7.3.1 Development of Speech Perception in Monolingual Infants

Most early speech perception studies have examined monolingual infants, that is, with little or no exposure to other languages. This research has overwhelmingly found effects of L1 experience on vowel and consonant perception in the second half-year of life (for reviews, see, e.g., Reference Best, Fernández and CairnsBest, 2017; Reference Best, Goldstein, Nam and TylerBest et al., 2016; Reference Werker, Yeung and YoshidaWerker, Yeung, & Yoshida, 2012). Before then, infants discriminate nearly all segmental phonetic contrasts of the world’s languages, including those lacking in their L1 environment. Thus, monolingual infants under six months do not yet appear to be perceptually attuned to native phonetic contrasts. But discrimination has declined by six months for most non-native vowel contrasts, and by nine to ten months for most non-native consonant contrasts examined. Most native contrasts continue to be discriminated well throughout that period. This general pattern of declining discrimination for most non-native contrasts but maintenance for most native contrasts is widely accepted and interpreted as evidence for perceptual attunement to native speech.

However, different developmental trajectories have been reported in monolingual infants for some consonant contrasts (see Reference Best, Fernández and CairnsBest, 2017; Reference Best, Goldstein, Nam and TylerBest et al., 2016; Reference Liu, Peter and TylerLiu, Peter, & Tyler, 2023; Reference Tyler, Best, Goldstein and AntoniouTyler et al., 2014). Discrimination of some native contrasts is only fair rather than good initially, improving during the second half-year (Reference Kuhl, Stevens and HayashiKuhl et al., 2006), while for others it is initially poor and shows even more delayed improvement with L1 experience, not evident until four years (Reference Polka, Colantonio and SundaraPolka, Colantonio, & Sundara, 2001). Conversely, initially good discrimination of some non-native contrasts is maintained across infancy and through to adulthood, rather than showing a decline (e.g., Reference Best, McRoberts and SitholeBest et al., 1988; Reference Best and McRobertsBest & McRoberts, 2003). The disparate developmental patterns for these contrasts appear largely compatible with PAM principles. For example, the Tigrinya ejectives /p’/–/t’/ that English-speaking adults assimilate to native stops /p/–/t/ ([pʰ]–[tʰ], a two-category assimilation) and the Zulu click contrasts that they perceive as nonspeech sounds (nonassimilable) are both discriminated well by English-learning infants from early on and show no decline (Reference Best and MolfeseBest, 1988; Reference Best, McRoberts, LaFleur and Silver-IsenstadtBest et al., 1995; see Reference Best, Goldstein, Nam and TylerBest et al., 2016). Meanwhile, the Zulu stop-ejective contrast /k/–/k’/ that English monolingual adults assimilated as good versus deviant exemplars of English /k/ ([kʰ]), and discriminated well (category-goodness assimilation), is discriminated at six to eight months but not at ten to twelve months (Reference Best and McRobertsBest & McRoberts, 2003).

By tuning in to the higher-order invariant properties of phonological categories, infant speech perception begins to specialize for efficient recognition and learning of meaningful words in the L1. The emergence of phonological constancy and phonological distinctiveness in the second year coincides with the vocabulary spurt, a period of rapid word learning that occurs once an infant’s expressive vocabulary (i.e., words they can produce) reaches around fifty words (Reference Nazzi and BertonciniNazzi & Bertoncini, 2003). For monolingual infants, this occurs at around seventeen to eighteen months of age. We suggest that it is the discovery of phonological constancy and distinctiveness, across phonetically variable productions of words (e.g., varying talkers, emotions, accents), that supports this rapid vocabulary expansion (Reference Best, Romero and RieraBest, 2015; Reference Best, Tyler, Gooding, Orlando and QuannBest et al., 2009).

7.3.2 Development of Speech Perception in Bilingual Infants

But how does an infant being raised bilingually optimally attune to the phonologies of two languages? It is tempting to ask simply whether bilingual infants show the same trajectories for their multiple languages as monolinguals show for their single L1, the focus of such research to date. But that question implicitly assumes a clear categorical distinction between monolingual versus bilingual infants, which is overly simplistic when given the multifaceted differences in language learning contexts between monolingual and bilingual children as well as among individuals on each side of what is actually a rather fuzzy dichotomy (e.g., Reference JohnsonJohnson, 2018). Even the speech heard on one or the other side of this blurry divide shows individual variation in quantity and quality, in the number of people speaking the language(s), their accents (native regional variants and/or L2), and sociocultural and family differences in spoken communication (Reference GathercoleGathercole, 2014; Reference JohnsonJohnson, 2018).

We also must consider whether bilingual infants acquire their languages simultaneously or sequentially, a fundamental distinction in adult bilingual research and theory. Simultaneous bilingual adults have acquired both languages “from infancy” (two L1s), whereas sequential bilinguals learned one language from birth (their L1) and the other later (their L2), either during the preschool years (early-sequential bilinguals) or later in childhood, or in adolescence/adulthood (late L2 learners). However, these definitions fail to capture the richly varied reality of infant bilingual contexts. While some do learn their languages simultaneously “from birth,” even these children’s circumstances can differ. For example, each parent may speak a different language to them, or both parents may speak both languages to them, or the parents may speak one language and an extended family member who regularly cares for the child from birth may speak another. But many, if not most, bilingual infants encounter their two languages sequentially. For example, they hear a single home language from birth, then begin to experience another one months later when they start spending regular time with people other than the parents, such as a nanny, other family caregivers, or daycare teachers. In such sequential cases the home language is often the parental heritage language, and the later-encountered L2 is used by the broader community/country, but the reverse situation can also occur (e.g., when a nanny or daycare teachers speak in languages other than the home/majority language; see Reference De HouwerDe Houwer, 2021; Reference GathercoleGathercole, 2014). Relatedly, one language may dominate, and dominance may shift over time. For example, the home language may be the “dominant” language when an infant is very young, but dominance shifts to the community L2 when the child enters daycare/preschool or interacts with L2-native peers in weekly play groups.

Bilingual infant speech perception studies have generally assumed a quasi-binary distinction that fails to consider either simultaneity-sequentiality or the other contextual differences among bilingual infants. The language background of participants has usually been grouped using arbitrary language exposure thresholds (e.g., those with > 85 percent exposure to a single L1 defined as monolingual, and those with < 65 percent exposure to one language and > 35 percent exposure to another as bilingual; Reference Bosch and Sebastián-GallésBosch & Sebastián-Gallés, 2001), but many infants fall between those two extremes. As a result, we still know little about the effects of the factors discussed here on bilingual infants’ perceptual attunement, or the vestiges of that attunement on adult bilingual speech perception.

7.3.3 Speech Perception in Bilingual Infants

Research on speech perception in bilingual infants is newer and less extensive than that on monolinguals, but has been increasing, along with reviews and theoretical analyses (e.g., Reference Byers-Heinlein, Grosjean and Byers-HeinleinByers-Heinlein, 2018; Reference Curtin, Byers-Heinlein and WerkerCurtin, Byers-Heinlein, & Werker, 2011; Reference Höhle, Bijeljac-Babic and NazziHöhle, Bijeljac-Babic, & Nazzi, 2020). The first such study examined discrimination of the Catalan-only vowel contrast between mid-high /e/ and mid /ɛ/ by monolingual and bilingual infants learning Spanish and/or Catalan (Reference Bosch and Sebastián-GallésBosch & Sebastián-Gallés, 2003). All groups discriminated Catalan /e/–/ɛ/ at four months, but by twelve months only Catalan monolinguals and bilinguals did so, not Spanish monolinguals, indicating perceptual attunement in all groups. At eight months, however, only the Catalan monolinguals discriminated it, indicating that the two monolingual groups were already perceptually attuned to their L1s. Despite their Catalan experience, the bilinguals showed a dip in discrimination at eight months, which they recovered by twelve months. In a follow-up study, an eight-month dip and twelve-month recovery was also found in bilinguals’ discrimination of a vowel contrast shared by Catalan and Spanish, high versus mid-high back rounded /u/–/o/, which both monolingual groups discriminated at all ages (Reference Sebastián-Gallés and BoschSebastián-Gallés & Bosch, 2009). The authors speculated that this indicates a temporary delay in bilinguals’ development of perceptual attunement. However, using a more sensitive anticipatory eye movement task at eight months they found that bilinguals, like Catalan monolinguals, can discriminate Catalan /e/–/ɛ/ whereas Spanish monolinguals failed (Reference Albareda-Castellot, Pons and Sebastián-GallésAlbareda-Castellot, Pons, & Sebastián-Gallés, 2011). Thus, the bilinguals’ earlier difficulty at this age appears to reflect greater sensitivity to task demands, rather than an actual loss of perceptual ability.

Analogously, English monolingual and Spanish-English bilingual infants discriminated the Midwestern American English /e/–/ɛ/ contrast at both four and eight months (Reference Sundara and ScutellaroSundara & Scutellaro, 2011) in a habituation task that was also more sensitive than Reference Bosch and Sebastián-GallésBosch and Sebastián-Gallés (2003). In another study, monolingual Dutch infants and infants being raised bilingually in Dutch and one of a range of languages (e.g., English, German, Spanish, or Turkish) completed the same sensitive habituation task with the Dutch /ɪ/–/i/ contrast, a purely spectral high versus mid-high front distinction lacking from the bilinguals’ other languages (Reference Liu and KagerLiu & Kager, 2016). Neither group discriminated it at five to six months, only the bilinguals discriminated it at eight to nine months, and then both groups succeeded at eleven to twelve months. The authors interpreted this pattern as a bilingual lead in perceptual attunement to this initially difficult native vowel contrast. That inference is consistent with monolingual English and bilingual Spanish-English infants’ brain responses when discriminating another difficult purely spectral English contrast, mid-high versus mid front lax /ɪ/–/ɛ/, indicating that bilingual infants may attend more to the distinction than monolinguals (Reference Shafer, Yu and DattaShafer, Yu, & Datta, 2011; Reference Shafer, Yu and Garrido-NagShafer, Yu, & Garrido-Nag, 2012).

Considering the task and vowel contrast differences, we draw the following inferences about bilingual infants’ perceptual attunement to vowel contrasts: they are neither delayed nor advanced relative to monolinguals in perceptual attunement to vowel contrasts, but instead differ in their attention to them, which is influenced by task demands. This attentional difference may explain the seemingly counterintuitive findings of a bilingual dip/delay at eight months for initially easy (discriminated by all groups at four to six months) native vowel contrasts (Catalan/Spanish /u/–/o/; Catalan /e/–/ɛ/; English /e/–/ɛ/), but a bilingual lead at eight to nine months for initially difficult (not discriminated by any groups at four to six months) native contrasts (Dutch /ɪ/–/i/; English /ɪ/–/ɛ/).

As for consonant contrasts, Canadian English monolingual versus French-English bilingual infants were examined on discrimination of stop consonant voicing distinctions that differ between languages in VOT (Reference Burns, Yoshida, Hill and WerkerBurns et al., 2007). As noted earlier, English uses short-lag versus long-lag aspirated VOTs to distinguish voiced versus voiceless stops, whereas, like Greek and Spanish, French uses near-zero-lag (Canadian) or voicing-lead (European) (Reference Caramazza and Yeni-KomshianCaramazza & Yeni-Komshian, 1974) versus short-lag VOT. The infants were habituated to a language-ambivalent short-lag [p] heard as /p/ by French-speaking but /b/ by English-speaking adults, then heard two test trials that assessed their discrimination of near-zero-lag [b] (Canadian French /b/) and long-lag [pʰ] (English /p/) from the habituation stimulus. At six to eight months both monolingual English and bilingual French-English infants discriminated the English [p]–[pʰ] but not the Canadian French [b]–[p] contrast, which was difficult for both groups. Bilinguals discriminated the French contrast from ten months onward, whereas English monolinguals again failed. This finding provides evidence of perceptual attunement for this consonant contrast by ten months for both groups, consistent with prior research showing perceptual attunement for consonant contrasts in monolinguals (see Reference Liu and KagerLiu and Kager [2015] for similar results with monolingual Dutch and bilingual Dutch-{French, Spanish, English, German, or Chinese} infants discriminating Dutch and English VOT contrasts). Analogous results were also found for a place-of-articulation difference between the allophones of /d/ in Canadian French (dental [d̪]) and Canadian English (alveolar [d]) (Reference Sundara, Polka and MolnarSundara, Polka, & Molnar, 2008). The [d]–[d̪] contrast was successfully discriminated by all three groups at six to eight months: bilingual French-English, monolingual French, and monolingual English. However, only the bilingual and the monolingual English infants discriminated the contrast at ten to twelve months, which is consistent with discrimination of the same contrast by monolingual English and bilingual adults (Reference Sundara and PolkaSundara & Polka, 2008). Neuroscientific studies on VOT comparisons with English monolingual and English-Spanish bilingual infants largely concur (Reference Garcia-Sierra, Rivera-Gaxiola and PercaccioGarcia-Sierra et al., 2011; Reference Ferjan Ramírez, Ramírez, Clarke, Taulu and KuhlFerjan Ramírez et al., 2017), and further suggest that bilinguals’ brain responses may show enhanced executive functioning and a longer period of “openness” to language experience relative to monolinguals (see also Reference Petitto, Berens and KovelmanPetitto et al., 2012; Reference Singh, Loh and XiaoSingh, Loh, & Xiao, 2017).

Our inference is that discrimination ability and perceptual attunement per se do not differ between bilinguals and monolinguals early in life. What differs instead is their experience-driven attentional biases to the phonetic details of native and non-native contrasts, which are affected by task demands as well as the other factors we discussed earlier (see also Reference StrangeStrange, 2011; Chapter 9, this volume). Thus, the simple “same or different trajectory” question is off target for understanding perceptual attunement in bilingual infants. The insights and inferences we have drawn here from research on the beginnings of language-specific perceptual attunement in monolingual versus bilingual infants provide developmental grounding for returning to consider our core question of how PAM principles might account for cross-language speech perception in adults who had acquired their two languages in the first two years of life, that is, mature early bilinguals.

8.2.1 Linguistic Perception (LP)

The term “Linguistic Perception” reflects the notion that human speech perception is a language-specific rather than a general auditory process. Reference EscuderoEscudero (2005, p. 7) defines speech perception as “the act by which listeners map continuous and variable speech onto linguistic targets.” Given that the very purpose of speech communication is to understand and to be understood, the listener’s task is to map the incoming variable acoustic cues (e.g., first formant or F1, second formant or F2, fundamental frequency, and duration) onto discrete and abstract linguistic representations (e.g., distinctive features, segmental categories, and suprasegmental structures) to ultimately extract the meaning intended by the speaker. The mapping patterns are language-specific in nature, since the number of linguistic representations and the use of acoustic cues vary substantially not only across languages but also across varieties or dialects of the same language.

Consider, for example, how the acoustic cues of F1 and F2 can map onto vowel categories. These cues, though physically continuous, should perceptually map to a different number of discrete categories depending on the language. Native English listeners need to make a fine-grained mapping of the two cues onto a dozen vowel categories so that they can identify and distinguish minimal pairs such as “heed,” “hid,” “hayed,” “head,” “had,” “hud,” “hod,” “hawed,” “hoed,” “who’d,” “hood,” and “heard,” although “a dozen” is a very rough approximation because the exact number of categories varies across different dialects of English. The mapping is much less dense for Arabic, which has only three qualitative contrasts (/i/, /a/, and /u/), though again dialectal variations exist. Languages also exhibit divergent mapping patterns even when they have the same number of categories. For example, native listeners of Greek, Hebrew, Czech, Spanish, and Japanese, all of which have a five-vowel system in their standard varieties, show distinct mapping patterns of the F1 and F2 cues per language (Reference Boersma, Chládková, Lee and ZeeBoersma & Chládková, 2011; Reference Escudero, Sisinni and GrimaldiEscudero et al., 2014b).

The language-specific nature of speech perception is formulated in the LP model as the optimal perception hypothesis,Footnote ² which posits that listeners learn the optimal mapping of acoustic cues onto appropriate sound representations that leads to maximum likelihood behavior (Reference BoersmaBoersma, 1998, p. 337). This means that the probability of correctly perceiving the intended linguistic representation based on the acoustic cues is maximized or, to put it another way, the probability of misperception is minimized. Native listeners’ perception of a language is optimal in that it tries to extract as many linguistic representations as required in the language (e.g., a dozen vowel categories in English, three in Arabic, or five in Greek, Hebrew, Czech, Spanish, and Japanese, with nonnegligible dialectal differences). It is also optimal in that the mapping patterns mirror the acoustic cues in the language (e.g., Japanese /u/ is generally more fronted than Spanish /u/, and so the perceptual usage of the F2 cue differs between the two languages).

How do native listeners acquire such language-specific, optimal perception? The LP model assumes a general learning device that is responsible for creating representations and adjusting cue usage, which is computationally implemented (see Section 8.2.3) by the Gradual Learning Algorithm (GLA; Reference Boersma and HayesBoersma & Hayes, 2001). An important attribute of the learning device is that it is distribution- and meaning-driven. It is distribution-driven in that it collects the statistical information concerning the acoustic cues in the ambient language and gradually adjusts the mapping patterns based on this information (Reference Boersma, Escudero, Hayes, Solé, Recasens and RomeroBoersma, Escudero, & Hayes, 2003), whereby the resulting perception exhibits what is known as the perceptual magnet effect (Reference KuhlKuhl, 2004). The device is meaning-driven in that it evaluates how the mappings signal lexical contrasts to determine the number of representations required for optimal perception in the language. The meaning-driven nature of the device implies that LP goes beyond simple acoustic-to-category mapping, since sound categories alone are meaningless unless they are associated with higher-level lexical representations. These learning mechanisms work alongside a complex structure involving multiple levels of representation and connections between them, as shown in the current LP model illustration in Figure 8.1 (Reference Escuderovan Leussen & Escudero, 2015).

Figure 8.1 Current full architecture of LP.

In Figure 8.1, the bottom-level representation, the [auditory] form, refers to the incoming acoustic signals as they arrive in the peripheral auditory system. The variable [auditory] form is then mapped to the following /surface/ form, which encodes the listener’s language-specific and invariant representations of speech sounds, including context-specific allophonic details. The /surface/ form is further abstracted into the third, |underlying| form, which encodes the canonical phonemic contrasts that may change the meaning of a word. Finally, the |underlying| form is connected to the <lexical> form, namely, to words and morphemes stored in the mind or brain. These representations, together with the connections between them, are acquired via distributional and meaning-driven learning. The [auditory]-to-/surface/ mappings (cue constraintsFootnote ³) are learned based on the distributions of acoustic values, while the connections between /surface/ and |underlying| forms (phonological constraints) are learned in relation to which <lexical> forms exist or not (lexical constraints).

Importantly, notice in Figure 8.1 that LP distinguishes prelexical perception and lexical recognition. Most psycholinguistic models of speech perception agree that lexical recognition guides perceptual learning, but it remains controversial whether the two processes are sequential (i.e., bottom-up) or interactive (i.e., bottom-up and top-down). The original LP model (Reference EscuderoEscudero, 2005, Reference Escudero, Boersma and Hamann2009) held a sequential view where perception precedes recognition, that is, the outcome of perception is faithfully passed on to recognition. According to this view, the lexical influences on perception are explained by offline (i.e., post hoc) learning from the lexicon (see the Merge model; Reference Norris, McQueen and CutlerNorris, McQueen, & Cutler, 2000). In contrast, the revised LP model (Reference Escuderovan Leussen & Escudero, 2015) allows for an interactive view as well, in which the lexicon can influence lower-level representations during the online (i.e., ad hoc) processing of speech (see the TRACE model; Reference McClelland and ElmanMcClelland & Elman, 1986). While the pursuit of this matter is beyond the scope of this chapter, the distinction and the connection between prelexical perception and lexical recognition will be discussed in Section 8.3.

8.2.2 Second Language Linguistic Perception (L2LP)

The L2LP model is a conceptual extension of the LP framework for L2 learners. The model consists of five theoretical ingredients, as shown in Figure 8.2, where straight arrows represent the ingredients’ sequential nature and curved arrows represent relationships between ingredients.

Figure 8.2 The five theoretical ingredients of L2LP.

As shown in Figure 8.2, the first ingredient is optimal perception in the listener’s L1 and the target L2. As mentioned, LP is language-specific, with the number of linguistic representations and the mapping of acoustic cues being unique to each language. This means that optimal perception for one language is not necessarily optimal for another and vice versa (see Footnote 2 for L2LP’s definition of “optimal”). The L2LP model proposes that a thorough analysis of optimal perception in each language, and specifically in each variety or dialect of a language,Footnote ⁴ is a crucial first step toward an adequate account of L2 perception, as L1 and L2 optimal perception define the initial state (Ingredient 2) and the end state or ultimate attainment (Ingredient 5) of L2 learning, respectively. To adequately and correctly predict and explain L2 development, the focus should be on the acoustic distributions of the target sounds and their closest L1 counterparts (closest in terms of acoustic-auditory proximity), together with their phonemic and allophonic status in the two languages (whether the sounds are lexically contrastive or not), but other factors such as the quantity and the quality of input and the learners’ cognitive capacity and skills are also relevant, as we shall see later.

The second ingredient of L2LP is the L2 initial state. The model’s Full Copying hypothesis, which derives from the Full Transfer hypothesis (Reference Schwartz and SprouseSchwartz & Sprouse, 1996), states that listeners start with a copy or duplicate of their L1 optimal perception at the onset of L2 learning. This results in the listener having a separate system or grammar for each of their L1 and L2, through which the sounds in the L1 and the L2 are perceived, respectively. Listeners at this stage are called “naïve” because no L2 learning has taken place yet, and their perception of target language sounds is commonly called “cross-linguistic” because L2 sounds are filtered by the L1. Note that both L1 linguistic representations and perceptual mappings are copied, which relates to the learning tasks in Ingredient 3.

Since the initial L2 grammar is seldom optimal for perceiving L2 sounds because of mismatches between optimal L1 and L2 perception, learners often struggle with misperception and miscommunication in the target language. The learners’ goal, then, is to modify the L2 grammar to solve the mismatch. Two kinds of learning tasks (Ingredient 3) are specified for this goal: a representational task to modify the number of categories (by forming new ones or disposing of existing ones), and a perceptual task to adjust the acoustic cue usage (by changing the weighting of Familiar cues and/or creating new mappings of Unfamiliar cues).Footnote ⁵ The L2LP model proposes that three types of learning scenarios emerge depending on the task(s): Similar, New, and Subset. These are illustrated with examples in Figure 8.3 and explained in detail in the paragraphs that follow.

Figure 8.3 Three types of learning scenarios in L2LP.

The Similar scenario occurs when the same number of representations are involved across the two languages. For example, L1 Canadian English listeners’ perception of the L2 Canadian French /æ/–/ɛ/ contrast falls into this scenario (Reference Escudero, Boersma and HamannEscudero, 2009).Footnote ⁶ While Canadian English also has /æ/ and /ɛ/ that differ in both F1 and duration (where /ɛ/ is generally shorter than /æ/), Canadian French /æ/ and /ɛ/ differ primarily in F1 with little durational differences. The weighting of the F1 and the duration cues is thus different between the two languages. Consequently, Canadian English learners of Canadian French tend to misperceive durationally short tokens of L2 /æ/ as /ɛ/, relying on their higher use of duration cues in the L1. The learners therefore have the perceptual task of adjusting the nonoptimal cue weighting to minimize the likelihood of L2 misperception. They do not have a representational task in this scenario because no addition or removal of categories is needed.

The New scenario occurs when L2 representations outnumber L1 representations. Unlike the Similar scenario, this scenario poses a representational task because a new sound category needs to be formed for L2 optimal perception. There are two subscenarios of New that differ in the perceptual task: one that involves an Unfamiliar acoustic dimension and the other that involves only Familiar acoustic dimensions. An example of the Unfamiliar New scenario comes from L1 Iberian Spanish listeners’ perception of the L2 Southern British English /iː/–/ɪ/ contrast (Reference Escudero and BoersmaEscudero & Boersma, 2004). This corresponds to a New scenario because the target L2 vowels, which contrast in both F1 and duration, map to the same L1 vowel /i/. The duration cue is Unfamiliar because Spanish does not employ duration for segmental contrasts.Footnote ⁷ The learners’ perceptual task is to create completely new mappings (e.g., long versus short) on this “blank slate” or “uncategorized” acoustic dimension. The mappings are then integrated into an existing category to create new ones (e.g., long /i/ versus short /i/) to accomplish the representational task. On the other hand, the Familiar New scenario occurs when new perceptual mappings are created along acoustic dimensions already utilized in the L1. An example of this scenario is L1 Tokyo Japanese listeners’ perception of the L2 American English /ɛ/–/æ/–/ʌ/ contrast (Reference Yazawa, Whang, Kondo and EscuderoYazawa et al., 2023b). This is also New because the three L2 vowels map to two L1 vowels /e/ or /a/. A notable difference from the case of Reference Escudero and BoersmaEscudero and Boersma (2004) is that the learners’ L1, Japanese, has phonemic vowel length, unlike Spanish. Given that all relevant acoustic cues for vowel identity (F1, F2, and duration) are Familiar in the L1, the perceptual task is to alter the existing mapping patterns along the known acoustic dimensions. This would result in the splitting of an existing category (e.g., /a/) to yield a new one (e.g., /æ/), as part of the representational task.

Finally, the Subset scenario occurs when L1 representations outnumber L2 representations, whereby an L2 sound perceptually maps to more than one L1 representation. The L2LP model is currently the only model that addresses this mapping pattern, which Reference Escudero, Boersma, Skarabela, Fish and DoEscudero and Boersma (2002) termed Multiple Category Assimilation (MCA). Examples of this scenario are L1 North Holland Dutch listeners’ perception of the L2 Iberian Spanish /i/ and /e/ (Reference Boersma, Escudero, Avery, Dresher and RiceBoersma & Escudero, 2008; Reference Escuderovan Leussen & Escudero, 2015) and L1 Australian English listeners’ perception of the L2 Iberian Spanish vowels (Reference Elvin, Escudero, Gutierrez-Mangado, Martínez-Adrián and Gallardo-del-PuertoElvin & Escudero, 2019). For Dutch listeners, the Spanish vowels /i/ and /e/ perceptually map to /i/, /ɛ/, or /ɪ/ in the L1, thus resulting in MCA. Here, the listeners can have a representational problem where three categories are perceived instead of two, which could lead to spurious lexical contrasts (i.e., /i/–/ɪ/ or /ɪ/–/ɛ/) in the L2. Even when they “know” from textbooks that there are only two such vowels in Spanish, they have a perceptual problem where their L2 initial grammar cannot help automatically mapping relevant acoustic cues to three categories. Thus, learners have a representational task to unlearn unnecessary categories and a perceptual task to alter the existing mapping so as not to perceive them.Footnote ⁸

The fourth ingredient of L2LP is L2 development, for which the proposal states that L2 learners have Full Access (Reference Schwartz and SprouseSchwartz & Sprouse, 1996) to the general learning device of LP (which is computationally implemented by the GLA; see Section 8.2.3) throughout their lifetime. Thus, L2 perceptual learning is assumed to be fundamentally similar to L1 learning, which is distribution- and meaning-driven. Studies have shown that distributional learning has immediate and long-lasting effects on adult L2 learners (Reference Escudero, Benders and WanrooijEscudero, Benders, & Wanrooij, 2011; Reference Escudero and WilliamsEscudero & Williams, 2014). The meaning-driven nature of L2 perceptual learning becomes evident when its relationship with lexical development is considered (Section 8.3). The hypothesized full access to an L1-like learning device does not guarantee that L2 learning occurs as quickly and as effortlessly as L1 learning, however. In fact, researchers have long noted that adults progress more slowly than children in L2 perceptual learning. The L2LP model attributes age effects to cognitive plasticity, which peaks in youth and then gradually decreases as one gets older. Crucially, Reference EscuderoEscudero (2005) also argues that the role of input outweighs that of plasticity, which explains why learners of the same age and linguistic background may not follow an identical developmental path, since the quality and the quantity of input are modulated by various factors including motivation. These factors have significant implications for predicting the end state of L2 learning (Ingredient 5).

The model’s final ingredient proposes that all L2 learners can ultimately acquire L2 optimal perception regardless of their age, provided that sufficient and appropriate linguistic input is continually provided to the learner. This holds true for all three learning scenarios, though different scenarios can pose different levels of difficulty depending on the number of learning tasks. Specifically, it is proposed that the New scenario is the most difficult, followed by Subset and then by Similar, as forming new categories is considered more difficult than deleting or reusing existing ones (Reference EscuderoEscudero, 2005, p. 125). Note that the L1 grammar remains intact because L2 development occurs in a separate copy of the grammar (see Ingredient 1), unless input is diminished or stops when moving to an L2-speaking environment, as implied by the higher weight of input for L2 development and L1 maintenance in Ingredient 4. Thus, L2LP predicts that learners can attain two separate optimal grammars for the two languages. This hypothesis may raise questions because bilinguals can show bidirectional interactions when they code-mix their two languages (Reference Antoniou, Best, Tyler and KroosAntoniou et al., 2011). The L2LP model explains such phenomena with the assumption of gradient and parallel activation of the two grammars, which derives from Reference Grosjean and NicolGrosjean’s (2001) language mode hypothesis. A recent L2LP study has confirmed perception modes in L1 Japanese learners of L2 American English, who adapt their cue weighting (duration versus F2/F1) for vowel perception depending on whether they listen to English or Japanese (Reference Yazawa, Whang, Kondo and EscuderoYazawa et al., 2020). Crucially, in this study, some learners showed L1–L2 intermediate cue weighting, implying that both grammars were activated to different degrees, which was also shown previously in Reference Boersma, Escudero, Avery, Dresher and RiceBoersma and Escudero (2008) and Reference Escudero, Boersma and HamannEscudero (2009). Within Ingredient 5, it is also proposed that for ultimate attainment and successful performance, bilinguals need to master language control (Reference GreenGreen, 1998) or selective inhibitory control (Reference Friesen, Luo, Luk and BialystokFriessen et al., 2015). This proposal can explain individual or group differences in performance and will be relevant for comparing results of different types of bilinguals in Section 8.3.2.

In sum, the five ingredients of L2LP constitute a comprehensive model of how L2 perception is acquired, starting from the initial state (Full Copying of L1 optimal perception), through learning tasks (Similar, Unfamiliar/Familiar New, and Subset) and development (Full Access to an L1-like learning device mediated by input and plasticity), to the end state (L1 and L2 optimal perception activated in different degrees). While these theoretical components alone can explain and predict the outcome of various L1–L2 learning scenarios, the model’s predictive power is further reinforced by employing computational and statistical methods, as described next.

8.2.3 Computational Implementation of L2LP

A unique strength of L2LP is that its theoretical component can be computationally implemented to simulate L2 speech perception. While sometimes conflated, simulation should be distinguished from modeling (Reference Maria, Andradóttir, Healy, Withers and NelsonMaria, 1997). A model is a representation of the construction and working of a system of interest, and modeling is the process of building a model. For example, Reference EscuderoEscudero’s (2005) work concerned the modeling of L2 speech perception, which resulted in the L2LP model. A model should be a close approximation to the real system it represents, incorporating its salient attributes, but it should not be too complex to understand. A good model, therefore, is a trade-off between realism and simplicity.Footnote ⁹

A simulation, on the other hand, involves the operation or implementation of a model through configuring it to virtually experiment with it. Simulations can serve at least two purposes. First, one can validate a model by implementing it computationally under known conditions and comparing the output with the real system output. For example, L2LP’s Ingredient 2 (the initial state) can be tested by simulating a virtual listener who learns to perceive Spanish as their L1. This virtual performance is then compared to that of real learners’ perception gathered experimentally. Second, simulations can predict the performance of a system under different configurations and over long periods of time, which would be too expensive or impractical to conduct in the real world. For example, the outcome of specific L2 learning environments can be predicted by reconfiguring the types of input and the learning period, such as one, three, six, and eighteen years of L2 Spanish input fed to an L1 Dutch grammar (Reference Boersma, Escudero, Avery, Dresher and RiceBoersma & Escudero, 2008) or a few months versus a few years of L2 English input to an L1 Japanese grammar (Reference Yazawa, Whang, Kondo and EscuderoYazawa et al., 2020). The main incentive for computational modeling in L2LP is thus to provide a direct test for a hypothesis before conducting an empirical study, resulting in the formulation of more accurate predictions.

Although L2LP’s theoretical components can be implemented with various computational methods, the model has often utilized Stochastic OT (Reference BoersmaBoersma, 1998) and the GLA (Reference Boersma and HayesBoersma & Hayes, 2001). More recently, neural networks have been used to extend these frameworks (Reference Escuderovan Leussen & Escudero, 2015). Stochastic OT is a probabilistic extension of OT that is used to represent the learners’ language-specific grammar. The GLA is an error-driven algorithm for learning optimal constraint rankings in Stochastic OT that represents the learning device, which has been shown to outperform other machine learning algorithms (Reference Escudero, Kastelein, Weiand and van SonEscudero et al., 2007). While we do not intend to provide detailed explanations of how Stochastic OT and the GLA work here, interested readers can find step-by-step instructions for implementing L2LP with these computational methods in the following studies: Reference Yazawa, Whang, Kondo and EscuderoYazawa et al. (2020) for a Similar scenario; Reference Escudero and BoersmaEscudero and Boersma (2004) for an Unfamiliar New scenario; Reference Yazawa, Whang, Kondo and EscuderoYazawa et al. (2023b) for a Familiar New scenario; and Reference Boersma, Escudero, Avery, Dresher and RiceBoersma and Escudero (2008) for a Subset scenario. These studies focus mainly on the acquisition of the cue constraints (see Figure 8.1), hence representing classic L2 perception research (i.e., cue-based segmental category identification and discrimination), but Reference Boersma, Benz and MattauschBoersma (2011) discusses how the phonological and lexical constraints can also be implemented. See also Reference Escuderovan Leussen and Escudero (2015) for how the constraints at different levels may interact and for an implementation using an approach more compatible with neural networks.

Other studies have utilized statistical methods to make L2LP’s predictions more specific (Reference Escudero, Boersma and HamannCurtin, Fennell, & Escudero, 2009; Reference Elvin, Vasiliev, Escudero, Gibson and GilElvin, Vasiliev, & Escudero, 2018b; Reference Elvin, Williams and EscuderoElvin, Williams, & Escudero, 2016). For example, Reference Elvin, Williams and EscuderoElvin et al. (2016) applied discriminant analysis to Australian English vowel production data to predict which acoustic cues (duration, formant means, and formant changes) would contribute to the identity of /iː/–/ɪ/–/ɪə/ and to what extent. The analysis, which has been used for assessing cross-linguistic phoneme categorization as well (Reference Escudero and VasilievEscudero & Vasiliev, 2011), found that /ɪ/ can be durationally distinguished from /iː/ or /ɪə/ while formant changes were essential for distinguishing /iː/ and /ɪə/. The statistical model results accurately predicted real Australian English listeners’ perception (Reference Williams, Escudero and GafosWilliams, Escudero, & Gafos, 2018), which resembled simulation results using Stochastic OT and the GLA (Reference YazawaYazawa, 2020). The key point here is that quantification of theoretical predictions is a crucial component in the explanatory adequacy of L2LP, whether the adopted method is computational, statistical, or both.

8.3.1 Learning to Distinguish L2 Sounds in Context: The Case of Minimal Pairs

Previous L2LP studies on lexical development of phonological contrasts employed a word learning paradigm where each novel word is explicitly and unambiguously paired with its corresponding referent. The method involves a learning phase where participants are presented with a picture of a novel object in tandem with the object’s auditory form, followed by a testing phase where they hear one of the learned words and select the corresponding visual object. Many studies using this paradigm have shown that adults and children can learn minimal pairs, that is, words that are distinguished by a phonological contrast, in their L1 or in a subsequent language (Reference EscuderoEscudero, 2015; Reference Escudero, Broersma and SimonEscudero et al., 2013, Reference Escudero, Simon and Mulak2014a; Reference Escudero, Hayes-Harb and MittererEscudero, Hayes-Harb, & Mitterer, 2008; Reference Escudero and KalashnikovaEscudero & Kalashnikova, 2020; Reference Giezen, Escudero and BakerGiezen, Escudero, & Baker, 2016). Results also show that word recognition accuracy is linked to how well the phonological distinction is perceived, confirming the L2LP proposal of a tight relationship between prelexical perception and lexical development. The mechanism underlying this rapid word learning ability is commonly called fast mapping (Reference Escudero, Smit and AngwinEscudero et al., 2023).

However, this type of explicit and intentional learning does not entirely reflect how the learning of new words proceeds in more naturalistic and immersive environments. Specifically, everyday situations typically pose high levels of ambiguity because a novel word may appear alongside many potential referents (Reference Mulak, Vlach and EscuderoMulak, Vlach, & Escudero, 2019; Reference Yu and SmithYu & Smith, 2007). Real-world ambiguity can be resolved by drawing conclusions from statistical regularities across instances or situations where the same word is presented, a mechanism known as cross-situational word learning (CSWL; Reference Escudero, Mulak, Fu and SinghEscudero et al., 2016a, Reference Escudero, Mulak and Vlach2016b, Reference Escudero, Smit and Angwin2023; Reference Yu and SmithYu & Smith, 2007). Studies have shown that adults (Reference Angwin, Armstrong, Fisher and EscuderoAngwin et al., 2022; Reference Escudero, Mulak, Fu and SinghEscudero et al., 2016a, Reference Escudero, Mulak and Vlach2016b; Reference Mulak, Vlach and EscuderoMulak et al., 2019) and children (Reference Escudero, Mulak and VlachEscudero, Mulak, & Vlach, 2016c; Reference Pino Escobar, Tuninetti, Antoniou and EscuderoPino Escobar et al., 2023; Reference Smith and YuSmith & Yu, 2008) use CSWL to learn words in their L1 and in subsequent languages (Reference Escudero, Mulak and VlachEscudero et al., 2016b, Reference Escudero, Smit and Mulak2022; Reference Junttila and YlinenJuntilla & Ylinen, 2020; Reference Tuninetti, Mulak and EscuderoTuninetti, Mulak, & Escudero, 2020). Importantly, CSWL differs from incidental word learning paradigms used in previous L2 vocabulary learning studies in that CSWL is not only unintentional but also ambiguous (see Reference Escudero, Smit and AngwinEscudero et al., 2023).

Reference Escudero, Mulak and VlachEscudero et al. (2016b) were the first to apply the CSWL paradigm to the learning of minimal pairs, demonstrating that adult monolingual Australian English listeners could track word-referent co-occurrences while at the same time attending to phonetic/phonological distinctions in spoken words. Participants were shown the eight words in Figure 8.4 (left) in pairs that formed nonminimal pairs (e.g., /bɔn/–/dit/), vowel minimal pairs (e.g., /dit/–/dɪt/), or consonant minimal pairs (e.g., /bɔn/–/tɔn/) in English. The experiment consisted of learning and testing phases (Figure 8.4, center and right). During learning, participants were presented with a series of trials with two auditory words and two visual objects, without any instruction about the nature of the task or the correspondence between the words and the objects. Each trial was ambiguous because the order of presentation of the auditory words was not synced with that of the visual objects. Participants were then asked to identify word-object mappings in the test phase. Performance at test was above chance for all pairs, but vowel minimal pairs were less accurate than consonant and nonminimal pairs, with no difference between the last two. These findings suggest that phonological-lexical encoding may be weaker for vowels than for consonants, at least in Australian English monolinguals, indicating that this unintentional and ambiguous paradigm can be used to explore the link between prelexical perception and lexical development proposed in L2LP in naturalistic environments (see Reference Escudero, Smit and MulakEscudero et al., 2022; Reference Escudero and Hayes-HarbEscudero & Hayes-Harb, 2022). The obvious next step was to examine how bilinguals fare at encoding phonological detail in ambiguous, everyday situations.

Figure 8.4 Illustration of the CSWL paradigm.

8.3.2 Different Bilinguals Learning Minimal Pairs: Advantages and Disadvantages

Most L2LP studies and most studies previously conducted within the field of L2 phonetics and phonology show that monolinguals outperform sequential bilinguals in their target language, which leads researchers and observers to conclude that the “optimal” end state in L2 acquisition is very hard to achieve. Additionally, the idea that monolinguals outperform bilinguals has been confirmed in most studies on lexical processing (see Reference Gollan, Kroll and RappGollan & Kroll [2001] for a review). Contrary to this common belief, Reference Escudero, Mulak and VlachEscudero et al. (2016b) found that sequential bilinguals had comparable word learning performance to monolinguals when tested in the same CSWL task described in Section 8.3.1. One reason for this discrepancy may be that the CSWL task allows sequential learners to perform well regardless of their linguistic background, yielding results that are different from those gathered with more conventional tasks and with those predicted by the L2LP model. However, another CSWL study (Reference Tuninetti, Mulak and EscuderoTuninetti et al., 2020) has shown that the relationship between L1 and L2 phonemes predicts the difficulty with which Australian English listeners learn Dutch and Portuguese words, indicating that CSWL results are in line with the L2LP proposal that perceptual difficulty is correlated with word learning and recognition difficulty depending on the listeners’ linguistic background.

Alternatively, the composition of the bilingual group may have influenced the results. That is, the sequential bilinguals in Reference Escudero, Mulak and VlachEscudero et al. (2016b) came from diverse linguistic backgrounds and had diverse onsets of acquisition for their L2 English,Footnote ¹⁰ and including this type of “heterogeneous” bilinguals may have obscured differences when compared to monolinguals. To test the L2LP developmental proposal that both perceptual difficulties related to linguistic background and acquired proficiency play a role in CSWL of minimal pairs, two studies were conducted using the same method reported in Reference Escudero, Mulak and VlachEscudero et al. (2016b), summarized at the end of Section 8.3.1. The first study tested simultaneous Mandarin-English bilinguals in Singapore (Reference Escudero, Mulak, Fu and SinghEscudero et al., 2016a), while the second tested a group of homogenous sequential bilinguals with L1 Mandarin who started learning L2 English at school and resided in Shanghai or Sydney at the time of testing (Reference Escudero, Smit and MulakEscudero et al., 2022). As expected, opposite group results were found, where simultaneous bilinguals performed better overall than monolinguals, while sequential bilinguals showed overall lower performance than monolinguals.

Both results are explained by the bilinguals’ linguistic background and their acquired proficiency. The L2LP model’s explanation for better performance in simultaneous bilinguals relates to the inclusion of language control and selective inhibition as part of ultimate attainment (see Ingredient 5), which states that with high proficiency and continuous input from both languages, a bilingual can perform equally to a monolingual of either language. It seems that the heterogeneous group of sequential bilinguals in Reference Escudero, Mulak and VlachEscudero et al. (2016b) had high enough L2 proficiency and did not activate L1 features that could have negatively affected their performance. In the case of the simultaneous bilinguals in Reference Escudero, Mulak, Fu and SinghEscudero et al. (2016a), their “advantage” for overall word learning is explained by their heightened ability to selectively inhibit or suppress the irrelevant language, which may enable higher performance when coping with the ambiguity of a CSWL task. This L2LP proposal is in line with studies showing a bilingual advantage for simultaneous bilinguals depending on their levels of inhibitory control, an ability connected to the general-domain executive functioning (Reference Friesen, Luo, Luk and BialystokFriessen et al., 2015; Reference Pino Escobar, Kalashnikova and EscuderoPino Escobar, Kalashnikova, & Escudero, 2018). Thus, the L2LP explanation extends the bilingual advantage to the domain of statistical learning of minimal pairs, which involves the encoding of phonological distinctions.

In contrast, the overall “disadvantage” for the homogenous sequential bilinguals is explained by the activation of an L1 Mandarin linguistic feature, namely, contrastive lexical tones, due to the pitch variations in the stimuli presented, since no negative evidence against the use of tonal contrasts was provided in the CSWL task (Reference Escudero, Smit and MulakEscudero et al., 2022). This possibility may have been enhanced by the words in the study being produced in infant-directed speech (IDS) because its properties can facilitate the learning of phonetic contrasts in adults and children (Reference Escudero and WilliamsEscudero & Williams, 2014; Reference Golinkoff and AliotoGolinkoff & Alioto, 1995; Reference Graf Estes and HurleyGraf Estes & Hurley, 2013).Footnote ¹¹ However, since IDS has more variable pitch than adult-directed speech across languages (Reference Igarashi, Nishikawa, Tanaka and MazukaIgarashi et al., 2013), the English words produced in IDS likely sounded as though they had different lexical tones to L1 Mandarin ears, challenging their word-referent mappings. A similar effect has been found by Reference Smit, Milne and EscuderoSmit, Milne, and Escudero (2022), in which participants’ music perception abilities negatively influenced their learning of English vowel minimal pairs via CSWL, presumably because of their enhanced sensitivity to pitch variations in vowels. Reference Escudero, Smit and MulakEscudero et al. (2022) explained that hearing tones in the English words could have resulted in these sequential bilinguals’ MCA of the vowels in the words, leading to a Subset scenario with spurious lexical contrasts and poorer performance. These findings suggest that not only segmental but also suprasegmental details should be considered in predicting and explaining vowel perception and word learning (Reference Escudero, Mulak, Elvin and TraynorEscudero et al., 2018; Reference Escudero and KalashnikovaEscudero & Kalashnikova, 2020).

8.4.1 The Role of Orthography

Many studies have shown that orthography influences speech processing in bilinguals (see Reference Bassetti, Escudero and Hayes-HarbBassetti, Escudero, & Hayes-Harb, 2015). Studies within L2LP have shown that the availability of orthographic forms as input to bilinguals can have various influences on speech perception and word learning, both facilitative and impeding (Reference EscuderoEscudero, 2015; Reference Escudero, Hayes-Harb and MittererEscudero et al., 2008, Reference Escudero, Simon and Mulak2014a; Reference Escudero and WanrooijEscudero & Wanrooij, 2010). For instance, Reference Escudero, Simon and MulakEscudero et al. (2014a) demonstrate that congruence between L2 learners’ orthographic systems influences performance in word learning, as the orthography of the learners’ dominant language is activated when reading L2 words (Reference EscuderoEscudero, 2015; Reference Escudero, Hayes-Harb and MittererEscudero et al., 2008). For both prelexical perception and lexical recognition, it has been shown that when the learners’ two orthographic systems match, learning is facilitated, but when they do not, learning is more challenging. Also, CSWL is more accurate for both monolinguals and bilinguals when words are presented orthographically rather than auditorily, suggesting that visual information facilitates unintentional and ambiguous word learning (Reference Escudero, Smit and AngwinEscudero et al., 2023). Thus, the role of orthography is clearly prominent in bilingual phonetics and phonology.

The L2LP model assumes that bilinguals’ mental lexicon contains phonological and orthographic representations of speech based on much previous research attesting the role of orthography in bilingual speech processing (Reference EscuderoEscudero, 2015; Reference Escudero, Smit and AngwinEscudero et al., 2023; Reference Escudero and WanrooijEscudero & Wanrooij, 2010). However, how exactly orthography fits within L2LP’s architecture (Figure 8.1) is yet to be formally modeled and computationally simulated. Ongoing research aims at bridging this gap.

8.4.2 Speech Production

There have also been attempts to extend L2LP to speech production (Reference Elvin, Vasiliev, Escudero, Gibson and GilElvin et al., 2018b; Reference Elvin, Williams, Escudero, Molsing, Becker Lopes Perna and Tramunt IbañosElvin, Williams, & Escudero, 2020), as the model claims that perception precedes production and is a prerequisite for the development of production skills (Reference Escudero and PenningtonEscudero, 2007, p. 110). Unlike other models of L2 speech acquisition that predict no “mastery” of L2 production, L2LP’s Ingredient 5 predicts that L2 learners can ultimately attain optimal perception (and, by extension, production). To test this hypothesis for production, Reference Yazawa, Konishi, Whang, Escudero and KondoYazawa et al. (2023a) examined 102 adult L1 Japanese speakers’ productions of L2 American English monophthongs using an L2 English speech corpus called J-AESOP (Reference Kondo, Tsubaki and SagisakaKondo, Tsubaki, & Sagisaka, 2015). All learners were late sequential bilinguals who had been learning English since the age of thirteen in Japanese schools and had never lived outside Japan. Despite the uniform linguistic background, the learners exhibited diverse levels, with some (if not most) showing near-native-like productions across all vowel categories, regardless of the perceptual similarity between particular L1 and L2 sounds. The result is consistent with L2LP’s prediction and provides a promising extension of the model to speech production.

Reference Liu and EscuderoLiu and Escudero (2023) applied L2LP’s predictions to the influence of dialectal variation in production, finding that those who speak two dialects of the same L1 had overall better performance in L2 production tasks than those who speak only one L1 dialect, despite their similar L2 learning backgrounds. This implies that the divergent performance of different types of bilinguals (Section 8.3.2) extends to “bidialectal” populations, confirming L2LP’s proposal to focus on each specific variety of a language and suggesting that the proposed inhibitory control advantages may also apply to the control of two dialects (Section 8.2.2). Further research can help to better understand how bilingualism and bidialectalism compare.

8.4.3 Applying L2LP to Language Training and Curriculum Design

Finally, L2LP’s theoretical proposals have significant implications for language learning and training, as detailed by Reference Elvin, Escudero, Gutierrez-Mangado, Martínez-Adrián and Gallardo-del-PuertoElvin and Escudero (2019). Specifically, its ingredients can be used to identify the specific difficulties learners may have because of cross-linguistic/dialectal differences and to predict their further development. The following are just a few examples of how L2LP can be applied to language training and curriculum design.

Many studies within L2LP have capitalized on the distributional nature of perceptual learning to demonstrate that difficult phonetic contrasts can be accurately perceived through very short exposure to the most frequent sound exemplars of a phonetic continuum (Reference Escudero, Benders and WanrooijEscudero et al., 2011). It has been shown that distributional training can enhance the perception of difficult vowel and tone contrasts (Reference EscuderoOng, Burnham, & Escudero, 2015), that individual differences prior to training modulate success (Reference Wanrooij, Escudero and RaijmakersWanrooij, Escudero, & Raijmakers, 2013), and that Similar contrasts are easier to train than New contrasts (Reference Chládková, Boersma and EscuderoChládková, Boersma, & Escudero, 2022, also for a review of neurphysiological studies testing L2LP). Reference Escudero and WilliamsEscudero and Williams (2014) also demonstrated that the effects of distributional learning in adult L2 learners can last a long time, as its effects remained for more than a year after training.

The CSWL task can be used to teach real words to L2 learners at different developmental stages. Reference Tuninetti, Mulak and EscuderoTuninetti et al. (2020) show that learners can easily learn twelve to eighteen words within a learning session, suggesting that this paradigm could be quite successful for classroom learning or self-paced learning at home, as suggested in Reference Escudero, Smit and AngwinEscudero et al. (2023).

Regarding L2 production training, Reference Colantoni, Escudero, Marrero-Aguiar and SteeleColantoni et al. (2021) devised innovative perception and production exercises for beginner university-level L2 learners of Spanish. The proposed teaching materials are based on key principles such as a focus on features with high functional load and shared by most varieties of the target language, which are direct applications of L2LP.

9.5.1 Neurobiology

An ideal method for studying the brain mechanisms underlying speech perception is via the electroencephalogram (EEG). The EEG reflects summation of excitatory and inhibitory postsynaptic potentials generated when neurons fire, and can be measured at the scalp. The timing and the topography (location on the scalp) of EEG modulations to a stimulus allow inferences about underlying brain mechanisms. Averaging portions of the EEG that are time-locked to a stimulus of interest, called event-related potentials (ERPs), is a common method used to isolate neural processes associated with the stimulus (Reference LuckLuck, 2014). Non-time-locked modulations can be examined using approaches such as time-frequency analysis (Reference Wagner, Ortiz-Mantilla and RusiniakWagner et al., 2022).

Studies using electrocorticography (ECoG) (undertaken in patients with intractable epilepsy) suggest that regions in the superior temporal gyrus (STG) are likely to be the locus of language-specific representations that form the basis of SPRs (Reference Yi, Leonard and ChangYi, Leonard, & Chang, 2019). These studies have identified STG regions that correspond to feature categories such as obstruent/sonorant, manner of articulation, place of articulation, and voicing (Reference Moerel, De Martino and FormisanoMoerel, De Martino, & Formisano, 2014; Reference Steinschneider, Nourski and FishmanSteinschneider, Nourski, & Fishman, 2013; Reference Yi, Leonard and ChangYi et al., 2019).

The ERPs used to study speech perception include auditory evoked potentials (AEPs), which are obligatory responses to acoustic input that occur in the first 200 ms from onset and have sources that are likely to correspond to those identified with ECoG methods. The AEPs named P1-N1-P2 are recorded at frontal-central sites (with a source in the superior plane of STG) and T-complex AEPs are recorded over lateral sites (presumed to have sources in the lateral plane of STG). Some studies suggest no effect of language-specific experience on AEPs (Reference Hoonhorst, Serniclaes and ColletHoonhorst et al., 2009; Reference Sharma and DormanSharma & Dorman, 2000); however, other studies do show modulation related to language experience (Reference Rinker, Shafer, Kiefer, Vidal and YuRinker et al., 2017; Reference Wagner, Shafer, Martin and SteinschneiderWagner et al., 2013).Footnote ³ Language-specific modulation of AEPs might be due indirectly to differences in how monolinguals and bilinguals allocate attention. For example, bilinguals compared to monolinguals showed greater attention to nontarget vowels (seen as a negative shift at the N1 and P2 latencies, called the Nd) in a task where the targets were tones or consonants (Reference Datta, Hestvik and VidalDatta et al., 2020) (see Figure 9.1). To date, too few cross-linguistic studies of AEPs to speech have carefully matched groups for bilingual experience; matching of experience would allow determining whether these effects are due to differences in how bilinguals and monolinguals deploy attention or to different tuning of phonological representations.

Figure 9.1 Auditory evoked potentials to a 250-ms vowel /ɛ/ at the onset of a sequence of ten stimuli (intersequence interval 1,250 ms) for monolingual AE adults (N = 14) and five to seven-year-old children (N = 14) and bilingual Spanish-English adults (N = 7) and five to seven-year-old children (N = 13) (both languages learned before age five years). The upper-right graph shows the adult P1-N1-P2 sequence at Fz and the upper-left graph exhibits the child P100-N2 sequence. The increased negativity (Nd shift) in the ERP of bilingual adults indicates allocation of attentional resources to processing the stimuli, even when instructed to ignore the auditory stimuli (see Reference Datta, Hestvik and VidalDatta et al. [2020]). The bottom-right graph shows the T-complex measures for left and right electrode sites (1 cm above the ears). The increased positivity of the Ta for bilingual adults may be the polarity reversal of the frontal Nd effect. The bottom-left graph shows the maturing T-complex, with the positive-going Ta peak beginning to emerge (see Reference Shafer, Yu and WagnerShafer, Yu, & Wagner [2015]). The bilingual children show more negative T-complex responses (see Reference Rinker, Shafer, Kiefer, Vidal and YuRinker et al. [2017]).

Neural mechanisms of speech perception have also been studied using attention-dependent measures, including N2b and N400, and positivities (P300/P3b). These measures reflect engagement of goal-directed, higher-level processes that are dependent on conscious perception, and that are expected to show a stronger correlation with behavioral perception than attention-independent measures (Reference Dehaene-Lambertz, Dupoux and GoutDehaene-Lambertz, Dupoux, & Gout, 2000; Reference Wagner, Shafer, Martin and SteinschneiderWagner et al., 2012). Task-related potentials can be used to track the underlying processes leading up to the behavioral response. For example, a common paradigm used to focus attention on the speech signal is an oddball task in which a frequent category is repeated and listeners are asked to respond to a rare (oddball) category (Reference Buchwald, Guthrie, Schwafel, Erwin and Van LanckerBuchwald et al., 1994; Reference García and FroudGarcía & Froud, 2018; Reference Xi, Xu and ZhuXi et al., 2021). Behavioral detection of the oddball elicits an increased negativity around 200 ms called the N2b followed by a large positivity (P300 or P3b) peaking between 300 and 600 ms. Easier discriminations allow for larger and earlier peak responses. In the absence of attention to the stimuli, a different measure called mismatch negativity (MMN) is elicited, as explained next.

The MMN ERP measure is ideal for testing the ASP model’s claim that native-language SPRs are automatic because: (1) it is sensitive to language experience; and (2) it can be examined when attention is directed away from the auditory stimulus. In the oddball paradigm, MMN is elicited to a rare event (deviant), and results in increased negativity at fronto-central sites (and positivity at mastoid sites), if the stimulus contrast is sufficiently salient (Reference Näätänen, Kujala and WinklerNäätänen, Kujala, & Winkler, 2011). Further, MMN is larger in amplitude and earlier in latency with increasing contrast salience. The sources of MMN are localized to the STG, but additional sources are sometimes found in the frontal cortex and hypothesized to reflect activation of the attentional network (e.g., orienting to a salient sound change). It is possible for MMN to be partially or wholly masked by other components, such as the N2b in tasks where attention is focused on detecting the stimulus change. Stable patterns can also be consolidated as an auditory representation, and deviation from the pattern will lead to MMN (Reference Sussman, Ritter and VaughanSussman, Ritter, & Vaughan, 1998). Figure 9.2 (right column) displays the MMN component in adults.

Figure 9.2 Mismatch responses (deviant minus standard) to the 250 ms vowel /ɛ/ (ISI 400 ms) from the same participants as in Figure 9.1. The MMN peak is 18 ms later for bilingual than for monolingual adults. Monolingual children show a pMMR followed by an MMN that is about 110 ms later than seen for adults. The bilingual children show a pMMR followed by a small MMN deflection.

More than a quarter-century of research has revealed that MMN is modulated by native-language experience (Reference Mah, Goad and SteinhauerMah, Goad, & Steinhauer, 2016; Reference Näätänen, Lehtokoski and LennesNäätänen et al., 1997). Reviewing this literature is beyond the scope of this chapter, but the following studies illustrate the utility of MMN in testing the ASP model.

The first studies using MMN to test the ASP model examined neural discrimination of Japanese vowel and consonant duration contrasts under different conditions of attention in native AE and native Japanese listeners (Reference Hisagi, Shafer, Strange and SussmanHisagi et al., 2010; Reference Shafer, Yu and WagnerHisagi, Shafer, et al., 2015). We predicted that AE listeners would show a smaller MMN when attention was focused away from the speech compared to attention focused on the speech. The AE listeners showed a smaller MMN to the vowel duration contrast when attention was focused on a visual shape task, but a comparable MMN to Japanese listeners when attention was directed to the auditory modality (by asking them to count vowel deviants), confirming the prediction.Footnote ⁴ For both groups, we also observed a large P3b response to the duration change when counting deviants, indicating behavioral detection of the oddball (Reference Hisagi, Shafer, Strange and SussmanHisagi et al., 2010). The findings for the consonant contrast, however, suggested that salience was important for native listeners, as well as non-native listeners (Reference Shafer, Yu and WagnerHisagi, Shafer, et al., 2015).

In another study, we examined L2 learners of English from Japanese, Spanish, or Russian backgrounds to further test whether listeners rely on L1 SPRs in the absence of focused attention (Reference Shafer, Kresh and ItoShafer et al., 2021). Neural discrimination of AE low vowels /ʌ/, /ɑ/, and /æ/ was tested in a paradigm using three natural tokens of each vowel (e.g., ɑpə1, ɑpə2, ɑpə3, ʌpə2, ɑpə1, ɑpə2, ɑpə3, æpə3, etc.; deviants are in bold). These three vowels are challenging to L2 learners whose language includes only one low vowel (represented by /a/), because their spectral F1 and F2 patterns are acoustically similar (compared to, e.g., /a/ versus /e/). The AE /ʌ/, however, is systematically shorter than /ɑ/ and /æ/, which provides an additional cue for listeners. Japanese listeners were expected to access this temporal cue at the automatic level because it serves as a primary cue in Japanese (Reference Strange, Akahane-Yamada and KuboStrange et al., 1998). The Spanish listeners were expected to show poor discrimination of all low-vowel pairings because Spanish does not distinguish vowel duration. The expectation was less clear for the Russian group because duration is used for stress. In addition, Russian listeners sometimes behaviorally judge /æ/ as more similar to Russian /e/ than /a/, which suggests that they are selecting different cues than Spanish listeners, who judge /æ/ to be most similar to Spanish /a/ (Reference StrangeStrange, 2011).

Spanish listeners did not reveal an MMN to the /ʌ/ versus /ɑ/ contrast when /ʌ/ was the standard; both the Japanese and the Russian listeners showed an MMN to this contrast, although later in latency than that of the English listeners. This result supported the prediction that Spanish listeners’ SPRs for vowels lacked duration information, and are insufficiently detailed in spectral information to support distinguishing /ɑ/ and /ʌ/. In contrast, Japanese listeners’ native-language SPRs included fine-grained temporal information. The finding of an MMN for Russian listeners suggests that SPRs for their native-language stress patterns facilitated discrimination. The later latency of the MMN for the Japanese and the Russian listeners than for the English listeners is consistent with reliance on the duration cue (which requires waiting for the end of the shorter stimulus), whereas the English speakers were able to discriminate the vowels on the basis of spectral differences from stimulus onset.

An asymmetry in discrimination provided deeper insight into the nature of SPRs. When /ɑ/ served as the standard and /ʌ/ as the deviant stimulus, all groups showed an MMN, suggesting that neural discrimination was easier for this presentation order. One explanation for this finding is that the repeated stimulus triggers the native-language SPR in long-term memory (without focused attention); in contrast, the deviant is registered as an iconic memory, which decays rapidly. Under this view, Spanish listeners’ representation for the repeated stimulus (standard) was Spanish /a/, whether the phonetic identity was AE /ʌ/ or /ɑ/; thus, their perception was /a-a-a/-[ɑ] for the phonetic sequence [ʌ-ʌ-ʌ-ɑ], but /a-a-a/-[ʌ] for the phonetic sequence [ɑ-ɑ-ɑ-ʌ]. In this latter case, the iconic memory trace of /ʌ/ as the deviant is sufficiently different (i.e., it is a poor exemplar of Spanish /a/) to generate the MMN. Increasing the ISI to induce greater decay of the deviant would serve to test this claim. For example, in a different study of Mandarin lexical tone contrasts, naïve AE listeners showed an MMN when the ISI was relatively short (600 ms), but not at a longer ISI (2,700 ms), whereas native Mandarin listeners showed MMN at both ISIs (Reference Yu, Shafer and SussmanYu et al., 2017).

These findings support the claim that listeners fall back on native-language SPRs in the absence of focused attention; however, several studies indicate that attentional resources may be needed to discriminate some native-language contrasts that are phonetically similar. Contrasts that are considered “marked” because they are phonetically similar are the most likely to need attentional resources. The Hindi dental-retroflex contrast and the Japanese short versus long consonant contrast appear to fall in this category (Reference Shafer, Yu and WagnerHisagi, Shafer, et al., 2015; Reference Shafer, Schwartz and KurtzbergShafer, Schwartz, & Kurtzberg, 2004). For the Japanese consonant duration difference, MMN was larger for both naïve English and native Japanese listeners in an auditory attend compared to a visual attend condition (although only Japanese listeners showed an MMN without attention to the auditory deviant). Another study by Reference Barrios, Namyst, Lau, Feldman and IdsardiBarrios et al. (2016) failed to find a larger MMN for Spanish listeners to a contrast that was phonemic in Spanish but allophonic in English compared to a contrast that was allophonic in Spanish but phonemic in English.

9.5.2 Neurodevelopmental Studies

The prior section demonstrated that MMN serves as an excellent measure for probing the nature of phonological representations stored in the brain. This knowledge allows us to now interpret the neural processes supporting speech perception development. Specifically, presence of MMN can be used to indicate whether and when in development phonological representations are sufficiently detailed to support discrimination. This section presents several key studies to illustrate what we have learned so far from neurodevelopmental measures.

First, it is crucial to recognize that brain regions mature at different rates in the developing child, and, thus, speech perception is likely to be accomplished differently across development (Reference Werker and CurtinWerker & Curtin, 2005). Studies of infant auditory development reveal relatively mature development of the peripheral and brainstem auditory systems; in contrast, neocortical regions are still immature at birth (Reference Moore and LinthicumMoore & Linthicum, 2007). Newborns can encode and discriminate speech, but the neural systems supporting auditory processing at the cortical level are immature in density of neurons, connectivity (through dendrites and synapses), and speed of processing (through myelination of axons, which connect neurons to each other). This immaturity is also apparent in neurophysiological measures used to test speech encoding and discrimination (e.g., Reference Shafer, Yu and WagnerShafer, Yu, & Wagner, 2015; Reference Yu, Tessel and HanYu et al., 2019). Cortical immaturity leads to the hypothesis that only very large acoustic-phonetic differences can drive attention-independent discrimination in newborns (Reference Moore and LinthicumMoore & Linthicum, 2007). In addition, infants have immature prefrontal lobes (supporting executive functions). In the first few months of life, infants are unable to voluntarily shift attention away from a highly salient or novel stimulus, such as a large sound change (Reference Hendry, Jones and CharmanHendry, Jones, & Charman, 2016). Thus, we predict that only large acoustic differences support an automatic discrimination process in the youngest infants, which will be indexed by MMN. This prediction is generally held up in infant studies examining neural measures of discrimination.

The first infant studies of neural speech discrimination observed a very large, sustained negativity to a Finnish /y/ versus /i/ contrast in newborn infants; the authors argued that this finding supported early maturation of the speech discrimination process reflected by MMN (Reference Cheour-Luhtanen, Alho and KujalaCheour-Luhtanen et al., 1995). Subsequent studies, however, suggest that MMN is observed in young infants only in relation to large acoustic differences. To smaller differences, infants showed an increased positivity to the deviant stimulus (called the positive mismatch response or pMMR), rather than increased negativity. This pMMR has been observed in relation to both tones and speech contrasts (Reference Dehaene-Lambertz and DehaeneDehaene-Lambertz & Dehaene, 1994; Reference Leppänen, Richardson and PihkoLeppänen et al., 2002; Reference Morr, Shafer, Kreuzer and KurtzbergMorr et al., 2002; Reference Trainor, Samuel, Desjardins and SonnadaraTrainor et al., 2001). Subsequent studies showed that MMNs in relation to subtle consonant and vowel contrasts are not consistently present in children until at least four years of age (Reference Lee, Yen and YehLee et al., 2012; Reference Shafer, Yu and DattaShafer et al., 2010). A negative response (that may be the MMN)Footnote ⁵ appears only to large acoustic differences, (e.g., 1,000 Hz vs. 2,000 Hz pure tones: Reference Morr, Shafer, Kreuzer and KurtzbergMorr et al., 2002; lexical tone: Reference Lee, Yen and YehLee et al., 2012; Reference Liu, Ong, Tuninetti and EscuderoLiu et al., 2018). In addition, the pMMR and the MMN can be elicited in the same children, with the pMMR preceding the MMN in time (Reference Shafer, Yu and DattaShafer et al., 2010). These studies also indicate slower neural discrimination by younger children. That is, MMN shifts earlier in time by about 10–12 ms per year from four to ten years of age and may not reach adult latencies until after puberty (Reference Datta, Hestvik and VidalDatta et al., 2020; Reference Shafer, Yu and DattaShafer et al., 2010). Figure 9.2 (left column) displays pMMR and MMN in children.

These developmental patterns suggest that pMMR indexes the magnitude of acoustic differences rather than phonological differences, and that it is generated by recovery from refractoriness of a neural population firing to the acoustic characteristics of the deviant (Reference May and TiitinenMay & Tiitinen, 2010).Footnote ⁶ Studies have shown that repetition of a stimulus leads to adaptation (suppression of neuron firing). Longer intervals between repetitions allow neurons to recover. The longer interval between deviants compared to standards allows for greater recovery, seen as greater positivity to the deviant stimulus for children. We hypothesize that the pMMR reflects an iconic memory mechanism.

Several studies with infants suggest that fine-grained phonetic differences initially need focused attention for discrimination. We observed that six-month-olds showed only a negative mismatch response (nMMR) to the fine-grained /ɛ/ versus /ɪ/ vowel distinction under conditions manipulating attention to the stimulus change (Reference Shafer, Yu and Garrido-NagShafer, Yu, & Garrido-Nag, 2012). For example, in Karen Garrido-Nag’s dissertation (Reference Garrido-NagGarrido-Nag, 2013), infant attention was manipulated by pairing a stimulus change with a reward (a smiling face). The MMRs to this “contingent” stimulus change were more negative than in another condition where the smiling face was not contingent on the change. Infants as young as three months of age appear to have sufficiently mature attentional networks (top-down feedback from frontal regions) to support the allocation of attention in speech perception (Reference Mersad, Kabdebon and Dehaene-LambertzMersad, Kabdebon, & Dehaene-Lambertz, 2021).Footnote ⁷

We argued that increasing negativity of MMRs to a vowel contrast /ɛ/ versus /ɪ/ with increasing age could be explained by infants’ representations becoming more detailed, and, thus, allowing discrimination whether attention was focused on or away from the speech stimuli (Reference Shafer, Yu and DattaShafer, Yu, & Datta, 2011). A study from Patricia Kuhl’s lab supports this explanation (Reference Garcia-Sierra, Ramírez-Esparza and KuhlGarcia-Sierra, Ramírez-Esparza, & Kuhl, 2016). They tested eleven to fourteen-month-old children from monolingual English and bilingual Spanish-English households and measured the amount of input in English and Spanish (by recording speech in the household and deriving the number of words in forty 30 s intervals). Monolingual children with high English input showed a significant negative MMR to an English contrast /pa/ versus /ta/, but a pMMR to a non-native Mandarin Chinese alveolo-palatal affricate (/tɕ^hi/) versus an alveolo-palatal fricative (/ɕi/). No MMR effects (positive or negative) were observed in relation to a Spanish prevoiced /da/ versus short-lag voiceless /ta/ contrast. Spanish-English bilingually exposed infants with high English or high Spanish input showed pMMR rather than negativity to their respective language voiced/voiceless contrast. This absence of negative MMR in bilingual infants may be explained by insufficient input in either language (the “high” input values for bilinguals matched low-input values for monolingual infants), or to bilingual experience leading to interference in setting up native-language representations (described as delayed neural commitment by Reference Garcia-Sierra, Ramírez-Esparza and KuhlGarcia-Sierra et al., 2016).

In summary, these neurodevelopmental studies demonstrate that MMRs provide a method for examining the nature of phonological representations in L2 learning. Section 9.5.3 reveals how research using the MMR method provides additional insight into the nature of SPRs for bilinguals who have learned two languages early in life.

9.5.3 Early Bilinguals

For bilingual listeners, conflicts will exist between the two language phonologies for some phonological representations. Bilinguals who have learned both languages before five years of age are particularly interesting because differences in speech perception compared to monolingual listeners cannot be attributed to closure of a critical/sensitive period for refining speech sound categories (Reference Reh, Arredondo and WerkerReh, Arredondo, & Werker, 2018).

Three different possibilities exist for bilingual speech perception (Reference WilliamsWilliams, 1977):

1. 1. Listeners show a bias toward one language (Language A); in this case, the “explicit” phonological representations will exist for Language A and allow automaticity of SPRs, but SPRs for Language B will require some attentional resources.

2. 2. Listeners compromise between the two languages; in this case, the phonological representations (SPRs) for both Language A and Language B differ from those used by monolingual listeners.

3. 3. Listeners can shift between the two systems by means of some cue that allows selecting language-specific SPRs; however, some attentional resources may be necessary for selecting the language.

Neurophysiological evidence can be found to support these three scenarios. For example, Reference Peltola, Tamminen, Toivonen, Kujala and NäätänenPeltola et al. (2012) found that proficient Finnish-Swedish bilinguals who were dominant in Finnish showed a larger MMN to a vowel change in the environment where it is phonologically contrastive (that is, Finnish context), but an attenuated MMN in the context where the vowel pair falls within the same category (Swedish). In contrast, balanced bilinguals, who had learned both languages from birth, showed evidence of compromising, with a smaller and a later MMN observed for both the Finnish and the Swedish contexts. The finding of difference in attention directed to speech between monolingual and bilingual adults suggests that bilinguals may be monitoring the speech signal for cues to determine which language to select (e.g., Reference Datta, Hestvik and VidalDatta et al., 2020) (see Figure 9.1). An Nd-attention difference was not observed between monolingual and bilingual children in Reference Datta, Hestvik and VidalDatta et al. (2020), but it may be that children have not yet developed this strategy.

10.2.1 Chronological Corollary of the OPM

The first corollary addresses overall chronological development. It claims that IL develops chronologically in the following manner: L2 increases, L1 decreases, and U increases and then decreases (Figure 10.1).

Figure 10.1 Chronological corollary of the OPM.

At the initial stage the learner speaks an L1, no L2, while U remains dormant; thus, L2 is essentially a tabula rasa. Therefore, IL = L1. One could claim that someone with the heaviest of foreign accents is not speaking the L2 at all, but merely speaking the L1 employing L2 loanwords. During acquisition, the L2 system gradually develops through various intermediate stages until it reaches the final idealized state when the L2 is completely mastered; that is, IL = L2. Very few, if any, learners reach this advanced stage (but see Reference MoyerMoyer’s [2021] “gifted language learner”). The chronological corollary pertains to normal phenomena that are neither similar nor marked. The stylized hypothetical in Figure 10.1 and those that follow characterize the idealized learner, but necessarily will vary with different learners and different phenomena. The relative frequencies of the L1, the L2, and U vary and the specific frequencies of these three factors are hypothetical.

The OPM claims that at the early stages, L1 influence is so strong that it prevents L2 mastery and U from surfacing. With continued exposure and improvement in perception,Footnote ² the learner realizes (consciously or unconsciously) that L1 substitutions are not accurate productions of the L2. Subsequently, the learner’s continuing but failed attempts to accurately produce the L2 cause this latent U to become activated, resulting in phenomena that are part of neither the L1 nor the L2. Thus, at an early stage, L1 obliterates the effect of U; hence, nothing in U is evident that is not already part of the L1; that is, U remains dormant. However, at a later stage, this nascent U awakens and replaces some L1 components. Then, still at a later stage, U decreases as the L2 replaces it.

Logically, the OPM should be true in light of widely observed IL patterns. The learner begins acquisition with 100 percent L1 and 0 percent L2. At the final stage the idealized learner’s IL is 0 percent L1 and 100 percent L2. Chronologically, this means L2 must increase and L1 decrease. Because at the beginning stage L1 is 100 percent and at the final stage L2 is 100 percent, it follows that U must be zero at both the initial and the final stages. However, U is also a well-documented factor (Reference Altenberg and VagoAltenberg & Vago, 1983; Reference JohanssonJohansson, 1973; Reference MusauMusau, 1993; Reference NemserNemser, 1971a). Therefore, U must rise and fall at stages other than at the very beginning or the very end (when it is zero) because mathematically L1 + L2 + U = 100 percent.

Although L1 and L2 acquisition involve some principles that are distinct from other forms of learning, all types of learning do share one common characteristic: transfer. In the early stages, the underlying cause for transfer being much more important than U follows from long-known principles of numerous learning theories of psychologists (Reference AusubelAusubel, 1963; Reference Ausubel, Novak and HanesianAusubel, Novak, & Hanesian, 1978; Reference BugelskiBugelski, 1942; Reference GagnéGagné, 1977). Extensive research has demonstrated that one relies on previous cognitive structures when learning new structures; that is, transfer occurs. Reference Ausubel, Novak and HanesianAusubel et al. (1978, p. 65) further claimed that all learning involves transfer and that subsequent learning “results in new transfer by modifying cognitive structure.”Footnote ³ Incorporating this research into an SLA framework, their conclusions predict that L1 transfer will dominate the early stages because the learner has mastered very little L2. As acquisition proceeds, the existing cognitive structure (the IL) is modified by the L2 experience, creating new cognitive structures. However, because the structures are new, they could not be L1, but necessarily are either U or L2. These new structures in turn cause further IL modifications. Therefore, as acquisition progresses, this new IL will show an increasing U and L2, while the influence of pure L1 transfer continues to decrease. Subsequently, as learning continues further, U substitutions will be replaced by native L2 forms, resulting in a decrease in U. These patterned progressions are precisely the OPM’s claims: L1 will decrease over time, L2 will increase, whereas U will first increase and then later decrease.

Consider the acquisition of English terminal coda voiced obstruents in NSs of Japanese and Portuguese (non-occurrent in either language). If transfer operates first, learners will epenthesize a vowel. However, if and when epenthesis ceases, devoicing is able to surface. Thus, these hypothetical stages for the acquisition of dog would be: do[gu] → do[k] → do[g] (L1 Japanese) and do[gi] → do[k] → do[g] (L1 Portuguese).

Although some substitutions are unmistakably either L1 or U, others can be both. Coda obstruent devoicing in Japanese learners of English is clearly U because Japanese has no coda obstruents. In contrast, this process in German speakers of English could be attributed both to U and to L1 because it is obligatory in native German. A number of researchers have proposed that substitutions persist longer when both transfer and U produce the same result (Reference Andersen, Gass and SelinkerAndersen, 1983; Reference Hecht and MulfordHecht & Mulford, 1982).Footnote ⁴ This claim follows logically from the OPM: In German speakers of English, L1 first causes devoicing to occur; subsequently, when learners eliminate L1 transfer, U surfaces. However, because both L1 and U cause devoicing, the process persists. A possible test case would be to compare devoicing in Japanese and German learners of English. The prediction would be that devoicing would persist longer in German NSs because it occurs in native German but not in native Japanese.

An early longitudinal study that provides evidence for the OPM is Reference WodeWode’s (1981) research on the L2 acquisition of English by his four German-speaking children (over a period of approximately six months, beginning at ages 3;11 to 8;11). Wode proposes gradual approximations and discrete jumps. The vowels [a] and [ɛ] were first substituted for English /ʌ/ and /æ/, indicating transfer from German. Later the substitutions became closer and closer to native English via gradual approximations. These gradual approximations can be attributed to U because they are neither native English nor native German. In contrast, discrete jumps occurred in r in all four children: [R] > [w] > [ɹ] > [ṛ] (Wode’s [ɹ] = “central frictionless continuant” and [ṛ] = “target-like retroflex”). The stages for vowel and r substitutions follow from the OPM: early substitutions were due to transfer, later substitutions were due to U, and, finally, L2 was mastered.

Reference AbrahamssonAbrahamsson’s (2003) longitudinal study further supports the OPM, where he reported deletion and epenthesis increased and then decreased in Chinese learners of Swedish. Other longitudinal studies provide evidence for the OPM, including English consonant cluster acquisition by Brazilian Portuguese speakers (Reference MajorMajor, 1994, Reference Major, Bayley and Preston1996) and Spanish r acquisition by English speakers (Reference MajorMajor, 1986a). In all three studies, the longitudinal patterns show an increase in L2, a decrease in L1, and an increase and then a decrease in U. For example, in the Spanish study, some speakers exhibited the following stages for /r/ acquisition: [ɹ] > [ɾ] > [ʀ] > [ʐ]Footnote ⁵ > [r].

For the past three decades, Optimality Theory (OT; Reference Prince and SmolenskyPrince & Smolensky, 1997, Reference Prince and Smolensky2004) has been a prominent theory in mainstream phonology. In OT terminology, the OPM can be formulated as follows: L1 constraint rankings comprise the early stages; later they are reranked into non-native L1 and non-native L2 (i.e., U); and finally they are reranked into native L2.Footnote ⁶ Reference Hancin-Bhatt and BhattHancin-Bhatt and Bhatt (1997), employing OT, found some substitutions due to language-specific rankings (L1 transfer) and others due to language-independent rankings (U). Reference Hancin-Bhatt and BhattHancin-Bhatt and Bhatt (1997, p. 368) argue that

10.2.2 Stylistic Corollary of the OPM

The second corollary deals with style. It predicts that as style becomes more formal, L2 increases, L1 decreases, and U increases and then decreases (Figure 10.2).

Figure 10.2 Stylistic corollary of the OPM.

It should be noted that the shapes of the graphs for chronology and style are identical (Figures 10.1 and 10.2). The L1 pattern follows logically from widespread research demonstrating that transfer is most prominent in casual styles but least prominent in formal styles, such as citation (Reference Dickerson, Dickerson, Corder and RouletDickerson & Dickerson, 1977; Reference NemserNemser, 1971b; Reference SchmidtSchmidt, 1977; Reference TaroneTarone, 1988; Reference Wilson and MøllergardWilson & Møllergard, 1981). Probably nearly every language teacher would agree that L2 speakers’ pronunciation of isolated words is usually much better than in conversation, where L1 transfer is more noticeable. A hypothetical example illustrates this. By changing the tone in Mandarin, the sequence /ma/ has four meanings (“mother,” “horse,” “hemp,” “scold”). An NS of English who studies Mandarin for six months and practices daily mā 妈 “mother” may shout “Mother!” and successfully sound native for this one word while not mistakenly calling her a horse, mă 马; nevertheless, it is highly probable that the same person will have a prominent L1 English accent in conversation.

Although L1 transfer decreases and L2 increases as style becomes more formal, this relationship does not overtly predict the rise and fall of U. However, this is implicit, because any IL component that is neither L1 nor L2 is by definition U. In the idealized or extreme case, a very formal style is pure L2 and a very casual style is pure L1. Therefore, as style changes from formal to casual, the U component has to appear and then disappear; that is, it increases and then decreases.

Reference WodeWode’s (1981, p. 228) study of his daughter Birgit’s L1 German-L2 English acquisition of /ɹ/ supports the stylistic claims of the OPM (Wode’s /r/ = IPA /ɹ/): “Until she got out of school at the beginning of June, she much preferred [R] as a substitute for [ṛ]]/[ɹ]in her casual spontaneous speech. In the imitation-like checkups she would frequently produce or attempt [w] or something [w]-like to substitute for the L2 /r/.” These “checkups” where she produced U substitutions ([w] or something [w]-like)Footnote ⁷ can be considered more formal than “casual spontaneous speech,” when she used L1 substitutions ([R]). These data thus support the OPM: L1 was greater in casual speech, but U was greater in more formal speech.

The patterns in Figure 10.2 may not hold true due to a number of factors, such as nervousness, comfort level, and competence in a given style. A former student of Major’s reported that his Japanese girlfriend living in the United States sounded more native-like (though formal) when she was less guarded and more relaxed than when she consciously attempted to sound casual, a style that she had not learned in Japan. Furthermore, mastery of different styles implies competence in the appropriate situations (i.e., communicative competence). Reference MajorMajor (2004) investigated English casual speech processes in NSs of American English and NSs of Japanese and found that although some Japanese speakers had mastered casual speech processes, they often used them in a formal style, for example [gɑɑɑʧɑɑɑ] got you when slowly reading a list of two-word phrases.

Because different processes can occur in different styles, a thorough knowledge of stylistic variation in both the L1 and the L2 is critical for a comprehensive analysis. If productions are more native-like in a casual style, this first might appear to be counterevidence to the OPM. However, this is not necessarily true: Reference ZampiniZampini (1994) pointed out that positive transfer can occur in a casual style but may not in a formal style. For example, in Brazilian Portuguese, stressed and pretonic /il/ and /iu/ are normally pronounced [iw], but [ju] in very casual speech (e.g., f[iw]mamos → f[ju]mamos “we film,” Reference MajorMajor, 1981, Reference Major1985). Brazilians tend to pronounce English few as [fiw], but in very casual speech they accurately produce [fju]. Reference MajorMajor’s (1994, Reference Major, Bayley and Preston1996) studies of consonant clusters also showed that L1 casual processes correctly predicted L2 accuracy, yet L1 formal processes did not. Thus, L1 transfer is the governing factor in both these styles and has no relevance to the stylistic corollary.

When L2 learners are unaware that L2 sounds actually occur in their L1s, it can be to their disadvantage. Brazilian Portuguese has seven nasalized monophthongs and diphthongs, including /ʌ̃/, /ʌ̃j̃/, and /ʌ̃w̃/ (e.g., /χʌ̃/, /mʌ̃j̃/, and /mʌ̃w̃/ “frog,” “mother,” “hand”), which are usually very troublesome for English speakers. Nonetheless, unbeknownst to most learners, these three sounds in fact do occur in English. Huh is [hʌ̃] and in very rapid casual speech “what’s the matter, honey?” and “I don’t know” are often pronounced [smæɹ̩hʌ̃j̃] and [ʌ̃w̃]. Such an awareness could be quite useful in pedagogy.

Further examples demonstrate the relevance of stylistically conditioned processes in SLA analysis. In certain environments, Japanese and Portuguese /i/ and /u/ become voiceless and subsequently deleted. Thus, to English speakers, Japanese sukiyaki is perceived as [skijaki] and Portuguese lápis “pencil” is perceived as [laps]. However, curiously and fortuitously, for Japanese and Portuguese speakers of English, greater accuracy in English can occur in casual speech. Thus, in careful speech a Japanese speaker of English may say [sukaj] for sky but [skaij] in running speech. Likewise, a Brazilian carefully pronouncing English laps may say [lapis] but in fluent speech correctly produce the coda cluster [ps]. A further example from Japanese demonstrates how one process in the same style produces a native-like utterance in one word but not in another word. In a careful formal style, a Japanese speaker’s pronunciation of sky and city would be [sukay] and [siti] (assuming the transfer process /s/ → [∫]/ – [i] does not operate). However, in running speech the following utterance would be likely: I love a blue [skaj] in a big [sti].

10.2.3 Similarity Corollary of the OPM

The third corollary addresses similarity. The OPM claims that in similar phenomena, IL develops chronologically in the following manner: L2 increases slowly, L1 decreases slowly, and U increases slowly and then decreases slowly (Figure 10.3).

Figure 10.3 Similarity corollary of the OPM.

Slowly means more slowly than in normal phenomena. Compare Figure 10.1 to Figure 10.3. In similar phenomena, the role of L1 is much greater than U compared to normal phenomena. Although L1 predominates all acquisition in the early stages, in similar phenomena L1 persists longer than in normal phenomena, indicating that similar phenomena are more difficult to learn.

A caveat is the definition of similar and difficult (Reference ArchibaldArchibald, 2021; Reference Major and KimMajor & Kim, 1996; Reference Yang, Chen and XiaoYang, Chen, & Xiao, 2020).Footnote ⁸ Although there are problematic cases, there are clearly unambiguous cases: [k] is more similar to [q] than it is to [t] or [s]. However, perceptions of similarity are also governed by the native languages (NLs) of the learners. French speakers typically substitute /s/ and /z/ for English /θ/ and /ð/, while Portuguese speakers substitute /t̪/ and /d̪/, even though both French and Portuguese have /s/, /z/, /t/, and /d/; Hindi speakers of English substitute Hindi retroflex /ʈ/ and /ɖ/ for English alveolar /t/ and /d/, but English speakers perceive Hindi dental /t̪/ and /d̪/ being closer to native English; English speakers typically pronounce fondue as fond[u], while Brazilian Portuguese speakers pronounce it fond[i].

Psychologists have demonstrated that transfer operates only when there are relevant phenomena to transfer. Reference Ausubel, Novak and HanesianAusubel et al. (1978, p. 165) claimed that past experience has “impact on relevant properties of cognitive structure,” but if the properties are not “relevant properties,” transfer cannot occur (compare with Reference Andersen, Gass and SelinkerAndersen’s [1983] “transfer to somewhere”). Transferring a tennis swing to badminton seems more likely than transferring shot-put techniques to the sport of curling. Reference ZoblZobl (1980, p. 43) proposed conditions of transfer in SLA, including the “selectivity of L1 influence on L2 acquisition,” also supported by others (Reference James, James and KettemannJames, 1983; Reference Young-ScholtenYoung-Scholten, 1985). In similar phenomena, L1 transfer is more prevalent than U because of perceptual saliency. Minimal differences are less likely to be noticed, resulting in transfer, that is, nonlearning. Accordingly, transferring English /k/ to Arabic /q/ seems more likely than transferring any other English sound to Xhosa clicks, where learners are more likely to produce non-English sounds, indicating that U plays a greater role.

However, seemingly there is a paradox regarding similarity. When L1 and L2 phenomena are maximally similar, transfer should persist and consequently L2 should not be learned. Nevertheless, at some point, as L1 and L2 approach identity, negative transfer will become positive transfer because the two phenomena have become virtually indistinguishable. Thus, learning ostensibly has taken place, but in fact it is merely positive transfer. This “similarity paradox” was noted by the prominent psychologist Reference OsgoodOsgood (1949). However, the definition of “indistinguishable” can be a further issue because some phenomena can be instrumentally distinguishable but humanly perceptually indistinguishable: Reference Eckman, Iverson and SongEckman, Iverson, and Song (2015) found statistically significant voice onset time (VOT) differences between /p/ and /b/ in Arabic speakers of English, even though trained phoneticians perceived no differences.

The U pattern of the similarity corollary follows from mathematical logic: because it is well-known that similar phenomena are acquired slowly while L1 influence is strong and persists, U must then increase slowly and decrease slowly because L1 + L2 + U = 100 percent. Thus, U must be a lesser factor than L1.

There is a plethora of research that demonstrates that similar phenomena are difficult to learn and are acquired slowly (Reference Best and StrangeBest, 1995; Reference Bohn, Wrembel, Kiełkiewicz-Janowiak and GąsiorowskiBohn, 2020; Reference Major and ChapelleMajor, 2018; Reference Major and KimMajor & Kim, 1996; Reference Yang, Chen and XiaoYang et al., 2020).Footnote ⁹ The Speech Learning Model (SLM, Reference Flege and StrangeFlege, 1995; SLM-r, Reference Flege, Bohn and WaylandFlege & Bohn, 2021) has convincingly demonstrated that “similar” sounds are more difficult to learn than “new” sounds because speakers classify similar sounds as “equivalent.” A study designed to test the SLM and the OPM supported both models (Reference JevringJevring, 2015; also see Section 10.2.4). Reference JevringJevring (2015, p. 34) states that “these findings support the Similarity Corollary of the OPM as well, since L1 transfer is a larger affecting factor on similar phenomena than universals and the L2.”

Well-known facts about L1 acquisition, L2 acquisition, and dialect variation add further evidence for the OPM. In L2 acquisition, L1 phenomena can be termed similar or dissimilar to L2 phenomena, but in L1 acquisition there is no L2 as a basis for comparison. Nevertheless, L1 and L2 acquisition share important similarities. In L1 acquisition, mergers are common when a phenomenon being acquired is similar to something already acquired; for example, mergers of /s/ and /ʃ/ and /ɑ/ and /ɔ/ are legion. For the L1 learner, this means that the phenomenon in the adult L1 that has previously been acquired is being substituted for the phenomenon that has not yet been acquired (e.g., [s] for [ʃ] and [ɑ] for [ɔ]). Accordingly, because of similarity, this substitution persists and therefore U cannot surface (as in SLA). Mergers that occur in L1 acquisition are also reflected in dialects in contact (a form of L2 acquisition) and historical change. For example, /ɑ/ and /ɔ/ mergers prevalent in children (e.g., caught/cot) are standard in many American dialects, in addition to mergers between tense and lax vowels before nasals and liquids (e.g., Engler/angler, hairy/Harry, sale/sell). The patterns of L1 and L2 acquisition are therefore analogous.

10.2.4 Markedness Corollary of the OPM

The last corollary deals with markedness. The OPM claims that in marked phenomena, IL develops chronologically in the following manner: L2 increases slowly, L1 decreases rapidly and then decreases slowly, and U increases rapidly and then decreases slowly (Figure 10.4). Slowly and rapidly mean more slowly and more rapidly than in normal phenomena. Research is prevalent demonstrating that marked phenomena are acquired slowly (Reference Altenberg and VagoAltenberg & Vago, 1983; Reference Anderson, Ioup and WeinbergerAnderson, 1987; Reference ArchibaldArchibald, 2021; Reference Carlisle, Yavaş, Kehoe and CardosoCarlisle, 2017; Reference Carlisle, Espinosa and YavaşCarlisle & Espinosa, 2015; Reference EckmanEckman, 1977, Reference Eckman, Malovrh and Benati2018; Reference MajorMajor, 1994, Reference Major and Chapelle2018; Reference Major and FaudreeMajor & Faudree, 1996).

Figure 10.4 Markedness corollary of the OPM.

Compare Figure 10.1 to Figure 10.4. In marked phenomena, the role of U is much greater than L1 compared to normal phenomena. Further compare Figure 10.3 to Figure 10.4. Both graphs indicate a slower rate of L2 acquisition compared to normal phenomena (Figure 10.1). However, Figures 10.3 and 10.4 differ in their relative proportions of L1 and U. In similar phenomena, L1 persists and U remains minimal throughout; in marked phenomena, U has a relatively greater role than L1, compared to similar phenomena.

In order to test the similarity and the markedness corollaries, it is crucial to compare phenomena that are equal in all relevant criteria except the one under scrutiny. For example, the claims could not be tested by comparing a marked and similar phenomenon to a less marked and less similar phenomenon. This is because both similarity and markedness slow acquisition but have opposite effects on U and transfer (Figure 10.3 versus Figure 10.4). Although it is virtually impossible to control for everything, the OPM is formulated ceteris paribus.

The markedness corollary follows logically from general learning principles. At the early stages of acquisition, learners soon realize that for a specific phenomenon there is often nothing from the L1 to transfer to the L2 (e.g., Zulu clicks; compare with Section 10.3 on conditions for transfer). This prompts learners to attempt something other than the L1, namely U, which causes U to rise rapidly. Not making a focused effort is employing the L1; however, making a concerted effort but without success is employing U. Because marked phenomena are difficult, learners continue attempts but fail repeatedly, resulting in various U substitutions rather than L1 substitutions. For example, very few non-native Highland Chontal speakers can master the very marked /tɬʼ/, a laterally released voiceless affricate ejective. Predictably, English speakers will often produce non-English/non-Chontal sounds, namely U substitutions, and continue to do so. When L1 continues to decrease while L2 accuracy is still not achieved, this results in a rapidly rising U component. However, because L2 is mastered slowly, U must subsequently decrease slowly.

The L1 pattern follows from mathematical logic: Because L1 + L2 + U = 100 percent and because it has been argued that U increases rapidly and then decreases slowly, this means that L1’s chronology must be the reverse of U: a rapid decrease with a subsequent slow decrease.

The claim that the role of U is greater in more marked phenomena than in less marked phenomena is supported by several longitudinal studies. Two studies designed to test the OPM’s markedness corollary supported its claims: a study of onset clusters in Spanish learners of English (Reference Carlisle, Espinosa and YavaşCarlisle & Espinosa, 2015) and Reference JevringJevring’s (2015) study of English fricatives of Swedish learners of English (compare with Section 10.2.3 on similarity). Both studies found a greater occurrence of U in more marked phenomena, compared to less marked phenomena. Earlier studies also confirmed this. Reference MajorMajor (1986a) reported more U substitutions for /r/ than for /ɾ/, the less marked sound, in a longitudinal study of L2 Spanish acquisition by English speakers. In a longitudinal study of Brazilian English investigating initial and final consonants and consonant clusters in nine different environments, a variable rule analysis (VARBRUL) found that the more marked the environment, the greater the probability of U (Reference Major, Bayley and PrestonMajor, 1996). For example, in codas, the probability of U was 0.854 for double stops, 0.590 for fricative plus stop, but only 0.083 for a single stop.

Further, L1 acquisition of marked phenomena parallels L2 acquisition. At an early stage, the L1 learner most likely substitutes the nearest equivalent previously acquired (e.g., [w] for [ɹ]). However, the L1 learner soon realizes that this early-stage substitution is not adequate and so attempts other substitutions; nonetheless, these substitutions are not native-like either. This necessarily means that they are U because the learner has only the portion of the L1 previously acquired and U; no other systems are available (e.g., an L1 Spanish learner’s non-native substitutions for the Spanish trill /r/). This chronological development is analogous to the L2 learner: after realizing that L1 substitutions are not satisfactory, the L2 learner employs U. Both the L1 and the L2 learners’ continual but inaccurate trials result in a rapid rise of U and its subsequent persistence.

10.2.5 Comparison of Normal, Similar, and Marked Phenomena

The unifying pattern for similarity and markedness is that the rate of L2 mastery is slower than for normal phenomena. However, they differ in their relative importance of L1 and U. After the initial stages, when L1 is large and U small for all phenomena, their patterns diverge. In similar phenomena, L1 persists and the proportion of U to L1 is relatively small throughout the various stages. In contrast, in marked phenomena, the patterns are reversed: U quickly becomes large and persists, resulting in a relatively large proportion of U to L1 throughout acquisition. Thus, the later stages of acquisition of similar and marked phenomena for L1 and U are near mirror images of each other (Figure 10.3 versus Figure 10.4). Two examples mentioned previously (Sections 10.3 and 10.4) illustrate the difference: an English speaker persistently substitutes /k/ for /q/ but produces non-English substitutions for /tɬʼ/. Table 10.1 compares Figures 10.1, 10.3, and 10.4.

In order to further investigate the OPM’s claims for similarity and markedness, it would be informative to investigate their interaction with a phenomenon that is both marked and similar to a phenomenon in the L1. Because these factors should be mutually reinforcing, an obvious prediction would be that acquisition would be even slower than for purely marked or purely similar phenomena. It would also be predicted that L1 transfer would be more important than U because the effects of similarity and markedness on U would tend to neutralize each other (because in marked phenomena U increases rapidly but in similar phenomena U increases slowly). Furthermore, in similar phenomena L1 tends to persist. A test case could be L1 English [h] and L2 Arabic [ħ]. The two sounds are very similar, but [ħ] is very marked. This prediction was supported by a longitudinal study of Arabic. Reference AlshalawiAlshalawi (1998) reported that NSs of American English substituted [h] for [ħ] more frequently than U substitutions at all stages, meaning that L1 transfer was more important.

It would also be relevant to compare the patterns of marked L2 phenomena and less-marked phenomena to similar L2 phenomena and less-similar phenomena. One possible test case could be L1 Spanish /a/ and L2 English /ɔ/, /æ/, and /ʌ/ (Spanish does not have any of these vowels). English /æ/ and /ʌ/ are more marked than /ɔ/, but all three are similar to Spanish /a/. Another possible study could be L1 Spanish /t̪/ and L2 English /θ/ and /t/; English /θ/ is more marked than English /t/, but both English /θ/ and /t/ are similar to Spanish /t̪/.

10.2.6 Summary of the OPM

While L1 transfer, similarity, markedness, and U have long been investigated, remarkably, prior to the OM (Reference Major, Ioup and WeinbergerMajor, 1987b), research had not addressed the interrelationships of these four factors. Expanding on the OM, the OPM proposes even more explicit interrelationships. The model argues that, chronologically and as style becomes increasingly formal, L2 native-like processes increase, L1 transfer processes decrease, and universal processes increase and then decrease. It further claims that similarity and markedness govern these patterns in predictable ways. In similar phenomena L1 transfer persists, whereas in marked phenomena U persists. Furthermore, it has been argued that L1 and L2 acquisition share important similarities.

10.3.1 Loan Phonology

Loan phonology usually involves a very rudimentary form of L2 acquisition, where L1 transfer completely dominates. For example, Koran in English is pronounced with initial [kʰ], not [q]. However, sometimes a new sound is adopted. Maori of the Cook Islands has no /s/, yet Jesus is pronounced [iesu]. Phonological changes can be limited to particular words. Some speakers with a classical musical and literary background carefully pronounce Bach [bax] and genre [ʒɑ̃ɹə]. The [x] may only occur in this one word Bach. In contrast, Reich is pronounced with final [k] and the deletion of [n] before r in genre does not occur in other words. Although [bax] Bach and [ʒɑ̃ɹə] genre may sound cultured and educated, [tʰæxɑməɾɹ̩] tachometer and [hɛ̃ɹi] Henry sound odd or amusing. The special status of loanwords is also evident in dialects of the same language. In the United States, many young white speakers of standard American English pronounce cool [kʰuw], yet these same speakers pronounce pool [pʰuɫ] (the slang use of “cool” as reportedly originated from African American jazz musicians; in African American English, coda /l/ → [w]). Though sound change is usually regular (according to Neogrammarians, e.g., IE p > Germanic f: pater > father), these examples indicate that change can be limited to certain words (compare with British /ɑ/ vs. /æ/ in can’t, can, pant).

Loan phonology is relevant to the OPM because of the occurrence of U. Some English speakers pronounce Bach and Van Gogh with final [h], rather than [x] in native German and Dutch, respectively. In attempting final [x], speakers produce final [h] (non-occurrent in English), thus resulting in a change in English phonotactics (note: in native Dutch, Gogh is pronounced [xɔx], not *[gɔx], as some English speakers do). Onset cluster deletion is a further example of U (e.g., Pfizer, psoas, and tsunami). Although coda cluster deletion in English occurs synchronically (e.g., clothes → clo[z]), onset deletion does not and is therefore a U process limited to loanwords.

We find U also in the pronunciation of the names of two well-known former professional athletes, Brett Favre and Patrick Roy. In US and Canadian English, Favre is pronounced [fɑɹv] and Roy [wɑ]. Metathesis, which occurs in Favre (also in hors d’oeuvre), is a diachronic process in English but not normally an adult synchronic process, except in speech errors; thus, it should be considered U (although very common in L1 acquisition, e.g., a[nim]al animal). The deletion of r in Roy ([ʁwa] in native French) is likewise a U process because L1 transfer would predict epenthesis, not deletion (e.g., Rwanda [ɹəwɑdə], not *[wandə], but Roy [wɑ], not *[ɹəwa]).

The Bach example is also relevant to both the chronological and the stylistic corollaries of the OPM. Some English speakers substitute [k], others use [h], while still others achieve native [x], representing stages of the chronological corollary: L1 transfer → U → native L2. In a classical radio broadcast, an announcer may carefully pronounce Bach with final [x], though [h] when less careful, yet [k] outside of work, perhaps not wanting to sound pretentious. This corroborates the stylistic corollary.

The similarity and the markedness corollaries pertain to loan phonology as well. In similar phenomena L1 transfer is relatively more important than U and in marked phenomena the situation is the reverse. Thus, English speakers typically pronounce coup d’état with aspirated (long-lag VOTs) [kʰ] and [tʰ] instead of the similar unaspirated French (short-lag VOTs) [k] and [t̪], whereas the same speakers carefully trying to pronounce the word Xhosa likely produce non-English/non-Xhosa sounds, rather than a [kʰ] or [h].

10.3.2 Isolation, First Language Attrition, and Assimilation

First language attrition,Footnote ¹⁰ assimilation, bilingualism, and isolation can result when different languages come into contact (compare with Reference WeinreichWeinrich’s [1953] classic work). Complete isolation is rare because monolinguals often have contact with bilinguals; therefore, both groups are mutually influenced.

Another possible outcome is L1 attrition, the extreme form being a complete L1 loss with subsequent assimilation to the dominant language. Attrition can be considered acquisition in reverse. Thus, the OPM proposes that the beginning stages of L1 attrition are the later stages of L2 acquisition (Figures 10.1, 10.3, and 10.4). Consider children acquiring the NL of their immigrant parents who are undergoing attrition. Most L1 acquisition input is casual speech (e.g., conversation versus carefully reading a word list). Because research has shown that L1 attrition in adults first occurs in casual speech, this then means that the first generation of speakers will show significant attrition, perhaps more than their parents, but also due to the dominant language in that country.

Further, L1 attrition has striking similarities to some of Jakobson’s work from more than eighty years ago (Reference JakobsonJakobson, [1941] 1968).Footnote ¹¹ He postulated interrelationships between child language, aphasia, and phonological universals, convincingly arguing that L1 attrition in aphasic adults mirrors L1 acquisition; specifically, phenomena that are acquired last are lost first. His claim – last acquired, first lost – parallels L1 attrition as well. Thus, L1 attrition is L2 acquisition in reverse. In accordance with the OPM, L2 will transfer to L1. Moreover, U is also a factor; for example, in L1 Spanish attrition in the US, speakers often merge English /ɛ/ and /æ/ (similar to many monolingual speakers in western USA where Ellen and Allan are homophones and penguin and sanguine rhyme). Since Spanish has neither vowel and these speakers do not substitute Spanish /e/ or /a/, this merger must be attributed to U. Thus, Figure 10.1 also depicts L1 attrition in reverse.

The vast majority of US monolingual NSs of English have immigrant ancestors who did not speak English natively. Consequently, American English is a product of L2 acquisition, leaving traces in speakers who have long since lost their ancestral languages. In Minnesota and North Dakota, there is speculation that the very back and monophthongal /o/ can be attributed to German and Scandinavian immigrants. A number of mergers in American English possibly could be traced to incomplete L2 acquisition, a form of attrition from the original British English being acquired. Among them are the merger of British /ɑ/, /ɒ/, and /ɔ/ to American /ɑ/ and /ɔ/ or to the single vowel /ɑ/ (especially in the West), and the loss in most American dialects of the distinction between /w/ and /ʍ/.

10.3.3 Bilingualism and Multilingualism

In bilingual acquisition in childhood, there are four components: U, language A, language B, and shared language A and B. For example, at some stage the child may produce only three vowels in both languages,Footnote ¹² though A and B have five and eleven vowels (e.g., Spanish and English). At a later stage, the learner exhibits evidence of two vowel systems (though still not native adult), while still retaining a core shared system characterizing both languages. At a later stage the proportions of A and B increase, resulting in a proportional decrease of shared A and B. At the final stage the two systems become separate. Two completely separate systems characterize the idealized learner. It is unlikely that such a person exists because both languages continually influence each other. Moreover, bilingual speakers are never equally competent in both languages in all environments. In order for this to happen, absolutely simultaneous bilingual acquisition would have to occur, a physically impossible situation, except hypothetically in cloned individuals. The result is two new language varieties (A_new and B_new) having components of the original A and B (i.e., mutual transfer), remnants of U, as well as a mutually shared core system. The same patterns also apply to multilingualism.Footnote ¹³ In the US Southwest, some Spanish-English bilinguals contrast /b/ and /v/ in Spanish, which are non-occurrent in standard Latin American Spanish.

Furthermore, these new varieties (A_new and B_new) can affect monolingual speakers of A and B. In the US Southwest, where there is widespread Spanish-English bilingualism, monolingual English speakers seem to have more of a syllable-timed rhythm (typical of Spanish) than monolingual English speakers from other parts of the US, who have a stress-timed rhythm. Such mutual influence can also result in phenomena that are different from monolinguals of both languages. In the classic study of bilinguals, Reference Caramazza, Yeni-Komshian, Zurif and CarboneCaramazza et al. (1973) found VOTs of word-initial stops (based on perception and production) in the English and the French of French-English bilingual Canadians to be intermediate between monolingual speakers of both languages. These values may be attributed to U because they are different from monolinguals (but see Reference MoyerMoyer, 2021).

10.3.4 Dialects in Contact

The differences between language and dialect are not always based on linguistic criteria but often on political, ethnic, and social criteria. For these reasons, SLA necessarily includes dialects in contact because a new dialect is indeed an L2. Thus, the OPM principles operate: at the initial stage L1 transfer predominates (native dialect). As acquisition proceeds, the L2s (new dialects) increase and the L1s (native dialects) decrease, while U increases and then decreases. Finally, typically remnants of the L1s and U remain, similar to an individual acquiring an L2.

It is widely known that the more intense and varied the contact, the more frequent are mergers (Reference TrudgillTrudgill, 1986). Some distinctions in the Eastern US are absent in the Midwest and even more absent in the West, where frequent migration continues to occur. In Los Angeles, San Francisco, and Phoenix, the majority of the residents are not native-born. The distinction between /æ/ and /ɛ/ before /ɹ/ in the Northeast (e.g., Harry/hairy) is absent in the Midwest and the West. In many areas of the US, writer and rider are pronounced r[ʌj]der and r[ɑj]der, respectively, but they merge in the West, resulting in a diphthong that is intermediate between the two. The /ɑ/–/ɔ/ distinction has been lost in the entire West. Many western speakers merge certain vowels before /l/: wheel/will, sale/sell, Ellen/Allan, colt/cult, pole/pull. In standard British English, the following pairs contrast phonologically: where/wear, tune/toon, dew/do, as well as the vowels in balm, bomb, and bought (/ɑ/, /ɒ/, /ɔ/). These distinctions rarely occur in the USA.

The similarity and the markedness corollaries of the OPM are also relevant to dialects in contact. The OPM claims that in similar phenomena, L1 plays a larger role than U, but in marked phenomena the opposite is true. Accordingly, because of the similarity of the mother dialects, L1 plays a greater role than U (e.g., Midwestern versus Western US English compared to Mandarin versus German).

Reference LabovLabov (1994, p. 363) coined “the Bill Peters effect,” which is germane to the OPM’s stylistic corollary. Bill Peters, an eighty-year-old man, exhibited a distinction between /ɔ/ and /ɑ/ in spontaneous speech (his native dialect) but in minimal pairs he merged them, typical of younger speakers. The dialect with this merger can be considered Peters’ L2 because he learned it only in later life. Labov further documented a number of similar cases. Peters’ greater accuracy with the new dialect (the L2) in formal speech than in spontaneous speech is evidence supporting the OPM.

10.3.5 Pidgins and Creoles

When individuals sharing no common language come into contact, a pidgin can form; if it becomes nativized it is termed a creole. Reference SchumannSchumann (1978) likened L2 acquisition to pidginization, giving evidence from the oft-cited Alberto’s acquisition of English.

The OPM claims that the patterns of L1, L2, and U hinge on whether the phenomena are normal, similar, or marked. When the L1s share similar phenomena, transfer is a larger component than when they have little in common, where U plays a greater role. There are important differences between pidgins and creoles and dialects in contact. In dialects in contact, L1 transfer is generally more important than U because of the similarity of the mother languages (Section 10.3.4); however, in pidgins and creoles, U is more important because of the frequent dissimilarity of the mother languages. In US Western dialects, which are the results of dialects in contact, a number of marked phenomena are present: /θ/ and /ð/, coda consonant clusters, and coda voiced obstruents, which are also present in the vast majority of the mother dialects. In contrast, many English creoles do not have any of these marked phenomena; moreover, marked phenomena are rare in all pidgins and creoles (e.g., Tok Pisin has no interdental fricatives and Haitian Creole has no front rounded vowels; Reference ValdmanValdman, 2015; Reference VerhaarVerhaar, 1995).

One reason why U is a prominent factor in pidginization is because speakers have little or no shared knowledge of each other’s languages; therefore, they resort to universal principles in order to communicate. Reference BickertonBickerton’s (1981) well-known bioprogram hypothesis argues for the importance of universals in pidginization. Successful communication far outweighs any attention to form, which is a moot point because when a pidgin is created, no established standard exists. Thus, a speaker employing U is more probable than employing an L1 with someone who has no familiarity with that L1. As a pidgin forms, speakers with different L1s produce different varieties. Because communication is paramount, these different varieties become increasingly influenced by general acquisition principles (U) and less and less influenced by their L1s. If and when a pidgin becomes more standardized (and/or becomes a creole), the U components can then be considered to be part of the language and no longer U (e.g., Haitian Creole has affricates but French does not; terminal coda obstruent devoicing is standard German but also is a common U process in L1 and L2 acquisition; Section 10.2.1). When children learn a pidgin as their L1 (thus forming a creole), they will hear different varieties and necessarily create a compromise, an intermediate category. This results in mergers (as with dialects in contact), which are widespread in all pidgins and creoles (e.g., Tok Pisin has five vowels but English has nine to eleven, depending on the dialect).

The preponderance of U is evident in Haitian Creole. In an extensive work, Reference ValdmanValdman (2015) describes a host of phenomena in Haitian Creole that do not occur in French. Among them are /ɣ/, /ɣ/ → [w] (< French /ʁ/), /j̃/ (< French /ɲ/, zonyon [zõj̃õ] “onion”), /ʧ/ and /ʤ/,Footnote ¹⁴ intervocalic voicing and deletion of voiceless stops, and hypercorrection. These are natural phonological processes that occur in a number of languages, both diachronically and synchronically (e.g., English /ʧ/, /ʤ/ < /kʲ/, /gʲ/; Portuguese casual speech minha [mĩɲa] → [mĩj̃a] “my” [fem.]).

Usually one language is dominant in pidgins and creoles. If a pidgin survives and becomes a creole, it may undergo decreolization or hypercreolization. If the creole experiences continual contact with the dominant language, decreolization is favored (e.g., Jamaican English) and thus the dominant language becomes increasingly more important than U. However, if contact is cut off, then hypercreolization is favored (e.g., Tok Pisin and Haitian Creole). The result is that U becomes increasingly more important because the creole continues to develop independently and is therefore influenced more by general linguistic U than by the original dominant language. However, other factors can be more important, such as group identity and solidarity. For example, though most speakers of Hawaiian Creole English have continual contact with standard English, the creole does not seem to be undergoing decreolization; rather, it seems to be becoming even more distinct from standard English, very likely because of its importance to Hawaiian identity.

Decreolization is analogous to SLA when L2 learners have continual contact with NSs and learners become increasingly more native-like (e.g., immigrants). In contrast, hypercreolization is similar to learners being cut off from contact with NSs while continuing to use the L2 with other NNSs (non-native speakers) with different L1s (e.g., a group of professionals with different L1s working together in a country where English is not spoken natively). This L2 continues to develop independently (as in decreolization), with less and less influence from other speakers’ L1s and the original NS version of the L2. Thus, this process is similar to hypercreolization and is continually taking place in international Englishes (see Reference Trudgill and HannahTrudgill & Hannah, 1994).

10.3.6 Summary of the OPM and Languages in Contact

In sum, SLA encompasses both an individual and languages in contact. The OPM claims that its principles apply not only to the individual but also to languages in contact, including loanwords, isolation and assimilation, bilingualism and multilingualism, dialects in contact, and pidgins and creoles. For example, in similar phenomena and in dialects in contact, L1 transfer is more prominent, but in marked phenomena and in pidgins and creoles, universals are more prominent.

11.3.1 Perception

Cross-linguistic connections in perceptual representations have a long-studied history,Footnote ¹ approached heavily from the perspective of second language acquisition. Two popular frameworks within the language learning space are the Perceptual Assimilation Model-L2 (PAM-L2; Reference Best, Tyler, Bohn and MunroBest & Tyler, 2007) and the Speech Learning Model (SLM; Reference Flege and StrangeFlege, 1995). These frameworks postulate connections between perceptual categories across languages such that category boundaries are nudged on account of cross-linguistic mutual influence. The revised SLM framework, the SLM-r (Reference Flege, Bohn and WaylandFlege & Bohn, 2021), is an updated and more dynamic model that allows for continual reorganization of cross-linguistic connections across the life span (Reference Flege, Bohn and WaylandFlege & Bohn, 2021). We recommend Reference Chang, Katz and AssmannChang (2019) as an overview on the phonetics of second language learning.

While early work indicated that bilinguals may exhibit perceptual behaviors that are between those of monolingual groups from the respective languages (Reference Caramazza, Yeni-Komshian, Zurif and CarboneCaramazza et al., 1973), subsequent work focusing on early bilinguals and highly proficient late bilinguals suggests that these groups perceptually shift their categorization boundaries based on the language they are perceiving. For instance, the threshold for stop voicing in different languages (e.g., /b/ vs. /p/) is shifted along a voice onset time (VOT) continuum for Spanish-English bilinguals, suggesting that listeners can rely on different, and thus separable, categories in perception and recognition.Footnote ² This behavior is called a double phonemic boundary (Reference Casillas and SimonetCasillas & Simonet, 2018; Reference Garcia-Sierra, Diehl and ChamplinGarcia-Sierra, Diehl, & Champlin, 2009; Reference García-Sierra, Ramírez-Esparza, Silva-Pereyra, Siard and ChamplinGarcía-Sierra et al., 2012; Reference Gonzales and LottoGonzales & Lotto, 2013). Bilinguals may also switch in and out of language modes based on previously heard speech, social and pragmatic environments, or linguistic cues. Those linguistic cues may be phonotactic or subphonemic in nature. Not all perception tasks point to the same linguistic knowledge. For example, Reference Antoniou, Tyler and BestAntoniou, Tyler, and Best (2012) examined early bilinguals’ performance on categorization, goodness ratings, and discrimination tasks. Language-specific patterns were found for categorization and goodness rating responses, but not discrimination, which appeared to be dictated by the more dominant language of the listener. While more work is necessary to fully understand the effects of perceptual modes on listening for bilinguals, it is clear that bilinguals, particularly early bilinguals, exhibit sensitivity to phonetic details.

11.3.2 Production

Across a range of methods, there is clear and robust evidence that bilingual speech production can show cross-linguistic influence (e.g., Reference Baker and TrofimovichBaker & Trofimovich, 2005; Reference BarlowBarlow, 2014; Reference FlegeFlege, 1987; Reference Fricke, Kroll and DussiasFricke, Kroll, & Dussias, 2016; Reference MajorMajor, 1992; Reference OlsonOlson, 2016; Reference Sancier and FowlerSancier & Fowler, 1997; Reference SimonetSimonet, 2011). This cross-linguistic influence suggests the existence of linked categories. However, early bilinguals can maintain distinct phonetic categories across their two languages (e.g., Reference GuionGuion, 2003; Reference Sundara, Polka and BaumSundara, Polka, & Baum, 2006; Reference MacLeod, Stoel-Gammon and WassinkMacLeod, Stoel-Gammon, & Wassink, 2009) and the evidence that bilinguals habitually produce compromised categories – those that average across the categories of their multiple codes – is not particularly strong (Reference CasillasCasillas, 2021). Multilingual individuals are socially sophisticated, capable of performing different degrees of cross-linguistic influence in pronunciation variation based on social demands (Reference Bullock, Toribio, Isurin, Winford and deBotBullock & Toribio, 2009). As such, a model of bilingual speech production must allow for the interaction of phonological and phonetic categories without the requirement that one language’s categories subsume the other’s. The degree of cross-linguistic influence also needs to be flexible synchronically and across an individual’s life span (Reference Simonet and AmengualSimonet & Amengual, 2020). The dynamic nature of the strength of cross-linguistic influence in production is apparent in studies that manipulate language mode (Reference GrosjeanGrosjean, 1998, Reference Grosjean2008). In experimental contexts where researchers attempt to present a unilingual mode, bilinguals can show, for example, phonetic patterns that are equivalent to monolingual norms (e.g., Reference Antoniou, Best, Tyler and KroosAntoniou et al., 2010; Reference Henriksen, Coetzee, García-Amaya and FischerHenriksen et al., 2021), whereas conditions that prompt a multilingual mode through, for example, code-switching elicit cross-linguistic influence in phonetic patterns (e.g., Reference Henriksen, Coetzee, García-Amaya and FischerHenriksen et al., 2021). The empirical evidence is mixed for whether language mode effects vary as a function of language dominance (e.g., Reference AmengualAmengual, 2018; Reference SimonetSimonet, 2014).

11.3.3 Social Meaning

The subphonemic variation in speech that is associated with social identity is rich, and well-studied within monolingual populations. There appears, however, to be limited research in this domain in bilingual populations. One notable study was Reference Szakay, Babel and KingSzakay, Babel, and King (2016), who queried the social associations across languages and dialects in the New Zealand context with a sample of English-Māori bilinguals. New Zealand English has two main varieties: Māori English and Pakeha English. These Englishes are associated with ethnically Māori and ethnically (white) European individuals, respectively. Māori English is acquired as a native variety of English and is positively associated with Māori culture and identity in New Zealand (Reference HolmesHolmes, 2005; Reference Maclagan, King and GillonMaclagan, King, & Gillon, 2008). Reference Szakay, Babel and KingSzakay et al. (2016) assessed whether the Māori social association with Māori English led to stronger L2 (Māori)-L1 (English) priming when Māori English was the target, as opposed to Pakeha English. We walk through this hypothesized social connection with the lemma snow. (See figure 1 in Reference Szakay, Babel and KingSzakay et al. [2016] for a schematization to accompany our description.) This lemma concept of “white flakes of frozen water vapor” exists in both languages and has distinct lexemes in each language, which can be orthographically represented as huka and snow. Stronger connections at the lemma level exist in the mapping of the L1 lexeme to the L2 lexeme through the lemma, often seen as the stronger effect of L1–L2 priming (e.g., Reference Jiang and ForsterJiang & Forster, 2001) compared to the weaker links in the L2–L1 direction. Of note in our example with snow, the goose vowel in New Zealand English exhibits pronunciation variation indexed to Māori and Pakeha identities, with realizations of snow as [snɵʉ̘] and [snɐʉ], respectively, in Māori English and Pakeha English.Footnote ³ The social link between the Māori English pronunciation of snow [snɵʉ̘] and the Māori huka [hʉ̙kɐ̝] was hypothesized by Reference Szakay, Babel and KingSzakay et al. (2016) to facilitate L2–L1 priming for Māori English targets. Indeed, they found that Māori effectively primed Māori English, but not Pakeha English, suggesting that a social category connection facilitated L2–L1 processing. Listeners were sensitive to the subphonemic variation in the Māori English that indexed Māori identity, which is also intricately connected to the Māori language itself.

This particular example of cross-linguistic dialect-specific L2–L1 priming is crucial because the lack of shared phonemes in Māori huka and English snow strongly suggests that the cross-language activation must come from subphonemic cues. Indeed, Māori English and Māori are known to share suprasegmental similarities (Reference SzakaySzakay, 2012). Listeners must be sensitive to these subphonemic suprasegmental differences in order to utilize them in the creation, maintenance, and use of a social Māori category across codes. This is a classic example of the fine-grained phonetic detail in an exemplar network being used to facilitate linguistic processing.

11.4.1 Phonetic Details and Abstract Categories

Exemplar models have historically been contrasted with abstract models, where mental representations of language are exclusively available in an abstracted form, leaving fine phonetic detail unrepresented (e.g., Cohort: Reference Marslen-Wilson, Bouma and BouwhuisMarslen-Wilson, 1984, Reference Marslen-Wilson1987; TRACE: Reference McClelland and ElmanMcClelland & Elman, 1986; Shortlist: Reference NorrisNorris, 1994). In these types of models, while token-specific information may be perceived in the speech signal, a normalization process results in a more abstract sublexical representation, with models differing in the nature of the abstract sublexical units (Reference Weber and ScharenborgWeber & Scharenborg, 2012). The dichotomy between abstract and exemplar interpretations was more explicit in early phonetic research. For instance, in a study by Reference Pallier, Colomé and Sebastián-GallésPallier, Colomé, and Sebastián-Gallés (2001), Catalan-Spanish bilinguals, varying in Spanish or Catalan dominance, were tested in a “medium-distance” primed auditory lexical decision task. Minimal pairs bearing Catalan-specific vowel contrasts were separated across eight to twenty trials (e.g., /netə/ “granddaughter” – /nɛtə/ “clean, fem.”) and matched with corresponding identity pairs (e.g., /netə/ – /netə/; the same word is repeated twice, but as acoustically unique instances). Reference Pallier, Colomé and Sebastián-GallésPallier et al. (2001) observed that Catalan-dominant speakers showed priming for identity pairs and not minimal pairs, while Spanish-dominant speakers showed priming for both identity pairs and minimal pairs. These findings were taken as evidence in support of abstract models and not exemplar models, since Spanish-dominant speakers could only have demonstrated minimal pair priming if the Catalan contrasts were being assimilated to native language categories, with minimal pairs in Catalan being perceived as identity pairs.

Recent work has provided growing empirical evidence in support of hybrid models that assume that linguistic knowledge is composed of both abstract and exemplar representations: listeners are sensitive to phonetic detail and make generalizations over and across both experienced items and abstract categories (Reference PierrehumbertPierrehumbert, 2016). Lexically guided perceptual learning paradigms offer a nice example of the need for both attention and retention of phonetic detail and generalization over categories. In these paradigms, listeners are exposed to novel pronunciations in the context of a lexical frame. This lexical frame provides the scaffolding necessary for a listener to connect the novel pronunciation to a particular phoneme or allophone. This exposure causes an adjustment of a category, such that listeners’ categories accommodate the new pronunciation variation. The empirical support for perceptual adaptation is robust (e.g., Reference Eisner and McQueenEisner & McQueen, 2006; Reference JesseJesse, 2021; Reference Kraljic and SamuelKraljic & Samuel, 2007; Reference McQueen, Cutler and NorrisMcQueen, Cutler, & Norris, 2006; Reference Norris, McQueen and CutlerNorris, McQueen, & Cutler, 2003; Reference SamuelSamuel, 2016), and the literature includes work on bilinguals (Reference Bruggeman and CutlerBruggeman & Cutler, 2020; Reference Chan, Johnson and BabelChan, Johnson, & Babel, 2020; Reference Cutler, Burchfield, Antoniou, Calhoun, Escudero, Tabain and WarrenCutler, Burchfield, & Antoniou, 2019; Reference Drozdova, Van Hout and ScharenborgDrozdova, Van Hout, & Scharenborg, 2016; Reference Reinisch, Weber and MittererReinisch, Weber, & Mitterer, 2013). Within bilingual populations, however, the results are somewhat mixed. While there is evidence that bilinguals show adaptation in their second language (Reference Drozdova, Van Hout and ScharenborgDrozdova et al., 2016; Reference Reinisch, Weber and MittererReinisch et al., 2013), some data suggest that a bilingual’s less-dominant language may not be so adaptable (Reference Bruggeman and CutlerBruggeman & Cutler, 2020; Reference Cutler, Burchfield, Antoniou, Calhoun, Escudero, Tabain and WarrenCutler et al., 2019; but see also Reference Chan, Johnson and BabelChan et al., 2020). What is important for the current purposes here is that a prerequisite for adaptation is sensitivity to phonetic details that connects with phonological categories. It may be that, in some bilingual scenarios, the appropriate connection strength may be lacking. This sets the stage for the crucial next step of this chapter: the introduction of a particular exemplar model.

11.4.2 Weaving Bilingualism into an Exemplar Model

It is apparent that a model of speech perception needs to provide the structure and the mechanisms to account for sensitivity to phonetic detail and the ability to abstract and generalize. Exemplar-based models typically abstract generalizations of experienced items. One must remember, however, that this does not mean that all experienced items are treated equivalently. Variation is socially weighted such that not all experienced items may contribute equivalently to the category (Reference Sumner, Kim, King and McGowanSumner et al., 2014). In many exemplar models, a production–perception feedback loop is central to the model’s architecture (e.g., Reference Ettlinger, Trouvain and BarryEttlinger, 2007; Reference Harrington, Kleber, Reubold, Schiel and StevensHarrington et al., 2018; Reference Pierrehumbert, Gussenhoven and WarnerPierrehumbert, 2002; Reference SóskuthySóskuthy, 2015; Reference Todd, Pierrehumbert and HayTodd et al., 2019; Reference TupperTupper, 2015; Reference Wedel and FatkullinWedel & Fatkullin, 2017). Note that none of these frameworks are designed to model bilingual knowledge, but at the heart of many is the tagging of phonetic variation with respect to its social and pragmatic contexts. While it is not a stretch to adapt such a concept to the tagging of language, what becomes complex and messy is the nature of the connections across languages within a bilingual.

Let us walk through the basic structure of Reference Todd, Pierrehumbert and HayTodd et al. (2019), a recent instantiation of an exemplar-based model, and note opportunities for bilingualism to factor into the model. We focus on the model by Reference Todd, Pierrehumbert and HayTodd et al. (2019) because it is computationally implemented (on a small scale) and thus well-specified, and it builds upon the seminal work by Reference Pierrehumbert, Bybee and HopperPierrehumbert (2001), which is, arguably, the most well-known exemplar model in linguistics. This multilayered framework builds on crucial insights from previous exemplar models (e.g., Reference Harrington, Kleber, Reubold, Schiel and StevensHarrington et al., 2018; Reference Johnson, Johnson and MullennixJohnson, 1997; Reference NosofskyNosofsky, 1986; Reference SóskuthySóskuthy, 2015; Reference WedelWedel, 2006, Reference Wedel2012; Reference Wedel and FatkullinWedel & Fatkullin, 2017), but is unique in its integration of these stages into a single architecture. Reference Todd, Pierrehumbert and HayTodd et al. (2019) advance an exemplar-based computational model where the role of the listener is emphasized in navigating and organizing phonetic variation. While this description sounds as though it affords the listener some amount of agency in the process, the point is that the processes of perceiving, categorizing, encoding, and recognizing are iterative, and transform the original signal into a mental representation that has been evaluated several times over. This provides multiple opportunities for bilinguals’ multifaceted linguistic knowledge to be integrated into an exemplar framework.

Let us begin with establishing that, in this model, categories and types are abstract generalizations of experienced phonemes and words, respectively. In terms of the architecture, we start on the listener side of the production–perception loop. In Reference Todd, Pierrehumbert and HayTodd et al. (2019), activation is the first stage in perception, where an item activates other similar exemplars within a distributional space. Note that the nature of the phonetic detail involved in this activation and similarity-matching process is unspecified.Footnote ⁴ The apprehension of an utterance results in activation, a process that excites other exemplars based on their distance in perceptual-acoustic space. What aspects of the signal are grounds for a similarity assessment in activation is an opportunity for bilingual interaction. A bilingual will have not only different but more categories and types available for activation than a monolingual. The similarity assessment can also change in the context of language learning. For example, as native speakers of Korean learn English, they adjust their spectral and duration cue weights for the /i/–/ɪ/ contrast in English (Reference Kim, Clayards and GoadKim, Clayards, & Goad, 2018). In learning to weigh the spectral cue more heavily, individuals may use the spectral information more prominently in the similarity assessment in the activation phase.

After activation comes identification, which is the process of recognizing the category, as informed by the phonological frame or word in which it is contextualized. The process of identification does not equal encoding, however, as encoding is achieved only if the item passes the discriminability and typicality evaluations. Discrimination is a comparison of the ratio of category activations to the discriminability threshold. The equation implemented by Reference Todd, Pierrehumbert and HayTodd et al. (2019) has an equivalent approach in Reference NosofskyNosofsky (1986). This part of the process is yet another opportunity where the bilingual experience is relevant. If a bilingual’s categorization threshold is lower, this will lead to an increase in discriminability, regardless of the activation threshold achieved. We assume that lexical biases are built upon robust abstractions at the phonological level. Second language learners exhibit less lexical bias (Reference Samuel and FrostSamuel & Frost, 2015; Reference Soo, Sidiqi, Shah and MonahanSoo et al., 2020), which may be a reflection of decreased thresholds for some types of bilinguals. Typicality is a comparison of the activation of the intended category to the typicality threshold. When activation is higher than the threshold, this increases the likelihood of that particular token passing the typicality evaluation. In Reference Todd, Pierrehumbert and HayTodd et al.’s (2019) model, both discriminability and typicality are probabilistic in nature. If a token is not well discriminated or is deemed atypical, it does not pass the thresholds in the discrimination and typicality evaluation stages. Correspondingly, it will not be stored and will, therefore, not be allowed to update the system. Here, an important question to consider in the context of bilingual listeners is what speech patterns are part of the typicality assessment. In many parts of the world, diaspora communities foster heterogeneous bilingual populations with exposure to a range of pronunciation variants. In Canada, for instance, the cultural mosaic means that Cantonese-English bilinguals are exposed to Cantonese, English, Cantonese-accented English, and English-accented Cantonese since birth. Which varieties benefit from their own category aggregates? Which compete or coalesce into a more or less cohesive system? Crucial questions like these need to be addressed in order to properly extend an exemplar model to bilingual listeners.

For bilinguals who speak related languages, these steps of identification, discrimination, and typicality evaluation are ideal opportunities for language interaction, as cognates can lead to cross-linguistic activation that blurs phonemic contrasts. Let us consider, for example, Reference AmengualAmengual (2016a), who examined Majorcan Catalan-Spanish bilinguals’ production and lexical processing of the /ɔ/–/o/ contrast in Catalan. Crucially, this contrast does not exist in Spanish. Participants – whether Catalan or Spanish dominant – produced less of a difference between Catalan /ɔ/ and /o/ in words that were cognates compared to noncognate items, suggesting phonetic interference in production. In terms of lexical processing, the results of a lexical decision task suggest coactivation of Spanish and Catalan lexicons. For congruent cognates, this had the effect of increasing listeners’ abilities to correctly identify words and nonwords, but the coactivation increased error rates for incongruent cognates.Footnote ⁵ Reference AmengualAmengual (2016a) identifies exemplar models as a framework that accounts for the necessary link between lexical and subphonemic structures within and across languages. To accommodate these findings, the similarity of the phonological frames in the cognate items must form part of the assessment of like sounds in the target window for typicality evaluation.Footnote ⁶ For cognate items, this coactivation would include both Catalan /ɔ/ and Spanish /o/, ultimately inviting the conflation at the lexical level, which manifests as increased acceptability of “incorrect” pronunciations in Catalan words. Ultimately, this coactivation would cause cross-linguistic influence on lexical items independent of the cognate status (Reference AmengualAmengual, 2016b).

A token is stored or encoded if it passes both the discriminability and the typicality evaluations. In Reference Todd, Pierrehumbert and HayTodd et al.’s (2019) model, the storage process overwrites a randomly selected exemplar within the same space. If bilinguals experience more variability in their environment, potentially by virtue of experiencing their languages spoken by individuals who vary in terms of their proficiency and accentedness, they could have consistently lower thresholds than monolingual individuals, leading to an exemplar space that has wider distributions.

We turn next to speech production, which first involves sampling from the pool of stored exemplars. The first pass is called type selection, which is the selection of the lexical item. target selection is the selection of an exemplar that provides an acoustic target. The sampling from the type and token space can pull a lexical item or pronunciation variant that has more or fewer sociocultural associations with one language or another, potentially affecting a bilingual’s target pronunciation. The language-specificity of the target selection may be influenced by language mode or the magnitude of language activation, where acoustic targets may be sampled from a single language distribution or a distribution composed of both languages or a single language distribution that is pulled by the competing language (Reference AmengualAmengual, 2018; Reference Brown and CoppleBrown & Copple, 2018; Reference Johnson and BabelJohnson & Babel, 2023; Reference Simonet and AmengualSimonet & Amengual, 2020).

The last two stages in production can nudge a pronunciation in different directions. First, a bias parameter conditions category behavior – in Reference Todd, Pierrehumbert and HayTodd et al.’s (2019) model this is a crucial aspect for modeling the dynamics of sound change, which we can broadly envision as a balance of stability (e.g., maintain categories) and change (e.g., a pull in a particular direction or dimension). In sound change, typically, socially structured variation within the community gets pushed or pulled in various ways, the direction of which may relate to the exact nature of the phonetic variation (Reference Harrington and SchielHarrington & Schiel, 2017; Reference Harrington, Kleber, Reubold, Schiel and StevensHarrington et al., 2018). Such a dynamic is not unlike the cross-linguistic influence in bilinguals (for a recent attempt to connect the sound change and cross-linguistic influence literatures, see Johnson and Babel, Reference Johnson and Babel2023). We see the bias stage as a prime opportunity for mutual influence in bilinguals. The final stage before an item is unleashed in the acoustic space is imprecision, which is random noise added to the target form. For bilinguals, we can imagine that an aspect of imprecision is an opportunity for the global activation of the nontarget language to bias and shift pronunciation. Language-specific articulatory settings and interspeech postures may manifest as cross-linguistic influence at either the bias or the imprecision stage (Reference Mennen, Scobbie, de Leeuw, Schaeffler and SchaefflerMennen et al., 2010; Reference Wilson and GickWilson & Gick, 2014;). This could provide an explanatory mechanism for phonetic drift in L1–L2 language dynamics (Reference ChangChang, 2012; Reference GuionGuion, 2003).

11.5.1 Similarity

The degree of cross-linguistic similarity is an important factor in exemplar-based models, as the similarity of languages and linguistic structures influences the kinds of connections that can be made within a bilingual (e.g., Reference Yao and ChangYao & Chang, 2016). Similarity is relevant at multiple dimensions within an exemplar model: phonetic forms, phonemic forms, and lexical similarity. A level of linguistic analysis on the part of the speaker-listener is relevant, as sound-based similarity relationships may be better accounted for at a certain level of abstraction, as opposed to phonetic similarity (Reference Chang, Yao, Haynes and RhodesChang et al., 2011; Reference FlegeFlege, 1987); see Reference Chang, Katz and AssmannChang (2019) for an overview.

The similarity between languages is continuous and multidimensional. Importantly, similarity along one dimension (e.g., phonemic categories) does not entail similarity along another dimension (e.g., existence of cognates). The scaling up of exemplar-based theories to account for bilingualism should also offer accounts and make predictions about bidialectalism, where cross-code similarity is presumably greater than cross-linguistic similarity. The question of how processes related to bidialectalism compare to those involving bilingualism is poorly understood as comparison between bilingual and bidialectal population groups has historically been avoided (Reference Chevrot, Ghimenton, De Houwer and OrtegaChevrot & Ghimenton, 2018). We suggest, however, that the omission of bidialectalism from theoretical discussions about similarity, cross-language competition, and exemplar models is a missed opportunity to better understand the similarity quantification(s) that individuals engage in.

11.5.2 Heterogeneity within and across Bilinguals

The challenge in an exemplar-based model for bilingualism is not simply the increased complexity from the addition of another language. There is also the fact that bilinguals may be vastly different from one another.

A large body of work both within and across languages and dialects has made clear that bilinguals are not all the same. Nor should a bilingual be considered two monolinguals – complete or incomplete – in one (Reference GrosjeanGrosjean, 1998). Bilinguals’ phonetic behaviors are informed by a number of (interacting) language factors, such as age and order of acquisition (e.g., Reference Canseco-Gonzalez, Brehm and BrickCanseco-Gonzalez et al., 2010; Reference Montrul and FooteMontrul & Foote, 2014), age of arrival (e.g., Reference Flege, Munro and MacKayFlege, Munro, & MacKay, 1995; Reference Flege, Yeni-Komshian and LiuFlege, Yeni-Komshian, & Liu, 1999), length of residence (e.g., Reference Flege and LiuFlege & Liu, 2001), sociocultural identity (e.g., Reference Szakay, Babel and KingSzakay et al., 2016), language usage patterns (e.g., Reference Bruggeman and CutlerBruggeman & Cutler, 2020; Reference Chan, Johnson and BabelChan et al., 2020), and language dominance (e.g., Reference Casillas and SimonetCasillas & Simonet, 2016; Reference Soo and MonahanSoo & Monahan, 2023), to name but a few. Altogether, the myriad factors that contribute to bilingual behavior and linguistic organization suggest that bilingualism is best conceptualized as a continuum.

That bilingual individuals vary is, of course, no surprise to any bilingual researcher. Heterogeneity even within controlled populations is the norm. Yet, in the context of exemplar-based models, this variation is particularly important. The factors described just now set the stage for the organization of the bilingual’s linguistic system and its evolution over time. The similarity assessments within a bilingual exemplar model and the cross-linguistic activation that colors mutual influence will be conditioned by the nature of the bilingualism, which for any individual will likely shift over time.