination of the former than the latter. Results cannot be attributed to the actual acoustic items used, since the same mapped onto different types for a second group of adults, whose native language was Spanish. In Spanish [] and [d] are in complementary distribution (classic allophony) and [] and [] are mostly in overlapping distribution (classic phonemic). Despite the fact that not all items fell on the extremes of the phonemicness/allophony continuum, perceptual results were the opposite across language groups in all tests but a measure of reaction time. Thus, this work seems to suggest that differences between phonemic and allophonic processing are found even when non-extreme points of the continuum are investigated.

Nonetheless, a different pattern emerges from work using electrophysiology. Hacquard et al. (2007) and Kazanina et al. (2006) both explore cases of complementary distribution, free variation, and overlapping distributions. While there are other effects in these studies (e.g., inventory size), overall, using an oddball paradigm, they find different processing results for complementary distribution (which patterns like overlapping distribution) and free variation (which patterns differently). For example, Kazanina et al. (2006), using Russian and Korean manipulated stimuli, find that while Russian listeners (for whom t/d are phonemic) show a significant mismatch response, Korean listeners (for whom t/d are allophonic) show no such response to the exact same stimuli. Hacquard et al. (2007) also examine whether the amplitude to the mismatch response in an oddball paradigm is related to the phonological status of the sounds in question using synthesized stimuli on vowel tenseness in continental French, Argentinean Spanish, and Puerto Rican Spanish listeners. Tense and lax [e]/[ε] are phonemic in continental French, allophonic in Argentinean Spanish and in free variation in Puerto Rican Spanish and this is reflected in the size of the mismatch responses. Furthermore, Argentinean listeners seem to discriminate the allophonic differences as well as they discriminate the phonemic ones based on the size of the mismatch response, but Puerto Rican listeners seem to discriminate the allophonic/free variation contrast more poorly than a phonemic contrast ([e]/[a]). Thus, theoretical descriptions and psycholinguistic evidence suggests that allophones and phonemes are different and that typology of the allophones may also be a factor in processing at least at some level. The next section assesses when these differences in processing come about over the course of development.

Infants’ Processing of Allophones

As mentioned above, a considerable body of literature suggests that perception of non-native (absent) sounds declines, whereas perception of native phonemes improves toward the end of the first year of life (Polka et al., 2001; Kuhl et al., 2006; Narayan, 2006). This improvement likely relates to the much richer and more abundant cues for the former: The child will accumulate more passive, phonetic exposure to the former; she may attempt these sounds; she may learn some words that have them, and so forth. The first question we would like to answer is when listeners become less sensitive to allophonic distinctions and more sensitive to phonemic ones. We review evidence from discrimination, phonotactic learning, phonotactic processing, and word learning suggesting that infants are sensitive to phonological status.

English-learning 2-month-olds discriminate allophonic variants (e.g., /t/ in “night rate” versus “nitrate”; Hohne and Jusczyk, 1994) showing an initial sensitivity to sounds that will eventually be treated as allophones later in life. Recent work suggests that, while young infants are sensitive to sounds that are allophones in their ambient language, this sensitivity declines with maturation and language-specialization. Specifically, Seidl et al. (2009) briefly familiarized English- and Quebec French-learning infants with a pattern that depended upon vowel nasality. Note that as mentioned earlier vowel nasality is phonemic in French, but allophonic in English. Infants in this study heard syllables in which nasal vowels were followed by fricatives, but oral vowels were followed by stops. Then they were tested on their ability to generalize this pattern to new syllables. English-learning 4-month-old infants were able to learn this novel phonotactic dependency involving vocalic allophones and behaved like French-learning 11-month-old infants, for whom nasality is phonemic. However, by 11months of age English-learners were no longer able to encode this abstract phonotactic regularity and showed no evidence of learning. It should be noted that these older infants are not completely impervious to allophones, since they use them to extract words from running speech at 10.5months (Jusczyk et al., 1999). Rather, these results suggest that the same exact sounds no longer function in the same manner across languages which use them as phonemes versus allophones.

It might be suggested that some of the contrasts that have been studied as allophones could be more perceptually difficult than ones that have been explored as phonemes. Specifically, allophonic alternates may simply be more difficult to discriminate because they represent subtle changes. For example, Pegg and Werker (1997) found that two phones that map onto different phonemes /t/ and /d/, but are extremely similar, are not discriminable by one-year-olds. In their study, they measured sensitivity to the word-initial realization of /d/ against the post-/s/ realization of /t/, which differ very subtly. However, an important point is that simple acoustic distance between the tokens used in any given test cannot explain developmental changes in allophonic sensitivity, since this sensitivity changes with age and language exposure. Even in the Pegg and Werker (1997) study, 6-month-old were, in fact, able to distinguish the very similar surface realizations of /t/ and /d/. Similarly, Dietrich et al. (2007) and Seidl et al. (2009) show that attention to the same contrast declines in languages for which they are allophonic, but not in languages in which they are phonemic. For example, Dietrich et al. (2007) show that 18-month-old Dutch-, but not English-learning toddlers interpret vowel length as lexically contrastive. Thus, it appears that while sensitivity to allophonic sounds initially exists in infancy, it appears to decline by 11months of age (Seidl et al., 2009) as infants converge on the native phonemic contrasts present in their input language and come to ignore the non-native ones which are not present in their input language (Werker and Tees, 1984).

It is worthy of note that most of the studies cited above have been conducted on English-learning infants (albeit with two exceptions, Dietrich et al., 2007; Seidl et al., 2009). If we are to draw any clear conclusions concerning the time course of allophonic sensitivity, we will need to expand this work cross-linguistically, since it may be that the time course is different across languages and may also be impacted by the kind of sound distinction explored. Unfortunately, such single language studies only allow for certain allophonic sounds to be tested, and confound potential differences in discriminability with phonological status.

Also worthy of mention is that many allophonic alternates in the studies mentioned above are predictable from the phonological context. For example, the aspiration of /t/ studied in Hohne and Jusczyk (1994) represents a clear case of complementary distribution or classic allophony. Exceptions to this are the cases of vowel nasalization utilized in Seidl et al. (2009) and the case of vowel length in Dietrich et al. (2007). Specifically, although vowels are nasalized before tautosyllabic nasal consonants in English, they are also often nasalized in other locations (e.g., within a word with another nasalized vowel), so complementary distribution does not entirely hold. Thus, although there are cases where nasalization of vowels is completely predictable on phonotactic grounds (before nasal Cs in the same syllable), we also see nasalization in other locations for coarticulatory reasons. To add more complexity to this picture, variation in nasalization has been reported across American English dialects, such that nasalization could become a sociolinguistically relevant feature (e.g., a marker of African American Vernacular English), more than a phonotactically relevant one. Similarly, in Dutch we see a case where vowel length is difficult to classify using our classic definition of allophony. Although there are minimal pairs with vowel length in Dutch, the presence if minimal pairs occurs unevenly across the inventory. For example, // has long and short minimal pairs that differ mostly in length (although there are slight vowel quality differences). All other vowels that have been described as contrastive in length show considerable changes of vowel quality with the addition of length, much as we see in English tense-lax pairs. Certainly, the infant literature is not rich enough to conclude that there are no differences among the different degrees of allophony. Nonetheless, current research suggests that in infants, as in adults, even degrees of contrastiveness may make a difference, with more allophonic pairs being processed less well than more phonemic pairs. Across all studies, however, it appears that younger infants attend to salient distinctions more than older infants when the distinctions are allophonic in the target language.

Mechanisms for Learning Allophones

Young toddlers treat allophones as distinct from phonemes. Further, some of the evidence reviewed suggests that they come to do so within the first year of life. How does such a young toddler come to treat allophones as distinct given that they clearly vary from language to language? Or more specifically, how do they come to attend less to allophonic sound pairs and attend more to phonemic sound pairs? There are several possible answers. Below, we describe computational models and laboratory studies documenting the ways by which allophonic treatment could come to be distinct from phonemic treatment.

Implications

Collectively, this work suggests that multiple mechanisms, likely including the computation of complementary distribution and the calculation of phonetic similarity, operate in concert to guide infants toward their functional interpretation of sounds that are present in the input, yet not contrastive. This review also bears on the more general question of how infants cope with phonetic variability that is not lexically meaningful such as variation between talkers’ voices and accents. Interestingly, infants become resilient to talker and accent changes also toward the end of the first year of life (Houston and Jusczyk, 2000; Schmale et al., 2010). Future work should investigate whether this similarity is merely superficial, or whether it is indicative of a perceptual reorganization allowing toddlers to recognize wordforms in the presence of lexically irrelevant variation. To answer this question, research should focus on how infants cope with deviations from canonical productions and how predictable those productions are. Moreover, the question of allophones is a categorical one, but many sources of variance are gradient and future work should explore whether these different kinds of variation are more or less learnable since it may be that gradient changes to the acoustic character of a sound are more variable.

A second consideration relates to the nuances in the concepts of phonemes and allophones laid out above, and predictions that can be stated on their learnability. Recent artificial grammar learning work suggests that infants tend to attend more to regular, neither entirely predictable nor entirely unpredictable, patterns (Gerken et al., 2011). In the domain of allophonic learning this might translate to different attention being allotted to patterns that are halfway between allophones and phonemes because of their very irregularity, a matter that could be investigated by assessing infants’ acquisition of different types of phonemes/allophones. Additionally, one could imagine that for infants the areas of the grammar in which the irregularity resides may be very important. For example, if the irregularity is lexically or morphologically based the language learning infant may not be immediately aware of it, and so would initially treat the pattern as if it were regular.

Additionally, differential processing of allophones and phonemes could inform translational research. For example, some work suggests that inappropriate learning of sounds in terms of these sound classes (e.g., perceiving equally well different phonemes and different allophonic alternates) correlates with reading ability and differs between normally developing and dyslexic children (Serniclaes et al., 2004). If we can pinpoint these differences in early development it may be possible to intervene while these infants are still at a very plastic stage of development. Thus, longitudinal studies exploring allophonic and phonemic processing may well contribute to early intervention at some point in the future.

Although we have steered clear of production in this review, it is certainly the case that accurately representing sounds as mapping onto distinct phonemes or the same phoneme should relate to production, since the target phonology for production will require the child to use the underlying sound in different ways in different environments. All signs indicate that this is a process that occurs quite early in development (Fikkert and Freitas, 2006). Still it is unclear to what degree the continuum between allophones and phonemes relates to production of those categories. We leave that question for a future review, but mention here that it is crucial to unite these two processes within the infant in order to truly understand the course of infant development.

It remains unclear how infants might make use of “phonetic similarity” in discovering allophones and distinguishing them from phonemes. For example, all vowels are more similar when compared with consonants, yet even young infants do not appear to have difficulty in distinguishing one vowel from another. It may be crucial to discern how acoustic similarity is judged vis-a-vis the infant. It is possible that lexical factors may also play a role in infant learning of phonological categories in a greater way than has been shown in learning models (Swingley, 2009).

Finally, it is clear that allophones may be relevant not just to phonological learning, but also to syntactic learning since allophonic alternates may mark phrasal edges (Selkirk, 1984; Nespor and Vogel, 1986; Seidl, 2000) and this marking may help infants to learn their syntactic structure if they are attentive to these edges (Nespor et al., 1996; Christophe et al., 1998). For example, if there is strengthening of contact at domain edges (Keating et al., 2003) or specific phonological processes at domain edges as mentioned above, e.g., a greater degree of aspiration or longer linguo-palatal contact the higher up you go in the prosodic hierarchy, then if infants are aware of the prosodic cues that they use for syntactic bootstrapping, this knowledge should inform or a least interact with their acquisition of the knowledge of allophones.

Conflict of Interest Statement

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Footnotes

References