2.1 A Field at the Crossroads of Different Disciplines

As noted in the introduction, scholars in the field of L2 phonology examine how language users acquire a phonological system in a language that is not their native or home language. The study of this acquisition process touches upon elements from different linguistic subdisciplines and requires insights from different domains, including L2 acquisition, phonological theory, phonetics and sociolinguistics, as visualized in Figure 2.

Figure 2 Second language phonology at the intersection of different disciplines.

First, L2 phonology research is embedded in the field of L2 acquisition. In any type of language acquisition, three main types of ‘actors’ play a role: the learner, the context and the interlocutor or (target) recipient of the message that the learner wants to get across. A myriad of factors related to the learner impact the L2 learning process, including the learner’s native language, age, age of first exposure, aptitude, motivation, length of residence in the country/region of the target language, amount of L2 use and socio-economic and educational background. The relative impact of these factors will depend on the context in which the learning process takes place. Important factors related to the context are the naturalistic versus instructed setting in which learning takes place, and the amount and quality of exposure, including the number of different interlocutors. Finally, there is a growing body of research on native listeners’ perception of (L2) accented speech, comprehensibility (listeners’ self-reported level of understanding) and intelligibility (speech actually understood by the listeners). The speaker’s pronunciation (segmental properties, prosody and fluency) and lexicogrammar (richness of vocabulary, as well as accuracy and complexity of grammar) have been shown to be the two main linguistic dimensions influencing a speaker’s intelligibility and comprehensibility (Saito, Trofimovich & Isaacs, Reference Saito, Trofimovich and Isaacs2017). However, research on L2 phonology has increasingly recognized the impact of the interlocutor. Crucially, a learner’s L2 speech can only be called (in)comprehensible or (un)intelligible to a certain listener (Munro, Reference Munro, Edwards and Zampini2008; Simon, Lybaert & Plevoets, Reference Simon, Lybaert and Plevoets2022). As a result, it is essential to recognize the role of the listeners in L2 phonological acquisition, including listener characteristics related to, for instance, native language, age, familiarity with L2 speech, social and sociopolitical attitudes, socio-economic status and educational background.

Secondly, in order to address issues in L2 phonology, questions that are fundamental to phonological theory need to be raised. Such questions can be related to the nature of phonological units, the way these units are organized and their level of abstraction. In order to build models of L2 phonological acquisition, the issue of what exactly needs to be learnt by the L2 speaker-listener needs to be addressed. When L1 learners start building phonological representations, they do so from scratch (‘bottom up’), on the basis of phonetic input. Conversely, adult L2 leaners have a full set of L1 representations available, including their phonetic implementations. As a result, they may additionally use a ‘top-down’ approach in their processing of an L2, looking for correspondences between speech elements in their L1 and in the L2 (Flege, Munro & MacKay, Reference Flege, Munro and MacKay1995: 22).

Thirdly, research in L2 phonology draws on the field of phonetics, which is the science of how speech sounds are articulated, how they transfer through the air in the form of waveforms and what acoustic properties they have, and how they are perceived by the listener. Most current L2 speech acquisition studies start from the acoustic input: learners acquire an L2 sound system by being exposed to acoustic-phonetic forms (Flege & Bohn, Reference Flege, Bohn and Wayland2021). As such, work in L2 phonology will necessarily draw on insights in the closely related discipline of phonetics. Much of the earlier work in laboratory phonology was devoted to the question of how phonology is related to phonetics (Cohn, Fougeron & Huffman, Reference Cohn, Fougeron and Huffman2017) and this relationship remains a crucial factor in current discussions of L2 phonological research.

Finally, as variation is an inherent property of speech, one of the challenges faced by L2 learners is learning how to deal with this variation, which is at the same time linguistic and socio-indexical in nature. For instance, a vowel in a particular language may have different phonetic realizations, from more open to more closed and from more front to more back, depending on the flanking consonants (e.g. more fronted preceding a front consonant), but also on the speakers’ regional or social backgrounds. When studying patterns of L2 phonological acquisition, we therefore also need to take into account insights from sociolinguistics, or – more specifically – from sociophonetics.

In sum, the field of L2 phonology is a discipline situated at the interface of research on L2 acquisition, phonology, phonetics and sociolinguistics. Non-native listeners need to build robust phonological representations in their L2, which can be impeded by their native language phonology (Section 3), different listening conditions (Section 4) or sociophonetic variation (Section 5). We will explore these challenges to L2 learning and zoom in on a group of learners that may help us to understand the nature and development of L2 phonological representations, namely child and adult atypically developing language users, who have been claimed to have fuzzy representations in their L1 (Section 6). In the next section (Section 2.2), we go back to a basic question in theoretical phonology that is key to the field of L2 phonology: what needs to be learnt?

2.2 What Needs to Be Learnt?

When people learn a language that is not their native language, they have to learn a new sound system. Even when the language they are learning is similar to their native language(s) in terms of, for instance, the grammar or the lexicon, there will undoubtedly be differences in the way in which segments are produced and interact with each other. The target language may have a sound that does not occur in the native language or not in the same positions (i.e. phonotactics may be different), it may have connected speech processes, such as assimilation patterns, which do not apply in the native language, or it may differ in the concrete realization of sounds. Hence, the question emerges what exactly needs to be learnt by learners who need to master the sound system of an L2. As Flege and Bohn (Reference Flege, Bohn and Wayland2021) point out, earlier speech learning models, such as contrastive analysis, made predictions regarding areas of difficulty in L2 acquisition, which were based on comparisons of the sound inventories at the phonological level: phonemes which did not occur in the L1 of the learner were predicted to be difficult. As these predictions were often not borne out, the idea that the phonetic level needs to be taken into account gained ground in the field of L2 speech acquisition research. Indeed, most current L2 acquisition studies now start from the acoustic input: learners are exposed – either in naturalistic or instructed settings – to acoustic-phonetic forms and acquire phonetic categories on the basis of this input. Most prevalent speech learning models – including the Speech Learning Model (SLM, Flege, Reference Flege and Strange1995) and its revised version (SLM-r, Flege & Bohn, Reference Flege, Bohn and Wayland2021), the Perceptual Assimilation Model (PAM, Best, Reference Best and Strange1995) and its adaptation for L2 speech learning (PAM-L2, Best & Tyler, Reference Best, Tyler, Munro and Bohn2007) and the L2 Linguistic Perception Model (L2LP, Escudero, Reference Escudero2005) and its revised version (van Leussen & Escudero, Reference van Leussen and Escudero2015) – aim to account for how learners develop categories for L2 sounds on the basis of the phonetic input. In fact, in L2LP, it is even argued that a detailed acoustic comparison of L1 and L2 sounds can reliably predict L2 perception, since the acoustic properties of listeners’ L1 speech sounds will be the starting point of initial L2 perception (Elvin, Escudero & Vasiliev, Reference Elvin, Escudero and Vasiliev2014). A useful overview of current speech learning and perception models including PAM, SLM, the Native Language Magnet Model and Automatic Selective Perception is provided by Chang (Reference Chang, Schmid and Köpke2019). Baese-Berk et al. (Reference Baese-Berk, Chandrasekaran and Roark2022), focusing on perception, point out that most speech perception models rely on ‘category learning’. They define a category as ‘an abstract and generalizable representation that enables listeners to perceive highly variable acoustic input and efficiently process it through a more parsimonious representation with fewer perceptual dimensions than the raw sensory input’ (3026).

Indeed, learners are exposed to a great deal of phonetic variation in the input and thus need to build representations at a more abstract level. In the next sections we aim to present a coherent overview of how this process takes place, how representations develop during the acquisition process and when the process may be hampered.

3.1 Developing Phonological Representations in First and Second Languages

Before we turn to the question of how L2 learners acquire L2 categories, it is interesting to consider how the process unfolds in L1 acquisition. After all, when adult language users learn an L2, the phonological system of their L1 will to a great extent determine the initial state of the L2 acquisition process.

Research on L1 acquisition has firmly established that after the first six months of life, children start to develop language-specific categories as a result of perceptual attunement to the native language (Kuhl, Reference Kuhl1992; Polka & Werker, Reference Polka and Werker1994; Werker & Tees, Reference Werker and Tees1984; for a review, see Werker, Reference Werker2018). This attunement is assumed to be the result of an increased knowledge of the phonological status of relevant L1 distinctions. In Kuhl’s Perceptual Magnet Model (Kuhl, Reference Kuhl1992), infants deduce this information through exploiting statistical properties, in the sense that highly frequent and hence familiar speech sounds would function as magnets by attracting similar sounds and thereby diminishing the discrimination between these similar sounds (Werker, Reference Werker2018: 710). Throughout childhood and adolescence, speakers’ phonological representations keep developing: as children grow older and turn into adolescents, the way in which they use acoustic cues while listening becomes more adult-like and perception gradually becomes more categorical. By the time they reach adulthood, they have normally built well-established, robust categories in their L1.

When adults then acquire an L2 they need to develop a new set of categories for the L2. A large body of research on L2 perception and word recognition has amply demonstrated that this is not a trivial task. In fact, acquiring new categories for contrasts which are absent in the L1 is known to be very difficult, even for highly proficient L2 learners. Examples are studies on the /r/-/l/ contrast in L2 English for L1 Chinese listeners (Aoyama et al., Reference Aoyama, Flege, Guion, Akahane-Yamada and Yamada2004; Cutler & Otake, Reference Cutler and Otake2004) or the /ɛ/-/æ/ contrast in L2 English for L1 Dutch learners (Broersma, Reference Broersma2005; Escudero, Simon & Mitterer, Reference Escudero, Simon and Mitterer2012; Simon, Sjerps & Fikkert, Reference Simon, Sjerps and Fikkert2014). Previous research has, however, also established that L2 learners may be able to create new categories by shifting the boundaries of L1 categories in the direction of the L2 (Elman, Diehl & Buchwald, Reference Elman, Diehl and Buchwald1977; Flege & Eefting, Reference Flege and Eefting1987). According to models such as SLM-r and PAM-L2 (see Section 1), the likelihood of successful category creation is predicted to depend on the L2 category’s relation to existing L1 categories. In SLM-r, the L2 sound’s degree of perceived phonetic dissimilarity from the closest L2 sound will determine whether a new category is formed or not, in addition to the quantity and quality of L2 input and ‘the precision with which the closest L1 category is specified when L2 learning begins’ (Flege & Bohn, Reference Flege, Bohn and Wayland2021: 65). In the framework, category precision is defined by the degree of acoustic variability that is produced by a speaker in multiple productions of the category. The authors refer to a study by Kartushina and Frauenfelder (Reference Kartushina and Frauenfelder2013), which showed that Spanish learners of French whose L1 Spanish /e/ productions were ‘compact’, in the sense of revealing relatively little token-to-token variability, were better at identifying French /ɛ/ than Spanish speakers with a less precise Spanish /e/ category – that is, in which there were more spectral differences between different /e/ realizations. Flege and Bohn (Reference Flege, Bohn and Wayland2021) point out that category precision is linked to the distance between the category and categories which are adjacent in phonetic space, but may also vary between different speakers, possibly as the result of differences in, for instance, auditory acuity and working memory (39).

As pointed out by Chang (Reference Chang, Schmid and Köpke2019), the L1 and L2 are linked not only at a segmental level, as predicted by SLM-r, but also at a broader, systemic level. Evidence for this comes from studies demonstrating the impact that an L2 can have on features of the L1, a process typically referred to as ‘phonetic drift’ (see also Section 8.2). Phonetic drift has been found in both immersion contexts (e.g. Chang, Reference Chang2012, Reference Chang2013 on English–Korean interactions in L1 English participants immersed in the L2 in South Korea) and in L1-dominant environments (e.g. Wojtkowiak, Reference Wojtkowiak2022 on the impact of L2 English on L1 Polish learners of English in Poland). On a segmental level, both vowels and consonants can be affected. An example of consonantal drift can be found in Sancier and Fowler’s (Reference Sancier and Fowler1997) case study involving a female adult L1 Brazilian Portuguese speaker travelling back and forth between Brazil and the United States of America. An analysis of this person’s speech revealed that her Brazilian Portuguese stops drifted towards those of American English during a stay in the US (in terms of Voice Onset Time (VOT)), while her English stops drifted towards those of Portuguese when in Brazil. However, phonetic drift has also been shown to occur at a level above the segment. Guion (Reference Guion2003) examined vowel qualities in the entire vowel inventory in a group of Quichua–Spanish bilinguals and found that bilinguals who had acquired Spanish produced Quichua vowels with an overall lower first formant – that is, more closed or higher in the vowel space. In a different context, Chang (Reference Chang2012) examined the English speech of a group of L1 English speakers who enrolled in elementary Korean classes and who were inexperienced, novice learners of the L2. Similarly to Guion’s (Reference Guion2003) results, the acoustic analyses of the vowels revealed a general raising of the English vowels in this group of speakers. In neither of the studies could the drift be explained as shifts in individual vowels, as the L1 vowels did not all assimilate towards or dissimilate from the closest L2 vowels. Instead, Chang (Reference Chang2012, Reference Chang, Schmid and Köpke2019) argues that the shift should be explained as the result of convergence towards the global formant levels of the L2, thereby transcending the level of the segment. Evidence for phonetic drift at the suprasegmental level can also be found in Mennen et al.’s (Reference Mennen, Reubold, Endes and Mayr2022) study on the impact of L2 intonational patterns on the L1. Specifically, they examined different intonation dimensions in the English speech of adult L1 English migrants to Austria. The results showed that L2 intonation patterns impacted those in the L1, which exhibited values in between the L1 or the L2 or completely assimilated to the L2. These results demonstrate a lasting plasticity of the L1 intonation in late sequential bilinguals.

Together, studies on phonetic drift in L2 learners or sequential bilinguals provide evidence for a strong linkage between the L1 and the L2. They demonstrate that phonetic drift takes place on an allophone-to-allophone level, predicted by acoustic similarity and perceptual assimilation mappings, but also on a systemic level, including vocalic drift of the entire inventory or drift on a suprasegmental level. A comprehensive review of older and more recent phonetic drift literature can be found in an overview chapter by Chang (Reference Chang, Schmid and Köpke2019).

3.2 How Robust Is the Learning?

One of the long-standing issues in L2 phonological acquisition has been that of complete or ultimate attainment, which refers to the question of whether L2 learners can reach a level of phonological competence in the L2 that is similar to or the same as that of L1 native speaker-listeners. Discussions on this issue were often related to testing the critical period hypothesis. This hypothesis, proposed by Lenneberg (Reference Lenneberg1967), states that there is a critical age after which full attainment of L2 phonology is impossible or – in milder versions of the hypothesis – more challenging, due to maturational changes in the brain. While the idea of a critical period is more readily accepted in the acquisition of an L1, it is less clearly outlined in L2 acquisition and several studies have provided compelling evidence against it (see, e.g., Zhang & Wang, Reference Zhang and Wang2007 on neural plasticity in speech acquisition). The hypothesis was tested in L2 acquisition studies comparing L2 learners’ performance with that of native speakers and examining the effect of variables like the learner’s age, motivation and the amount of instruction on the L2 performance (e.g. Moyer, Reference Moyer1999). After previously focusing on ‘end state’ learning in the original version of the SLM, the SLM-r argues that in order to examine whether L2 learners acquire new categories, studies should focus on the early stages of L2 acquisition rather than on highly experienced learners. Using production data by Italian immigrant learners in Canada, they argue that the patterns displayed by these learners indeed stabilize, but that ‘this does not place an upper limit on the human capacity for learning speech when phonetic input changes’ (Flege & Bohn, Reference Flege, Bohn and Wayland2021: 28).

The ultimate attainment debate is now no longer central in L2 phonological acquisition research, but a related question that also lies at the basis of this debate is how far L2 learning actually goes. After all, if L2 learners can reach L1 levels of attainment, this implies that the L2 learning is as profound as L1 learning and that the phonological representations developed by L1 and L2 learners are equally robust. However, while researchers agree that L2 learners are – at least to some extent – able to acquire L2 speech sounds, there is as yet no consensus on how far the learning goes. Do L2 learners effectively create new, abstract categories or is the learning more superficial – that is, do learners temporarily shift some category boundaries when listening to L2 speech, without actually creating robust L2 categories that can be used in perception as well as in production? Barrientos (Reference Barrientos2021) addresses this question in her study on the perception of the American English vowel contrast between /ɑ/ (‘cop’ /kɑp/) and /ʌ/ (‘cup’ /kʌp/) by L1 Spanish listeners. The results showed that the advanced learners displayed an increased sensitivity towards acoustic properties relevant in the L2 and an improved discrimination between the two vowels. However, the author argues that they did not develop a phonemic split at the representational level, meaning that the learning did not go far enough for the L2 listeners to develop two separate phonemes for /ɑ/ and /ʌ/. The claim is based on the absence of an S-curve in the L2 learners’ vowel discrimination plots, which suggest that there are no two distinct phonemic representations at the end points to which vowels are categorically mapped.

Answers to the question of how robust L2 learning is can also be sought in training studies, in which L2 learners typically receive intensive phonetic training on L2 contrasts which are absent in the L1. A large number of empirical studies point towards the positive effect of such training sessions on perception and/or production, often using the High Variability Phonetic Training (HVPT) paradigm, in which listeners are exposed to multiple talkers (but see Zhang et al. Reference Zhang, Cheng, Qin and Zhang2021 and Xie, Liu & Jaeger, Reference Xie, Liu and Jaeger2021, who address the question of whether multiple talkers are strictly necessary). However, questions about the robustness of these training effects have arisen, specifically about whether training leads to long-term effects on phonological representations or rather to short-term effects on perception. In a recent study (Simon, De Clercq et al., Reference Simon, De Clercq, Degrave, Decourcelle, Alves and Imaregna Alcantara de Albuquerque2022), we examined the effects of multimodal HVPT sessions with two groups of adult L1 French learners of Dutch in Belgium: a group of university students and a group of final year secondary school students. The study focused on five Dutch sound contrasts which are absent in French – namely (1) /i/ vs /ɪ/, (2) /ɑ/ vs /aː/, (3) /ə/ vs ø, (4) /x/ vs /k/, and (5) /h/ vs ø. The design comprised a pretest, a post-test and a delayed post-test, with five training sessions as the intervention for the experimental group and no training for the control group. The training sessions consisted of perceptual identification tasks with feedback and metalinguistic information on Dutch sounds, including articulatory and acoustic information in the form of videos and waveforms. Specifically, we examined the robustness of the training effects, defined as the generalizability of the training to novel tokens and talkers, the effect of training in non-optimal conditions (i.e. with noise added to the signal), and – crucially – the long-term retention (Simon, De Clercq et al., Reference Simon, De Clercq, Degrave, Decourcelle, Alves and Imaregna Alcantara de Albuquerque2022, in an edited volume on Second Language Pronunciation, Alves & Albuquerque, Reference Alves and Alcantara de Albuquerque2022). Long-term retention of learning is considered to be a signal that there have been long-term modifications to the phonetic or phonological categories as a result of the phonetic training (Rato & Oliveira, Reference Rato, Oliveira, Alves and Albuquerque2022). The results revealed variability along most robustness variables, which were to a large extent related to learner characteristics and type of L2 contrast. For instance, we observed training effects in the group of university students (on which most HVPT studies are based) along all three robustness variables, including long-term retention, but not in the group of secondary school pupils. This may have been related to the learners’ proficiency levels or to their motivation to participate in the study. A systematic review of twenty-seven training studies which include the testing of generalization and/or retention is provided by Rato and Oliveira (Reference Rato, Oliveira, Alves and Albuquerque2022), which appeared in the same edited volume. The authors conclude that all reviewed studies that tested generalizability (twenty-five out of twenty-seven studies) found carry-over effects to new stimuli, phonetic contexts, talkers and/or tasks. The same trend was observed for retention, which was tested in only eleven out of the twenty-seven papers. However, in only four out of eleven studies retention was reported for all conditions.

It should be noted that the HVPT paradigm represents one specific type of training, which is both supervised and attentive, in the sense that categories are made explicit as part of the training design. When participants are, for instance, trained on the different vowels in ‘bed’ and ‘bad’, the category labels are explicitly used during the training. This can be done through the use of the phonetic symbols (in this case /ɛ/-/æ/), through linking the phonemes to letters (e.g. <e> for/ɛ/ and <a> for /æ/), or through the consistent use of keywords representing the sounds (e.g. BED and BAD). In this kind of training, participants’ attention is explicitly drawn to the contrastive speech sounds. However, learning can also take place in an unsupervised way by exposing learners to a large number of speech sounds whose distribution provides information on the categories. Chládková, Boersma and Escudero (Reference Chládková, Boersma and Escudero2022: 1) define distributional learning as ‘an unsupervised statistical learning mechanism that works through exposure to the probability distributions of speech sounds in one’s environment’. They note that distributional learning was originally described as an unsupervised learning mechanism used by infants to develop categories for native speech sounds. Maye, Werker and Gerken (Reference Maye, Werker and Gerken2002) conducted a laboratory study with American infants growing up in an English-speaking environment. The study showed that infants exposed to a bimodal distribution between a prevoiced [d] and a voiceless unaspirated [t], which both fall within the same native laryngeal category (English /t/), were better at discriminating between these two non-native categories compared to infants who had been exposed to a unimodal distribution during the training phase. Distributional learning paradigms have also been used with adults, with varying degrees of success, suggesting that adult learners may employ distributional learning to perceptually adapt to specific accents or to speech in noisy environments (for an overview of these studies, see Chládková et al., Reference Chládková, Boersma and Escudero2022). In their own study, Chládková et al. (Reference Chládková, Boersma and Escudero2022) examined the effect of an unattended distributional training in the L1 on the creation of another boundary location between existing vowels (i.e. a shift of the boundary between /i/ and /e/ in Spanish) and on the creation of a new boundary (i.e. a new boundary between short and long vowels, which are not contrasted in Spanish). On the basis of neurological responses measured through electroencephalography (EEG), the authors concluded that perceptual learning took place in the case of boundary shift, but not (or hardly) in the case of a new boundary creation.

3.3 Concluding Remarks

In sum, adult L2 learners face a difficult task, as they have already built robust phonological representations in their L1 when they are first exposed to the sound system of an L2. The learners’ task is particularly challenging when the L2 contains sound contrasts which are not phonemic in the L1, in which case learners need to create an additional category or categories. Although the task is challenging, research has demonstrated that (advanced) learners may be able to create new categories by shifting the boundaries of existing categories. Such shifts may take place as the result of continued exposure to L2 speech over an extended period of time and the process may be aided by a growing L2 vocabulary, as their perception will be guided by the lexicon (see Tyler, Reference Tyler, Nyvad, Hejná, Højen, Jespersen and Sørensen2019 on phonological acquisition in the foreign language classroom). One way in which this learning process may be enhanced is through explicit phonetic training. Indeed, an accumulated body of research on L2 phonetic training suggests that phonetic (HVPT) training can lead to robust learning of non-native categories, which can be carried over to new contexts and talkers and which can be retained after a longer time period following the last training. However, a cautionary note is in place here: although there is now ample evidence that training can alter L2 listeners’ perceptions, it is clear that it is by no means a trivial task. First, as discussed, long-term retention has been examined in only a small subset of training studies and is generally not found in all conditions – that is, for all groups and for all contrasts learners were trained on. Secondly, as noted by Rato and Oliveira (Reference Rato, Oliveira, Alves and Albuquerque2022), there may be a bias in the training studies that get published. In all fields of research, experimental studies that do observe positive effects generally have a higher chance of being submitted and/or accepted for publication compared to studies that report null results of an intervention (Fanelli, Reference Fanelli2012). This phenomenon is sometimes referred to as ‘the file drawer effect’, because the unpublished manuscripts are imagined to be gathering dust in researchers’ desk drawers. Rato and Oliveira (Reference Rato, Oliveira, Alves and Albuquerque2022) therefore suggest that it is in this context valuable to examine results in so-called grey literature, such as proceedings and unpublished reports.

In Section 4, we discuss how phonological representations need not only be robust, but also flexible, meaning that listeners can deal with the high variability in the phonetic realization of categories that they may be exposed to.

4.1 The Challenge of Variation: Non-optimal Listening Conditions

Variation is an inherent characteristic of spoken language and a listener hence needs to be able to flexibly adapt to varying acoustic signals. Speech variation comes in various forms and occurs at different levels, and not all types of variation present a challenge to the listener. In fact, most types of variation are generally completely unproblematic. For instance, a major type of variation that is inherent in any speech produced by more than one individual is due to anatomical differences between talkers. The acoustics of speech are largely formed by movements of the articulators in the vocal tract and by the size of the vocal tract and larynx. As the size of the vocal tract and larynx may considerably differ between speakers, the acoustic output will necessarily vary between talkers. For some automatic speech recognition and speaker verification models to work, these talker differences need to be cancelled out or normalized (Johnson, Reference Johnson2020). This is equally the case for linguistic studies that aim to compare, for instance, vowel formants across speakers. For that purpose, different mathematical models have been proposed to normalize vowels, some of which use vowel-internal and some vowel-external (i.e. comparison between vowels) normalization procedures. Another distinction is that between talker-intrinsic and talker-extrinsic normalization, in which data are normalized on the basis of one versus across multiple speakers (see, e.g., NORM: Vowel Normalization Suite 1.1, uoregon.edu). As Johnson (Reference Johnson2020) points out, it remains an open question whether or not listeners use vocal tract length in speech perception. What is important from our perspective is, however, that listeners clearly have no difficulties with this type of variation and can map all the different acoustic outputs of these unique voices onto more abstract speech representations. In fact, listeners are extremely fast in detecting, for instance, speaker gender: a study by Owren, Berkowitz and Bachorowski (Reference Owren, Berkowitz and Bachorowski2007) showed that listeners are fairly accurate in speaker gender/sex judgements after exposure to less than 15 ms of speech. In a large-scale study including 100 adult voices and 618 listeners, Lavan (Reference Lavan2023) showed that judgements on physical characteristics of talkers (age, gender and health) emerge fastest, after only 25 ms exposure, compared to traits (‘trustworthiness’, ‘dominance’, ‘attractiveness’), and social characteristics (‘educatedness’, ‘poshness’, ‘professionalism’), which are formed after 800 ms of exposure to single vowel productions.

However, other types of variation may pose difficulties even to native listeners. Conditions in which speech variation emerges that causes challenges for speech perception or speech recognition are sometimes called ‘adverse conditions’ (Guediche et al., Reference Guediche, Blumstein, Fiez and Holt2014; Mattys et al., Reference 71Mattys, Davis, Bradlow and Scott2012) or ‘non-optimal conditions’ (Simon et al., Reference Simon, Lybaert and Plevoets2022). These adverse or non-optimal conditions can take multiple forms, depending on whether they relate to the source of speech, the environment or transmission of speech, or the receiver, as categorized by Mattys et al. (Reference 71Mattys, Davis, Bradlow and Scott2012). Both source and environmental degradation lead to variation in the acoustic signal that reaches the listener. One interesting type of source gradation occurs when the speaker has an unfamiliar accent and produces ‘accented speech’, defined – following Munro and Derwing (Reference Munro and Derwing1995a) – as speech that noticeably deviates from native speaker norms. By contrast, environmental degradation refers to distortions due to, for instance, noise resulting from competing talkers in the background or a noisy environment. Under both conditions, speech perception is more difficult compared to optimal listening conditions, in which the listener is familiar with the speaker’s accent and the communication is set in a quiet environment. In the case of competing background talkers, a listener’s brain is – to a greater or smaller extent depending on the individual – able to filter out irrelevant speech through selective auditory attention, a phenomenon commonly known as ‘the cocktail party effect’, a term coined by Cherry (Reference Cherry1953; see also Kattner & Ellermeier, Reference Kattner and Ellermeier2020). In the case of accented speech, by contrast, it is not the context that is responsible for the degradation of speech, but the speakers themselves. In that case, the deviations from the listener’s norm will normally be systematic, in the sense that speakers will tend to deviate from the norm in similar ways, such as in the fronting of a vowel for which the listener expects to have a more posterior place of articulation or the substitution of a dental fricative by a stop. Even though there may be variability in the phonetic realization of particular segments in regional or non-native accented speech (e.g. Turner, Reference Turner2024 on the production of French /u/ and /y/ by L1 English speakers), listeners are typically quite good at adjusting their perception to relatively systematic deviations from the expected realizations, a process commonly referred to as ‘perceptual adaptation’.

4.2 L1 Listeners’ Perceptual Adaptation to Accented Speech

When we are listening to speech, our perception is aided by our knowledge of the lexicon. An early study that provided evidence for this statement is the seminal work by Ganong (Reference Ganong1980). In this study, listeners were exposed to auditory continua from, for instance, the word task to the non-word dask. Results revealed a significant lexical effect – that is, for ambiguous tokens, listeners showed a tendency to categorize sounds in such a way that the stimuli formed words rather than non-words. Two decades later, Norris, McQueen and Cutler (Reference Norris, McQueen and Cutler2003) were presumably the first to provide experimental evidence for the effect of lexical knowledge on perceptual learning. In their study, L1 Dutch listeners were first familiarized with Dutch words containing ambiguous realizations of either [f] (one group) or [s] (another group) as the final fricative. For instance, the /f/-final word ‘witloof’ (chicory) was realized as [wɪtlo?] and the /s/-final word ‘naaldbos’ (pine forest) was realized as [na:ldbɔ?], with [?] being a sound in-between [f] and [s]. The results of a subsequent phoneme categorization task revealed that listeners who had been trained on [f]-ambiguous words (type ‘witloof’) classified ambiguous tokens more often as [f], while the reverse pattern was true for the [s]-ambiguous group. A control group, who had been trained on non-words ending in an ambiguous fricative (and who could hence not be guided by lexical knowledge), showed a pattern intermediate between the two experimental groups, without a bias for either [f] or [s]. These results provide clear evidence that listeners in the experimental groups had used their lexical knowledge (e.g. that ‘witloof’, but not ‘witloos’ is a word in Dutch) and that lexical information is thus able to modify phonetic categorization and change the perceptual system.

The question which arises is what precisely happens when listeners adapt their perception to unfamiliar accents, be they regional, native or non-native accents. In exemplar-based models (see, e.g., an early publication by Johnson, Reference Johnson, Johnson and Mullennix1997), episodic traces of talker- or context-specific elements are stored in the lexicon and retrieved when speech from the same talker or context is being processed. As such, as Dahan, Drucker and Scarborough (Reference Dahan, Drucker and Scarborough2008) point out, talker adaptation is an inherent part of such theories. By contrast, in theories that work with representations that have abstracted away from specific experiences with talkers or contexts, talker adaptation requires a temporary adjustment of these representations (Dahan et al., Reference Dahan, Drucker and Scarborough2008: 711). In their study, Dahan et al. (Reference Dahan, Drucker and Scarborough2008) exposed American English listeners to an American English accent with which the participants were unfamiliar. In that accent, the vowel /æ/ is raised before a /g/ (as in ‘bag’), but not before a /k/ (as in ‘back’). Results showed that identification of /k/-final words was facilitated when participants had previously been exposed to /g/-words. In other words, perceptual adaptation took place even in the absence of exposure to the acoustic signal, suggesting that the absence of vowel raising in /k/-words was stored in the lexicon. The authors take these results to suggest that perceptual adaptation actually brings about at least a temporary adaptation of representations in the mental lexicon.

Bent and Baese-Berk (Reference Bent, Baese-Berk, Pardo, Nygaard, Remez and Pisoni2021) discuss two main types of perceptual adaptation to non-native accented speech: the first involves the remapping of sound categories, the second the shifting of category boundaries. A remapping is necessary when the speaker substitutes one category by another one – for instance, when replacing English /θ/ by an alveolar plosive /t/ in words like think and bath. In that case, the listener needs to learn this rule and map the [t] sound onto the phonological representation of the dental fricative. Shifting boundaries is necessary when the speaker makes a contrast between two categories, but places the boundary in a different position compared to native speakers. A typical example would be the placement of the VOT boundary between two laryngeal categories of stops. If native Dutch speakers transfer their laryngeal contrast into English, they would produce voiced plosives with pre-voicing and aspirated stops as short-lag ones (Simon, Reference Simon2010). Native English speakers listening to Dutch-accented English would then need to shift the boundary of the voiceless plosive category from a high positive VOT to a VOT close to zero.

When exposed to accented speech, a listener thus essentially has to interpret the acoustic-phonetic signal in a new way, so that it is mapped onto the phonological category intended by the non-native speaker. The ease with which the listener performs this task may depend in part on the extent of familiarity of the listener with non-native accented speech. Studies on the effect of long-term linguistic experience with the perception of accented speech seem to show mixed results, as reviewed by Bent and Baese-Berk (Reference Bent, Baese-Berk, Pardo, Nygaard, Remez and Pisoni2021). Kennedy and Trofimovich (Reference Kennedy and Trofimovich2008) showed that English as a Second Language (ESL) teachers were better able to understand non-native accented speech compared to naïve listeners who reported little to no contact with L2 English speakers. Interestingly, the researchers manipulated the accented sentences in such a way that listeners could or could not rely on the lexical context. The results revealed that ESL teachers did not rely more on top-down information, such as the lexical context, but were in fact better at mapping the unexpected acoustic-phonetic signal onto the intended phonological representations, leading to a better word recognition. It should be noted that Kennedy and Trofimovich’s (Reference Kennedy and Trofimovich2008) study is based on a limited number (twenty-four in total) of participants. However, these findings tentatively suggest that if listeners are frequently exposed to non-native accented speech, their phonological representations become more flexible, in the sense that they become better at shifting the boundaries, expanding the boundaries or reweighing cues compared to listeners who are less frequently exposed to unfamiliar accents.

It is, in this context, noteworthy that in more naturalistic contexts, a high degree of exposure should be sufficient for perceptual adaptation to take place. When there is sufficient exposure and listeners have some time to adjust their perception to regular deviations from the expected norm, accented speech will not typically present a challenge to L1 listeners. For L2 listeners, by contrast, the situation may be different. In Section 4.3, we discuss why this may be the case.

4.3 The Double Challenge for Second Language Listeners

While native adult listeners are generally successful at adapting to speech in the face of variation in the acoustic signal, non-native listeners may experience more problems. Indeed, Mattys et al. (Reference 71Mattys, Davis, Bradlow and Scott2012) mention an incomplete language model as one of the ‘adverse conditions’ in listening and argue that, leaving aside children who are still developing their L1 system, this case is best represented by non-native listeners, who have an incomplete knowledge of the language, not only at the phonological, but also at the lexical, morpho-syntactic, grammatical and idiomatic levels. As our literature review confirmed, perceptual adaptation by native listeners seems to be facilitated by, first, the lexicon and, second, by experience or familiarity. Since non-native listeners typically have a smaller vocabulary and are less familiar with the target accent, because they have received less naturalistic exposure, they are less aided by (most crucially) the lexicon or previous experience, compared to native listeners. That the lexicon plays a role in perception by (L2) listeners has been demonstrated in studies providing L2 listeners with subtitles while listening to non-native speech. A study by Mitterer and McQueen (Reference Mitterer and McQueen2009), for instance, showed that L2 subtitles but not L1 subtitles facilitated the perception of unfamiliar Australian and Scottish accents by L1 Dutch listeners. Subtitles provide lexical information and when the subtitles were in the L2, they helped listeners with word recognition.

Although there seems to be consensus on the observation that listening to accented speech is generally more difficult in an L2, relatively little research has so far been carried out on whether or how L2 listeners deal with accented speech in the language they are learning. An exception is a recent study by Lee and Baese-Berk (Reference Lee and Baese-Berk2021), who examined whether non-native listener groups with different L1s adapted to L2 speech in a high- and low-variability talker condition. The results revealed that, first, L2 listeners showed adaptation to the American English accent after only limited exposure (as measured through reaction times on word recognition), and, second, that high variability in the sense of different talkers initially posed an additional processing cost for L2 listeners. However, in the end, exposure to multiple talkers enabled cross-talker generalization. An earlier study by Bruggeman and Cutler (Reference Bruggeman and Cutler2019) investigated perceptual adaptation to ambiguous speech sounds in the L1 (Dutch) and the L2 (English) of Dutch emigrants to Australia, who were all highly proficient in English and used Dutch with a limited group of relatives at home. The results revealed perceptual adaptation to novel L2 talkers, but not to novel L1 talkers. These findings suggest that perceptual adaptation to accented speech is possible, but flexibility needs to be practised through regular exposure to variation.

It should be noted that listening in an L2 is not always more difficult than listening in an L1, and that non-native listeners are not always disadvantaged compared to native listeners. Several studies have indeed reported an ‘interlanguage speech intelligibility benefit’ (henceforth ISIB), referring to a perceptual benefit when the speaker and the listener have the same L1 and hence share an interlanguage system, with similar L2 phonological representations. As discussed by Hayes-Harb et al. (Reference Hayes-Harb, Smith, Bent and Bradlow2008), the ISIB can refer to two different situations. First, some studies have demonstrated that non-native listeners are equally good or better at perceiving non-native speech compared to native speech, if the listeners share an L1 with the non-native talkers. Hayes-Harb et al. (Reference Hayes-Harb, Smith, Bent and Bradlow2008) refer to this as a benefit for talkers, or ISIB-T. For instance, Bent and Bradlow (Reference Bent and Bradlow2003) conducted a study on the perception of English speech produced by L1 English, L1 Chinese and L1 Korean talkers. The listeners had either English, Chinese, Korean or a set of other languages as their L1. The results of a sentence transcription task showed that the relative intelligibility of each talker depended on the language background of the listeners. Crucially, non-native talkers with the same L1 as the listeners were equally intelligible to these listeners as native talkers. Secondly, non-native listeners may have an advantage over native listeners when listening to L2 speech that is produced by a speaker with whom they share an L1, a process referred to as ISIB-L: a benefit for listeners. Evidence for ISIB-L was found by, among others, Hayes-Harb et al. (Reference Hayes-Harb, Smith, Bent and Bradlow2008), who analysed the perception of Mandarin-accented English speech by L1 English and L1 Mandarin Chinese listeners. They found that English speech produced by low-proficiency talkers was more intelligible (as measured through a forced-choice identification task) to low-phonological-proficiency listeners of L1 Mandarin Chinese, compared to L1 English listeners. To conclude, while there is sufficient evidence to claim that an ISIB can take place, both for listeners and for talkers, it is worth pointing out that it certainly cannot always be observed and that it seems to be modulated by factors such as talker and listener phonological proficiency and degree of accentedness (Fishero, Sereno & Jongman, Reference Fishero, Jongman and Sereno2023; Xie & Fowler, Reference Xie and Fowler2013).

As mentioned, another type of listener with a yet incomplete language system is children whose L1 is still developing. Just like non-native listeners, child native listeners have a smaller vocabulary and less experience with the target language compared to adult native listeners. Research on perceptual adaptation in child native listeners has demonstrated that by the age of twenty-eight months, toddlers can recognize words and sentences produced in an unfamiliar dialect, even without prior exposure to the dialect (van Heugten & Johnson, Reference van Heugten and Johnson2016). Bent and Baese-Berk (Reference Bent, Baese-Berk, Pardo, Nygaard, Remez and Pisoni2021) discuss the question of what precisely children do when they adapt to these unfamiliar accents: do they apply rules specific to the accented speech of the talkers, or do they apply general expansion rules, loosening up the criteria for category membership? A study by White and Aslin (Reference White and Aslin2011) seems to suggest the former, as children were – after exposure – successful in recognizing words with shifted vowels and could generalize to words with similar vowel shifts, but were not more liberal in accepting deviant pronunciations in general. However, as pointed out by Bent and Baese-Berk (Reference Bent, Baese-Berk, Pardo, Nygaard, Remez and Pisoni2021: 445), children (and adults) may use different strategies in lab experiments commonly used to examine perceptual adaptation compared to more naturalistic exposure to longer stretches of non-native speech. In more naturalistic non-native speech, there is never just one deviating feature, such as a vowel shift, but rather a cluster of features that differ from the default. In such circumstances, it is possible that listeners would apply more general strategies in which they relax the membership criteria in order to cope with the accented speech. This latter view is taken by Melguy and Johnson (Reference Melguy and Johnson2021) in their study on L1 adult listeners who were presented with Chinese-accented English speech. American English listeners were asked to transcribe Chinese-accented sentences in noise and were assigned to different visual conditions: a blank silhouette, a European face, an East Asian face or a South Asian face. Participants who believed the speaker to be non-natively accented significantly outperformed participants who reported to hear a native accent, but no advantage was observed for listeners who correctly identified the speaker as Chinese-accented compared to other foreign accents. The authors suggest that these results point in the direction of a general relaxation of category boundaries or a general expansion of categories. The results of another study by the same authors (Melguy & Johnson, Reference Melguy and Johnson2022) point in the same direction: listeners trained on ambiguous sounds showed a general category expansion, though this expansion proved to be limited by phonetic similarity – that is, if the sounds in testing were too dissimilar from the ones in training, listeners did not include them in the target category.

4.4 Concluding Remarks

In sum, although non-optimal listening conditions may require temporal adjustment by native listeners or may slow down speech recognition, they often present little actual challenge to intelligibility thanks to listeners’ rapid perceptual adaptation. This is especially true for variation resulting from talker-specific characteristics, regional accents or mildly non-native accented speech. For more strongly accented non-native speech, listeners can also improve their understanding, but gains may be more modest. As noted by Bent and Baese-Berk (Reference Bent, Baese-Berk, Pardo, Nygaard, Remez and Pisoni2021), these results point towards the flexibility of the perceptual system, but also to the remaining existence of differences in the extent to which non-native speakers are intelligible.

When L2 learners listen to either regionally or non-native accented speech, they may be doubly burdened: they need to map the L2 sounds onto the intended categories, which may differ from their L1 categories, and they need to adjust their expectations about the phonetic realizations of these L2 categories on the basis of the input. Whether L2 listeners apply certain rules when listening to accented speech, such as shifting the boundary between particular vowels, or apply general expansion rules, relaxing the criteria for category membership, is a matter in need of further investigation.

In Section 5, we delve a little deeper into the matter of accent variation and more generally discuss what the study of social and regional language variation brings to the field of L2 phonology.

5.1 Sociophonetics and Phonological Theory: Is There a Match?

Sociolinguistically relevant similarities and differences between accents of different speakers and speaker communities form the object of study in the field of sociophonetics, a discipline at the interface between sociolinguistics and the phonetic study of speech and one that is currently a growing area of new research (see Kendall et al., Reference Kendall, Pharao, Stuart-Smith and Vaughn2023 for an overview of recent advancements in the field). Foulkes and Docherty (Reference Foulkes and Docherty2006: 411) define sociophonetic variation as ‘variable aspects of phonetic or phonological structure in which alternative forms correlate with social factors’, which ‘include most obviously those social categories which have been examined extensively by sociolinguists and dialectologists: speaker gender, age, ethnicity, social class, group affiliations, geographical origin, and speaking style’. Whenever we listen to speech, we derive not only linguistic, but also socio-indexical information: these are invariably present together in any utterance. For instance, when listening to the speaker in Speech sample 4, listeners will deduce that the speaker is male and many will also be able to identify his accent as Southern American English. Similarly, many listeners will hear that the speaker of Speech sample 5 is not a native speaker of English and some will deduce from the speech signal that his native language is Spanish. Finally, segmental as well as prosodic features in Speech sample 6, taken from the Phonologie du français contemporain (PFC) corpus, may allow listeners to identify this French speaker as coming from the Marseille region in the south of France. Coquillon (Reference Coquillon2007) contains a detailed discussion of phonetic features in this speaker’s speech that are typical for Marseille French, such as the long nasal vowels ending in a nasal consonant.

As noted by Kendall and Fridland (Reference Kendall and Fridland2021), the field of sociophonetics is necessarily empirical, but is driven by broad questions which are deeply rooted in the field of linguistics, such as:

By addressing such questions, it is argued, sociophonetic studies may be able to build a bridge between empirically driven analyses of speech and cognitive-linguistic questions about the production and perception of speech. As Thomas (Reference Thomas, Chambers and Schilling-Estes2013) notes, ‘[e]xperiments involving sociolinguistic variants can also be used to test the cognition of phonology’ (110). Foulkes and Docherty (Reference Foulkes and Docherty2006) highlight the minor role sociophonetics has played in the development of phonological theory. But how easy is it to actually make the link between sociophonetics and phonological theory? Which sociophonetic experiments can be set up to get insight into language users’ phonological system?

In a review paper, Foulkes and Docherty (Reference Foulkes and Docherty2006) assess to what extent sociophonetic studies can inform theory building in phonetics and phonology. They argue that traditionally, social factors, responsible for differences between speakers, have not been included in studies aimed at advancing our knowledge of phonology. They ascribe this exclusion to the influence of generative-based models, in which an ideal speaker-hearer is assumed and variation is considered irrelevant for theory building (see Goldrick & Cole, Reference Goldrick and Cole2023 for a discussion). However, usage-based approaches in the framework of exemplar theory may be more apt at dealing with sociophonetic variation. Thomas (Reference Thomas, Chambers and Schilling-Estes2013) suggests that sociophonetic experiments on style-shifting and the intelligibility of dialects ‘could help resolve the extent to which Exemplar Theory accounts for phonological cognition’ (122), though exactly how the link should be made is left to explore.

Exemplar-based models of memory have a long history in cognitive psychology, but have been proposed for the study of language for some decades as well now (Johnson, Reference Johnson, Solé, Beddor and Ohala2007). As Johnson (Reference Johnson, Solé, Beddor and Ohala2007) points out, there is no one ‘Exemplar Theory’, but rather different exemplar-based models, which have in common that they rely on the assumption that language users store traces of all the instances they encounter in their memory without any abstraction. As these instances or exemplars (or traces thereof) are stored with detailed phonetic information, exemplar-based approaches to phonology are often considered to be suitable frameworks for modelling sociophonetic variation. In exemplar-based models, acoustic input is processed by comparing it to previously encountered stimuli with which the new input is compared. As a consequence, frequency of exposure and recency of exposure are crucial elements in these usage-based learning models (Frisch, Reference Frisch, Hannahs and Bosch2017). As the relation between linguistic form and sociophonetic variation is arbitrary, language users indeed need to be exposed to the variation to a sufficient degree and the exposure needs to be sufficiently recent. As type and amount of exposure to particular categories of phonetic variation – such as different accents – will differ over a language user’s lifespan, it may mean that the ‘representation of phonological knowledge may be less stable, even in adulthood, than is typically assumed’ (Foulkes & Docherty, Reference Foulkes and Docherty2006: 434). However, if one assumes a hybrid model, in which there are two layers, there may be one layer with stable phonological knowledge, and one layer with flexible forms, ever-changing as exposure and frequency of exposure change. Such a model is proposed by Pierrehumbert (Reference Pierrehumbert2016), who argues that an abstract level of representations is needed in phonology to account for language users’ ability to generate and process novel forms, but needs to be complemented by detailed phonetic characteristics. Phonetic characteristics of words are highly variable, as they are affected by several context effects. Such contextual factors include the position of the word in the phrase relative to the element that receives phrasal stress, the frequency and predictability of the word in a phrase, the individual speaker’s anatomical and dialectal features, as well as social context factors, such as the speaker’s social class or register within a particular social context (Pierrehumbert, Reference Pierrehumbert2016). As Pierrehumbert (Reference Pierrehumbert2016) argues, ‘[t]he ability to perceive indexical features, produce them more or less in different contexts, and generalize them to new words and new interlocutors means that they must be cognitively represented’ (43). If they are cognitively represented and stored in memory, they need to be part of phonological theory. A theory assuming abstract phonological representations does then not need to be incompatible with a usage-based model. Instead, a hybrid model is needed which works with a layer of abstract representations complemented with detailed phonetic information signalling indexical properties.

5.2 Acquiring Sociophonetic Information

A usage-based approach to the storage and processing of speech has implications for models of language learning, both as an L1 and as an L2. As Foulkes and Docherty (Reference Foulkes and Docherty2006) point out, ‘the interweaving of sociophonetic and linguistic information in speech is so complete that no natural human utterance can offer linguistic information without simultaneously indexing one or more social factor’ (419). Therefore, an acquisition or learning model in which both types of information are processed and stored together is desirable. In the next sections, we review research on the acquisition of socio-indexical features in an L1 by children acquiring their native language and by adults moving to a new dialect region, before turning to the question of how L2 learners gain knowledge of sociophonetic information in an L2 and acquire L2 sociophonetic features in perception and production.

5.2.4 Acquisition of Second Language Accent Variation

Besides gaining metalinguistic knowledge of accent variation in an L2, necessary for accent recognition, L2 learners also need to acquire sociophonetic features in both perception and production. Research has shown that the learning process may be impacted by the acquisition of accent variation in an L2. One type of perceptual issue mentioned by Thomas (Reference Thomas, Chambers and Schilling-Estes2013) that can be addressed in sociophonetic studies is the question of ‘how dialectal differences affect the phonological categorization of phones (Janson Reference Janson1983), as the boundaries may differ by dialect’ (121). If this question is relevant for L1 users, it certainly also pertains to L2 users who are exposed to different regional and social accents in the L2. The variation in the L2 may manifest itself in two different target accents, as in the study by Escudero and Boersma (Reference Escudero and Boersma2004), who looked at the differential vowel acquisition by L1 Spanish listeners of L2 Southern British English and Scottish English. However, multiple accents may also be present in the target language to which learners are exposed, as is, for instance, the case for many EFL learners in Europe, who are exposed to both British and American English accents.

While studies on L1 perception have convincingly shown that listeners are especially apt at dealing with regional variation in their native language (Bent & Baese-Berk, Reference Bent, Baese-Berk, Pardo, Nygaard, Remez and Pisoni2021; Dahan et al., Reference Dahan, Drucker and Scarborough2008), the extent to which L2 listeners deal with different accents in the L2 has been investigated much less. In a recent study (Simon, De Clercq & Degrave, Reference Simon, De Clercq and Degravein prep.), we examined the perception of two Belgian Dutch accents by L1 Dutch listeners and L1 French learners of Dutch in Belgium. The accents compared were the acrolectal Standard (Belgian) Dutch accent and the regional Antwerp accent. Since L2 listeners have fewer higher-order resources to rely on when exposed to unfamiliar accents and since their phonological representations may not be robust enough to be flexible, the prediction was that L2 learners would be less apt at dealing with regional variation in the L2 compared to L1 listeners. The results, however, revealed that the French listeners did not experience more difficulty in the categorization or transcription of regionally accented vowels than Standard Dutch vowels, though they were outperformed by native listeners in both tasks (see Figure 3).

Figure 3 Boxplots representing average scores per participant for the L1 group (left plot) and L2 group (right plot) on the categorization and transcription tasks

(Simon et al., Reference Simon, De Clercq and Degravein prep.).

In addition, we predicted that participants would perform better in the transcription task, in which they were exposed to longer stretches of speech, than in the categorization task. The prediction was based on the fact that longer stretches of speech contain more acoustic-phonetic information and thus opportunities for listeners to familiarize themselves with and adapt to the unfamiliar regional accent. However, the results of a mixed linear regression model did not reveal any task effects (Simon et al., Reference Simon, De Clercq and Degravein prep.). Our results suggest that regional accents are not necessarily more problematic for L2 listeners than standard, acrolectal accents. Rather, the type of acoustic cues signalling phonemic contrasts and the correspondences between L2 and L1 vowels seem to be more important in the listeners’ success at identifying vowels than the standard versus regional accentedness per se. Chappell and Kanwit (Reference Chappell and Kanwit2022) examine the role of L2 proficiency, study abroad experience and explicit instruction on L2 Spanish learners’ acquisition of a specific phonetic variable, namely coda /s/ reduction (e.g. esquina, ‘corner’, realized with [h] instead of [s]), which occurs in large parts of the Spanish-speaking world (e.g. Caribbean, Central and South America and southern Spain) and is a salient marker of regional and social identity (Chappell & Kanwit, Reference Chappell and Kanwit2022: 189). They observed that more proficient learners more often linked /s/ reduction to regional origin and social status than less advanced learners. Additionally, participants who had taken a phonetics course in which their attention was explicitly drawn to regional variation in /s/ reduction, and participants who had spent time studying Spanish abroad in a community in which /s/ reduction is produced more often identified this feature as typical for Caribbean Spanish.

Solon and Kanwit (Reference Solon and Kanwit2022) similarly examined the acquisition of Spanish intervocalic /d/ deletion, in words like ‘cantado’ (‘sung’), realized as [kantao] (cf. also Solon, Linford & Geeslin, Reference Solon, Linford and Geeslin2018). The form with deletion is widespread in the Spanish-speaking world, though deletion rates reported for different varieties range from 10 per cent to almost 40 per cent (Solon & Kanwit, Reference Solon and Kanwit2022: 810). Particularly male speakers and individuals of lower socio-economic status have been reported to produce the variant with deletion. American English learners of Spanish performed a contextualized preference task in which they heard a male and female speaker of Colombian Spanish produce sentences with the target word produced either with /d/ deletion or without. Again, the authors observed that the more proficient learners more often preferred the forms with /d/ deletion, suggesting that they had developed an awareness of this feature as being linked to certain regional varieties of Spanish. In addition, the more advanced learners also started to become sensitive to other variables constraining the production of /d/ deletion, including amongst others speaker gender, in the sense that, just like native speakers, they associate /d/ deletion more often with male than female speakers.

5.3 Concluding Remarks

In this section, we discussed whether research in sociophonetics can further our understanding of processes of phonological acquisition, and specifically, whether it can provide insight into the nature and acquisition of phonological representations in an L2. Traditionally, sociophonetic information was not considered relevant for theory building in phonology. More recently, however, awareness has grown that social and linguistic information are intertwined and invariably present in natural speech and are hence acquired from the very beginning of the acquisition process. This has led to the idea that any acquisition or learning model in which both types of information (linguistic and social) are processed and stored together is desirable. One such model that has been proposed is a hybrid exemplar-based model in which a layer of stable phonological knowledge, based on abstract phonological representations, is combined with a layer of fine (socio)phonetic detail that is updated as new exemplars are encountered in the acoustic input.

Given that previous studies on the acquisition of sociophonetic features have mostly been conducted on children acquiring their native language and on mobile speakers acquiring a new accent in their L1, the field of L2 sociophonetics is still relatively small. This is surprising, given that the acquisition of socio-indexical information in an L2 is part and parcel of the acquisition process, as learners need sensitivity to and knowledge of sociophonetic variables to navigate different kinds of social situations and interactions. Recent studies by Chappell and Kanwit (Reference Chappell and Kanwit2022) and Solon and Kanwit (Reference Solon and Kanwit2022) have shown that L2 learners can pick up socio-indexical features in an L2, but that the extent to which they do depends on linguistic experience and L2 proficiency. While recency and frequency of exposure play an important role in L2 sociophonetic acquisition and ‘full mastery’ is said to be unattainable for L2 learners in the absence of these, it should be noted that the same may hold for L1 speaker-listeners. Native speakers of Dutch in the Netherlands, for instance, will lack sociophonetic knowledge about Belgian Dutch regional and social accents, unless they have received sufficient exposure to them. Besides amount of exposure, and as illustrated by Clopper and Bradlow (Reference Clopper and Bradlow2009), the L1 exerts an influence on the perception of L2 sociophonetic information, as L2 listeners were shown to sometimes attend to cues which are relevant in their L2, but not in the L1 and vice versa ignored relevant L2 cues which play no socio-indexical role in their L1.

The extent to which sociophonetic variability affects intelligibility is another issue that has received little attention in research so far. From a number of recent studies, it emerges that regionally accented speech may not have a negative impact on the actual intelligibility, though more studies are needed to corroborate this finding (cf. Simon et al., Reference Simon, De Clercq and Degravein prep.). It would, however, be in line with earlier and older studies that have investigated the intelligibility of non-native accented speech by L1 listeners. In an oft-cited study on native listeners’ judgements and understanding of Mandarin Chinese speakers’ L2 English speech, Munro and Derwing (Reference Munro and Derwing1995a) observed no strong correlation between accentedness (level of ‘foreign accent’) and intelligibility, in that even speakers who were judged to have a strong accent were highly intelligible (see also Munro & Derwing, Reference Munro and Derwing2020, where they provide retrospective commentary on their Reference Munro and Derwing1995a paper). Of course, the listeners in that study were listening to their L1, which – as noted before – is crucially different from listening in an L2.

In the previous sections, we have discussed how L2 learners have to build robust phonological representations (Section 3), while at the same time being flexible enough to deal with the large amount of phonetic variability present in speech (Section 4). We have seen that building this type of robust but flexible representations is a difficult task for L2 learners, who have already acquired a set of phonological representations for their L1. In Section 6, we take a look at the population of atypically developing language users, who are reported to differ from typically developing speakers in the formation of L1 phonological categories (Section 6.1). We then explore whether observations made about L1 phonological representations in clinical linguistics studies may also apply to (typically developing) adult L2 learners (Section 6.2).

6.1 Phonological Representations in Atypically Developing Language Users

The nature of phonological representations is also a research topic in clinical linguistics. Children or adults with atypical development in communication and/or reading and writing are reported to perform poorly on phonological awareness (PA) tasks such as phoneme deletion tasks (e.g. ‘drop the first sound of the word black’), word learning tasks or non-word repetition tasks, all of which require language users to parse words into distinct and robust phonological units (Bishop & Snowling, Reference Bishop and Snowling2004; Swan & Goswami, Reference Swan and Goswami1997). The low performance on PA tasks is often linked to a central phonological impairment (though see Pan & Chen, Reference Pan and Chen2005 for evidence on the lack of universality of the role of PA), which would equally affect lexical retrieval (slower for impaired readers) and phonological memory (problems with perceiving or encoding speech sounds for temporary storage) (cf. Long et al., Reference Long, Fox and Jacewicz2016). However, there does not seem to be consensus on the nature of the impairment – that is, where exactly do atypically developing language users differ from typically developing ones? A review of the literature on the topic reveals three main points of view.

First, many scholars have argued that the impairment is due to ‘fuzzy’ or ‘low-quality’ representations, meaning phonological representations which show overlap between phonemic categories (Claessen & Leitão, Reference Claessen and Leitão2012). Such overlap would make it difficult for language users to establish robust grapheme-to-phoneme correspondences, inhibiting reading ability. Similarly, Maillart, Schelstraete and Hupet (Reference Maillart, Schelstraete and Hupet2004) argue for ‘underspecified’ representations in children with atypical development. They conducted a French lexical decision task with seventy-five French-speaking children, of whom twenty-five were diagnosed with a developmental speech disorder, while the remaining fifty showed typical language development. The children were aged four to seven and matched for lexical age, based on their performance on a receptive vocabulary test. The results showed that both groups of children performed equally high on real words (i.e. they accepted them with more than 90 per cent accuracy), but that the atypically developing children had difficulty rejecting pseudo words which differed from real words through slight modifications, specifically through the addition or deletion of a segment in initial or final position. When larger modifications were made, affecting the number of syllables in the word, the atypically developing children were well able to reject the non-words. In addition, children’s performance increased with receptive vocabulary level. The authors take the results to suggest that atypically developing children have more holistic, underspecified phonological representations, making them capable of detecting larger modifications, but not more subtle, segmental changes. The results further point towards a gradual refinement of phonological representations, as the size of the (receptive) lexicon increases, thereby confirming earlier studies on the relation between vocabulary growth and the development of phonological representations.

A second point of view, however, is advocated in a more recent study by Li et al. (Reference Li, Braze, Kukona, Johns, Tabor, van Dyke, Mencl, Shankweiler, Pugh and Magnuson2019). They set up an eye-tracking study with native speakers of English with a wide range of reading-related skills, as assessed through an extensive test battery. The participants were presented with unambiguous words (e.g. ‘neck’ [nɛk]) and words that had misleading coarticulation on the vowel and can be considered ‘phonological competitors’ (e.g. [nɛ_kt], cross-spliced from ‘neck’ and ‘net’). The results revealed a negative correlation between phonological skills, as assessed through reading-related tasks (amongst others a non-word repetition and phoneme deletion task), and mismatch negativity responses; in other words, individuals with lower phonological skills were more sensitive to sub-phonemic deviations compared to those with higher phonological skills. If these individuals’ phonological representations had been underspecified, they would – quite to the contrary – be less distracted by sub-phonemic details. Therefore, the authors argue that the phonological representations in adults with low reading skills are over- rather than underspecified. This does not mean that typically developing listeners do not use allophones in perception. Mitterer, Reinisch and McQueen (Reference Mitterer, Reinisch and McQueen2018) report on a series of three experiments on selective adaptation. Selective adaptation refers to the phenomenon where prolonged exposure to a specific linguistic stimulus (like a particular sound or word) causes a temporary change in the perception of similar stimuli (Eimas & Corbit, Reference Eimas and Corbit1973). They found that listeners only displayed selective adaptation if adaptors and test stimuli shared allophones. If they only shared context-independent phonemes (e.g. two different allophones of Dutch /l/), no selective adaptation was observed. The difference between typically and atypically developing listeners may thus lie in the type of sub-phonemic detail they attend to and in its relevance in the specific language they are listening to.

Finally, some scholars have argued that the phonological impairment may, in fact, be due to impaired access to phonological representations rather than to a deviant degree of specificity in the phonological representations themselves. In other words, the impairment would result from a processing issue rather than from a problem at the representational level. This argument is proposed by Long et al. (Reference Long, Fox and Jacewicz2016) in a study on children and adults with dyslexia, which the authors point out is part of a phonological impairment, rather than a cognitive deficit related to decoding written language. Participants performed a randomized forced-choice identification task with speech samples from two different areas, Ohio (the participants’ residence) and North Carolina. The results revealed that, overall, children were outperformed by adults, but also and more crucially, that individuals with dyslexia performed significantly more poorly compared to participants with average reading abilities. In other words, both children and adults with dyslexia had not acquired socio-indexical features related to different dialects to the same extent. They also did not show a response bias for their own dialect: in the case of ambiguous stimuli, the dyslectic children and adults did not favour their own Ohio dialect as the response option, in contrast to the participants with average reading abilities. These findings are in line with earlier studies demonstrating that individuals with dyslexia are deficient in the extent to which they can adapt to multiple, unfamiliar talkers in their native language (Perrachione et al., Reference Perrachione, Del Tufo and Gabrieli2011). Since the dyslectic adults in Long et al.’s (Reference Long, Fox and Jacewicz2016) study were, however, able to categorize talker dialects, the authors surmise that the phonological representations are actually intact (rather than under- or overspecified), but that access to these representations is impeded as the result of a less efficient working memory, which would also be responsible for difficulties with talker normalization. To date, research on the acquisition of socio-indexical speech features by children with atypical development seems relatively limited. An unpublished thesis by Laase (Reference Laase2015), however, seems to point in the same direction. This study examined whether children diagnosed with Williams syndrome, a genetic disorder that causes learning and cognitive disabilities, had acquired salient acoustic (vowel) properties of their native regional dialect of American English. The results revealed that participants in the clinical group acquired only some features of their regional variety in production: the children’s vowel spaces were relatively idiosyncratic and contained only a selection of the typical regional accent features, compared to the typically developing children. The acquisition of socio-indexical language features thus appeared to be impeded in children with Williams syndrome.

6.2 Fuzzy Representations in Second Language Learners?

Although the causes of decreased PA in individuals with atypical language development are far from well understood, a question that arises is whether the low-quality or fuzzy representations argued by some to occur in atypical development may also emerge in L2 learners with typical development. This idea is somewhat appealing, as L2 learners’ problems with perceiving and producing L2 contrasts that do not occur in the L1 may be due to either the lack of a phonemic split or to phonological representations whose boundaries are blurred and lack robustness. In 2022, the notion of fuzzy phonological representations in L2 learning was the topic of a special issue in Bilingualism: Language and Cognition, with a keynote by Bordag, Gor and Opitz (Reference Bordag, Gor and Opitz2022), seventeen commentaries by peers and an authors’ response. In the same year (Reference Bordag, Gor and Opitz2022), Frontiers in Communication invited contributions to the research topic ‘Fuzzy lexical representations in the non-native lexicon’, leading to the online publication of one editorial and twelve research papers. While a comprehensive discussion of this whole body of research would obviously lead us too far, we briefly outline the main points here.

Bordag et al. (Reference Bordag, Gor and Opitz2022) introduce the Ontogenesis Model of L2 Lexical Representation (OM), which they argue is best described as ‘a model of the non-native lexicon’, which captures properties of L2 lexical units and seeks to map the development of L2 representations. A key idea of the model is that sometime along their developmental history (or ‘ontogeny’) representations may reach their ‘optimum’, meaning that they are ‘properly encoded and fully specified’ (185). The argument is that while L1 representations typically reach this optimum, lexical representations in the L2 typically do not reach this state but instead remain fuzzy, and move dynamically along the continuum from ‘weak’ to ‘robust’, sometimes coming closer to the optimum, sometimes moving further away from it, for instance during periods of little L2 exposure or use. The concept of ‘fuzziness’ is hence central to the model and defined as ‘inexact or ambiguous encoding of different components or dimensions of the lexical representations that can be caused by several linguistic, cognitive, and learning-induced factors’ (186).

In the OM, a lexical representation includes semantic, phonological and orthographic knowledge. In fully specified lexical representations, these three domains are well interconnected, facilitating lexical retrieval. In L2 representations, the connection may be looser. Each of the dimensions may start developing at a different point in time and may follow a different path in the acquisition process, with some reaching the optimum faster than others, which may be slower or never reach ultimate attainment.

In the large number of commentaries to this proposal, strong points as well as weaknesses of the model are discussed. Summarizing, the strength of the model lies in its attempt to provide a unified account of different dimensions of lexical representations, including phonological, orthographic and semantic properties, and that it includes a model for the development of these dimensions for L2 learners. However, the model faces a number of important challenges, identified in several commentaries. One issue is that the notions on which it crucially relies need to be more properly defined. What is the nature of fuzziness, exactly? Can it develop over time and does it lie in one feature, a whole syllable, or is it something more global (Darcy, Reference Darcy.2022)? What exactly is the ‘optimum’ that representations need to reach (Ellis, Reference Ellis2022; Lemhöfer, Reference Lemhöfer2022)? Additionally, and critically, the model at this point fails to formulate testable predictions. Escudero and Hayes-Harb (Reference Escudero and Hayes-Harb2022) argue that for the model to have explanatory power it needs to incorporate a model of the initial state – that is, what do the learners’ representations look like at the start of L2 acquisition, what is the role of perception in mapping acoustic, orthographic and semantic input onto representations, and which stages are predicted in the development of representations towards the optimum? Nicol (Reference Nicol2022), also in a peer commentary on Bordag et al.’s (Reference Bordag, Gor and Opitz2022) paper, further claims that the problem for advanced L2 learners may lie in the processing rather than in the representations. She (Reference Bordag, Gor and Opitz2022) argues that ‘[i]n order to convincingly argue in favor of fuzzy form representations in the L2 mental lexicon, it is essential to present converging evidence not just from comprehension tasks – where WM [working memory, our note] limitations may play a role – but also from production tasks’ (231). The assumption here is that, if learners’ productions deviate from the target forms in the same direction as predicted by perception data, the L2 forms differ from the L1 ones as the result of representational differences.

6.3 Concluding Remarks

In sum, research on atypically developing children and adults diagnosed with low reading abilities points in the direction of a general phonological impairment, which manifests itself as poorer performance on phonological awareness tasks, lexical retrieval tasks and tasks involving phonological memory. While much earlier research seemed to lay the cause of this impairment with ‘fuzzy’ or underspecified representations, other studies point towards a heightened sensitivity to sub-phonemic (and phonologically hence irrelevant) variation resulting from overspecified representations. However, research on the acquisition of socio-indexical features, such as regional dialect properties, suggests that individuals with low reading abilities may in fact have intact representations (as they are able to recognize regional accents), but have problems with processing, leading to a slower and less accurate recognition. This hypothesis is in line with earlier studies which have shown that atypical readers show poorer performance on talker normalization tasks – that is, have difficulty recognizing multiple talkers, presumably due to the higher processing costs involved in speech processing.

In L2 phonology, the idea of fuzzy representations has been proposed in the OM by Bordag et al. (Reference Bordag, Gor and Opitz2022), which aims to provide a unified account for how representations include different dimensions, including phonological, orthographic and semantic information. However, although the idea is appealing, it is clear that the concepts need to be defined in more detail and a more elaborate model of the initial state and developmental paths needs to be proposed, in order to formulate testable predictions. In addition, and in line with the literature on atypical development in child L1 learners, it remains difficult to tease apart processing difficulties from representational issues. Production studies complementing perception experiments may be needed to disentangle the two (Lemhöfer, Reference Lemhöfer2022).

7.1 Perception Studies as a Window into the Mental Lexicon

In Sections 3 to 5, we have discussed which aspects and properties, both linguistic and socio-indexical, need to be acquired by L2 learners and how these features may develop over time as a result of continued exposure to (variable) input and growing proficiency in the target language. Assuming that at some level of a language’s phonology there is a layer of abstract phonological representations or categories (see Section 2.2 for Baese-Berk et al.’s Reference Baese-Berk, Chandrasekaran and Roark2022 definition), these representations gradually need to become more robust, with well-defined boundaries. At the same time, they need to be flexible and capable of shifting to accommodate for the variable acoustic input. In addition, sociophonetic knowledge in the form of socio-indexical features need to be attached to these categories in a surface layer containing fine phonetic detail. A crucial question in the field of L2 phonology is then how it can be established whether learning has taken place, and more specifically, whether categories have been formed. This general question can be split up into three more specific ones:

1. 1) How do we know whether robust representations have been formed in the L2?

2. 2) How do we know whether flexible representations have been formed in the L2?

3. 3) How do we know whether socio-indexical features of the L2 have been acquired?

To address question (1) about the formation of robust categories in L2 learners, researchers have typically used perception studies to gain insight into the emergence of phonological categories in L2 learners, as these are thought to offer a better window into learners’ grammar than production studies, in which the output may reflect articulatory constraints rather than internal structure or representations (e.g. articulatory difficulty to produce interdental friction in English /θ/). Often used tasks in experimental perception studies include identification and discrimination tasks. In identification tasks, listeners are presented with auditory stimuli and are asked to identify the target sounds by clicking on one of several visual representations on the screen. For instance, they hear the English word ‘but’ [bʌt] and are presented with a selection of orthographic forms, including, for instance, ‘bet’, ‘bat’, ‘bit’, ‘boot’ and ‘but’, from which they are asked to pick one. The results of L2 identification tasks provide us with information on how L2 listeners map the acoustic input of target sounds onto L2 phonemic categories and can hence provide insight into the presence or absence of L2 representations. A myriad of studies on different languages have used L2 identification tasks to get insight into L2 learners’ development (see Nagle and Baese-Berk, Reference Nagle and Baese-Berk2022 for a discussion of this task). In addition, studies have used cross-language categorization tasks in which listeners are asked to map target L2 sounds onto L1 categories on the screen. The aim of these tasks is to get a better understanding of the perceptual mappings and correspondences between the L1 and the L2. Lexical decision tasks, in which listeners are presented with an auditory stimulus and have to judge whether the stimulus is a real word (e.g. ‘cave’) or a non-word (e.g. ‘tave’), are similar to identification tasks, in that learners need to access their knowledge about L2 phonological categories in order to reach a decision.

Discrimination tasks reveal whether listeners are able to perceive differences and similarities between L2 sounds. One commonly used discrimination task is the AXB task, in which listeners hear a sequence of three auditory stimuli and are asked to indicate whether sound ‘X’ is most similar to sound ‘A’ or ‘B’. (Variants have been proposed, such as the XAB and the ABX tasks, in which the target stimulus comes at, respectively, the very beginning or the very end). Oddity tasks also exist in different forms. In a four-interval-oddity task, listeners hear four sounds in a row and are asked to press a button to indicate which one of these is deviant or, in other words, ‘sounds different from the others’. Often, the auditory stimuli – in both types of tasks – are not acoustically identical; instead they are different realizations by the same speaker or tokens produced by different speakers. When this is the case, and there is a sufficiently long interstimulus interval (1,000–2,000 milliseconds), these tasks are argued to tap phonological knowledge, as learners are not able to just rely on a comparison between sounds at the auditory level (Mora, Reference Mora2007).

The second question is how it can be established whether learners have built representations which are flexible enough to cope with phonetic variation in the acoustic input. In Section 4.2, we reviewed previous studies on perceptual adaptation to accented speech. Often, these studies make use of lexical decision tasks, preceded by a familiarization phase, in which listeners are exposed to unfamiliar phonetic realizations or unfamiliar boundaries. A typical example of the latter would be to familiarize listeners with a VOT contrast that doesn’t exist in their L1, such as presenting a three-way VOT contrast (e.g. Thai) to listeners of English, which has only a two-way contrast between short-lag and long-lag aspirated stops. These studies provide us with information on learners’ ability to shift phonemic boundaries after exposure to accented speech. Insightful as these studies are, they do not provide us with information on the flexibility of phonological representations of L2 listeners when they are exposed to, for instance, accented speech. The question remains unanswered whether in those cases, listeners need to adapt their perception or whether their category boundaries need to be flexible enough to deal with this kind of variation from the start.

While perceptual adaptation studies are grounded in the psycholinguistic research tradition, the perception of accented speech is often couched in intelligibility research, which has a history in research on L2 acquisition. With the shifting focus in L2 pronunciation teaching – away from native speaker models and towards intelligibility as a more feasible or desirable pronunciation goal – intelligibility research has been on the rise in the past decades. While early studies on intelligibility (see Section 4.5) were mostly concerned with the extent to which L2 speakers’ pronunciation was intelligible to native listeners, more recently, intelligibility research has moved away from being centered around native speaker-listener models. The idea that intelligibility depends not only on the speakers and their speech, but is equally determined by listeners has gained ground. For instance, the listener’s L1, their familiarity with the speaker’s language or accent and the extent of contact the listener has with the language will all play a role in the extent to which a speaker will be comprehensible or intelligible (Simon et al., Reference Simon, Lybaert and Plevoets2022). A study by Winke, Gass and Myford (Reference Winke, Gass and Myford2013) showed that language test-raters who had studied Spanish or Chinese as a foreign language were significantly more lenient in their comprehensibility ratings of, respectively, Spanish- and Chinese-accented English speech samples (see also, e.g., Lippi-Green, Reference Lippi-Green1994 on the impact of listeners’ goodwill in communicative success). A number of studies have examined to what extent regional and/or non-native speech is intelligible to L2 learners (see Section 5.2.3). Typically, such intelligibility tasks involve transcription of sentences (sometimes nonsensical, to reduce predictability effects based on semantic context) or individual words (e.g. fill-in-the-blank task). These tasks provide evidence for how well L2 listeners deal with non-native speech in non-optimal conditions (see, e.g., Verbeke & Simon, Reference Verbeke and Simon2023a, on the intelligibility of different native and non-native varieties of English). As recently discussed by Baese-Berk et al. (Reference Baese-Berk, Levi and Van Engen2023), however, it is unclear to what extent intelligibility research can actually help us assess whether perceptual acquisition has taken place. In fact, the authors argue that intelligibility data can typically not provide information on the causes of listening challenges and hence tell us little about the L2 learners’ phonological representations. In order to take that step, different tasks are needed in which the stimuli are controlled and which test the intelligibility of words differing in specific alternating target segments.

The third question just formulated is how we can assess whether learners have acquired socio-indexical features in the L2. It is in this respect useful to consider the distinction made in the literature between conscious knowledge of socio-indexical features and subconscious perception of those features (Schmidt, Reference Schmidt1990). Montgomery and Moore (Reference Montgomery and Moore2018) investigated the former in L1 English listeners, using a self-developed Salient Language In Context (SLIC) tool that allows participants to listen to speech samples in different accents while clicking a button whenever they notice a social feature that is particularly salient to them. This behavioural procedure, in which listeners can also review their clicks and explain why certain features were notable, explicitly addresses L1 listeners’ ability to notice sociophonetic features in different accents. As noted in Section 5.2.3, to date few studies have investigated the acquisition of sociophonetic features by L2 learners. In a recent set of studies Dailey (Reference Dailey2024) explored L1 and L2 French listeners’ perception of sociophonetic cues to formal and informal registers in Metropolitan French. Specifically, the studies focus on optional liaison in phrases like ‘plats [z] italiens’ (‘Italian dishes’), associated with formal register, and post-obstruent liquid deletion in words like ‘table’ [tab] (‘table’), typical for informal, everyday speech. Using a matched-guise technique, Dailey found that, unlike the L1 listeners, the L2 listeners only perceived speech containing liaison as more or less formal when they explicitly noticed the sociophonetic feature. Liquid deletion, by contrast, was associated with informal, everyday speech by L1 and L2 listeners even when they did not consciously notice the feature. Dailey (Reference Dailey2024) argues that the liquid deletion results suggest that the L2 listeners were able to acquire the social meaning of this feature, possibly helped by the fact that liquid deletion is a reduction phenomenon, which is cross-linguistically typical of casual speech. In order to explain the absence of a liaison effect and the minimal noticing of optional liaison displayed by the L2 listeners, Dailey refers to the stereotypical character of optional liaison in French. The L1 French listeners must have picked up the feature as part of being raised in a French educational and cultural context, in which this feature is associated with formal speech produced by prestigious speakers with upper-class professions. Even the highly advanced L2 listeners remained unfamiliar with this stereotype and did not make social inferences from the feature.

To conclude, the development of robust phonological representations in L2 learners can be investigated through perception tasks, such as identification, categorization or discrimination tasks, which provide insight into the phonemic or phonetic categories formed by the learners. Whether L2 learners’ phonological representations are flexible enough to deal with phonetic variability due to, for instance, regional accents, is typically examined though intelligibility tasks, such as transcription tasks involving accented speech fragments, or in lexical decision tasks, in which a familiarization phase is followed by a test phase. Studies examining the acquisition of socio-indexical features by L2 learners are not abundant yet, but use metalinguistic noticing tasks or matched-guise techniques. In a hybrid model of phonological representations, in which sociophonetic features are associated with abstract phonemic categories, the acquisition of social meaning is part of the phonological acquisition process. If L2 listeners display associations of particular phonetic cues with social features, such as register or regional background, this means that they have acquired at least part of the phonological representation.

7.2 The Role of Production Studies

So far, all the evidence we have cited for category formation comes from perception studies, as these are generally assumed to offer a more direct window onto learners’ mental representations compared to production studies. The reason is that in production tasks, articulatory constraints may lead to discrepancies between the speech sounds speakers aim to produce and their actual pronunciations. A typical example in L2 English research is the production of (inter)dental fricatives (e.g. onset consonants in ‘think’ /θ/ and ‘they’ /ð/), which are known to pose articulatory challenges to L2 learners. Even when L2 learners have developed robust mental representations for these speech sounds and strive to pronounce the sounds in line with these representations, they may fail to produce appropriate friction at the dental place of articulation, if the combination of these articulatory gestures is unfamiliar to them from their L1 (as would be the case in, e.g., Dutch, German and French). Perception studies, by contrast, can provide us directly with information on which cues are detected by the listener in the acoustic input and how these cues are mapped onto phonological representations (Archibald, Reference Archibald2023).

However, this does not mean that production data cannot reveal anything about the acquisition of phonological representations. A recent study by Turner (Reference Turner2024), for instance, examines the effect of a residence abroad in a French-speaking country on the acquisition of the French sounds /u/ (as in ‘vous’, you) and /y/ (as in ‘vue’, view) by L1 speakers of English, which lacks this vowel contrast. In the study, categorical changes in production are taken to reflect phonological acquisition, while gradient changes reflect changes in phonetic properties of the sounds in question. Specifically, if the learners increasingly produce what for them is a /u/ vowel in French /u/ items and an /y/ vowel in French /y/ items, this is taken as evidence for phonological acquisition. By contrast, gradient, phonetic acquisition is said to take place if the learners’ phonologically accurate /u/ and /y/ vowels become acoustically more target-like during the residence abroad (Turner, Reference Turner2024). The results of the production study revealed a significant phonetic development in French /y/, as well as a developing phonological representation for French /u/ over the residence abroad. Different input factors, such as quantity and quality of input over the stay abroad, amount of previous naturalistic exposure and length of time learning French, had a differential impact on L2 development. For instance, the quantity and quality of input during the residence abroad more strongly determined phonological compared to phonetic development.

In addition, many studies have addressed the relationship between L2 speech perception and L2 speech production (see Nagle & Baese-Berk, Reference Nagle and Baese-Berk2022 for a comprehensive overview and discussion). If a tight link between the two modalities is assumed, this implies that L2 production would develop hand in hand with L2 perception, meaning that both perception and production studies would be able to provide us with evidence of category formation. However, while it is clear that there is a link between perception and production in L2 learners, the nature of this link is still in need of further investigation. Previous studies have, for instance, demonstrated that L2 learners may be able to produce L2 sounds in a target-like manner without being able to reliably distinguish between them in perception. An often-cited study is the research by Sheldon and Strange (Reference Sheldon and Strange1982), who demonstrated that some L1 Japanese learners of English were able to produce English /r/ and /l/ before being able to accurately perceive these speech sounds, at least in pre-vocalic consonant clusters. In addition, asking participants to produce words during L2 speech sound learning (i.e. learning at the sub-lexical level) has been shown to have a disruptive rather than facilitatory effect on perceptual learning (Baese-Berk, Kapnoula & Samuel, Reference Baese-Berk, Kapnoula and Samuel2024). These findings have led some scholars to propose separate developmental routes for perceptual and production representations. While a review of studies on the perception–production link would lead us too far afield, the general picture that emerges is that the two modalities may be more or less closely linked to one another at different levels of representation (cf. also Baese-Berk et al., Reference Baese-Berk, Kapnoula and Samuel2024).

In the following sections, we single out two types of production studies that might be helpful in exploring the extent to which production studies can provide us with information on the acquisition of mental representations: phonetic convergence studies and articulatory phonetics studies.

8.1 Methodological Opportunities

8.1.1 Spontaneous Speech and Second Language Phonology Research

There is a trend in speech research these days to advocate the use of spontaneous data, which would enhance the ecological validity of studies on the perception and production of speech sounds. Ecological validity is defined by Kihlstrom (Reference Kihlstrom2021: 466) as ‘the extent to which experimental findings can generalize to the “real world” situation that a researcher wishes to understand’ (cf. also Verbeke, Reference Verbeke, Buysschaert and Lefèvre2024 for a discussion on ecological validity in language and speech research). The question relevant to us in this Element is whether the use of spontaneous speech in L2 experiments will further our understanding of how L2 phonology is acquired. In other words, do experiments that use spontaneous speech have a good or even better potential of pushing the field of L2 phonology forward? In a recent Cambridge Element in Phonetics devoted entirely to the topic of spontaneous speech, Tucker and Mukai (Reference Tucker and Mukai2023) emphasize investigating spontaneous speech, pointing out that it is a central part of the wide range of speech types that speakers and listeners use: ‘If the goal is to understand communication from the perspective of ecological validity (i.e., a sample that reflects actual speaker and listener experience), then it is important to understand how this range of speech types impacts speech production and perception’ (2). The authors grant that it is difficult to determine the boundaries of what can be called spontaneous speech, because there are multiple dimensions to speech types: a careful-to-casual dimension, a fast-to-slow dimension and a reduced-to-unreduced dimension. As a working definition of spontaneous speech, they propose the following: ‘Spontaneous speech is speech produced by a speaker in an informal, dynamic, unrehearsed, casual manner’ (5).

As spontaneous speech is much more prevalent in everyday communication than formal, rehearsed and planned speech, it is clear that it should also be the object of investigation in research on L2 phonology. Studies in this field have often opted for the use of highly controlled laboratory speech, both for the study of L2 production and for L2 perception experiments. The advantages of this type of speech in these studies are obvious: the quality of the acoustic signal is high (background noise is absent) and – even more importantly – the produced speech can be controlled in several respects. Depending on the type of task, it is possible to monitor speech rate through instructions or by setting time limits, or to control the phonetic context of target segments. For instance, numerous L2 perception experiments on vowel acquisition have used stimuli in which vowels are included in fixed consonantal frames (such as ‘hVd’ and ‘bVt’) (e.g. Escudero et al., Reference Escudero, Simon and Mitterer2012). These studies are able to answer questions about how the vowels are perceived, without different consonantal contexts biasing the results. The disadvantage of such controlled stimuli lies, of course, precisely in their controlled nature: speakers do not normally produce vowels in these consonantal frames at the same speech rate, and listeners are not normally exposed to such forms. This does not need to be problematic, when the researcher’s goal is to gain insight into mechanisms underlying speech perception, or when the research question is of a theoretical nature and the main goal of the study is to test, for instance, predictions made by different speech perception models (see, e.g., Chang, Reference Chang, Schmid and Köpke2019 for a comparison of, e.g., PAM-L2 and SLM-r). If that is the case, there is great value in using more controlled laboratory data for perception experiments.

Using highly controlled speech may become problematic when the research question itself pertains to the acquisition of aspects of spontaneous speech in an L2. A well-known property of spontaneous speech is that it is subject to reduction processes which may be absent or at least less strongly present in formal, rehearsed speech. English, for instance, is well known for vowel reduction processes, causing words like ‘family’ to be pronounced as [fæm.li] or ‘suppose’ as [spəʊz], but such processes can be found in many languages. In a recent study, we aimed to answer the question to what extent phonetic reduction in several accents of English affects the perception and the intelligibility for L2 listeners with different native languages. Specifically, we investigated whether L1 English listeners and L2 listeners with either Dutch or Spanish as their native language differ in the extent to which reduced forms in different accents of English are intelligible to them. While vowel reduction is common in Dutch, it occurs to a far lesser extent in Spanish, which also lacks a schwa in its inventory. In order to assess the intelligibility of reduced and unreduced forms in context, participants performed a transcription task in which the target stimuli occurred in sentence-long utterances. These audio fragments were excised from longer broadcast interviews or talk shows, which were all publicly available online. In line with previous research (e.g. Ernestus & Baayen, Reference Ernestus, Baayen, Trouvain and Barry2007) and our hypothesis, reduced words were recognized less accurately compared to unreduced words, and L2 listeners had relatively more difficulty with reduced words than L1 listeners (see Figure 4). In contrast to what we had predicted, the L2 Spanish listeners did not have more trouble understanding reduced forms, suggesting that the presence of similar reduction processes in Dutch and English did not create an advantage for the Dutch listeners.

Figure 4 Effect display of two-way interaction between listener group and reduction condition

(Verbeke, Mitterer & Simon, Reference Verbeke, Mitterer and Simon2025).

Because the target words were embedded in a semantic context which might have facilitated recognition, we set up a follow-up experiment consisting of a lexical decision task with the same stimuli in which we measured participants’ reaction times. For that purpose, we needed to create non-words. In order for these non-words to sound as authentic as possible, we excised syllables from spontaneous words and merged them together. For instance, the non-word submantic (*/səbˈmæn.tɪk/) was created from submerge (/səbˈmɜːdʒ/) and romantic (/rəˈmæn.tɪk/). Results revealed that listeners needed more time to recognize reduced target words as real words compared to unreduced forms. Again, there was no difference between Spanish and Dutch listeners in terms of the effect of reduction: both groups were slowed down to a similar degree. The study provides insights into the role of phonetic reduction in the intelligibility of spontaneous speech by L2 listeners. Methodologically, it provides an example of a study that walks the line between using maximally spontaneous speech in perception experiments, and keeping control over the context and task. Although we decided to use authentic data from talk shows, the stimuli could not be randomly selected. In fact, they had to be carefully selected, so that we kept control over variables such as sentence duration, number of syllables, speaking rate, predictability of the target word within the sentence context (checked using a cloze task) and signal-to-noise ratios. As a result, the selection of these authentic stimuli was time-consuming.

One important implication of our decision to use authentic, spontaneous speech from talk shows is that it was impossible to include reduced and unreduced forms of the same words in the experiments, as these could simply not be found for the same speakers of the different accents. Arguably, it may be possible to collect such data in a laboratory setting, in which case speakers would be asked to produce sentences containing reduced and unreduced variants of the same word. In Xu’s (Reference Xu2010) forum piece in Journal of Phonetics, several myths about lab speech are debunked, such as that lab speech is slow and careful, clear and articulate. As Xu argues, participants in the lab seem to have little difficulty following instructions about, for instance, the level of formality they are to use, and so it may be possible to elicit relatively spontaneous sounding stimuli differing in presence or absence of reduction, though these may lack a certain authenticity (cf. also Wagner, Trouvain & Zimmerer, Reference Wagner, Trouvain and Zimmerer2015 on stylistic diversity in speech research).

Besides studies that specifically focus on the acquisition of properties of spontaneous speech, such as reduction, another type of L2 research for which spontaneous speech may be preferred over highly controlled speech is sociophonetic research, discussed in Section 3. In L2 sociophonetic studies, which, as we have argued, may contribute to our understanding of L2 phonological acquisition, we are interested in whether and how L2 learners acquire socio-indexical features of L2 accents. As the production of socio-indexical features is context-sensitive (e.g. more features in informal speech with members of the same community), the use of authentic, spontaneous speech in this field seems mandatory.

However, in a recent paper by van de Velde et al. (Reference van de Velde, Pinget, Voeten, Demolin, Kristiansen, Franco, de Pascale, Rosseel and Zhang2022), the authors argue that in sociolinguistic research spontaneous speech has recently lost some of its previously unassailable status. Traditionally, variationist linguists greatly valued the study of speech data gathered in authentic communicative contexts. Some downsides of using authentic speech are, however, that variables are unequally distributed in speech samples (referred to as the frequency problem), that specific combinations occur infrequently (the co-occurrence problem) and that groups of speakers show different patterns of variation (the interaction problem) (van de Velde & van Hout, Reference van de Velde and van Hout2000). Given that experimental designs do not run into the same problems, van de Velde et al. (Reference van de Velde, Pinget, Voeten, Demolin, Kristiansen, Franco, de Pascale, Rosseel and Zhang2022) advocate an approach which they call ‘laboratory sociolinguistics’, characterized as ‘a branch of variationist sociolinguistics that makes use of laboratory techniques and quantitative research methods’ (563). Recent studies examining the acquisition of sociophonetic variables in an L2 or in a second dialect using lab data produced with different degrees of spontaneity include Montgomery and Moore (Reference Montgomery and Moore2018), Vriesendorp (Reference Vriesendorp2021) and Dailey (Reference Dailey2024).

In sum, we argue that L2 speech research would benefit from the inclusion of spontaneous speech data, especially when the research questions revolve around the acquisition of aspects of spontaneous speech, such as reductions or assimilation processes, or in L2 sociophonetic studies.

We agree with Tucker and Mukai (Reference Tucker and Mukai2023: 30) that studying speech perception using spontaneous speech is challenging and therefore argue that spontaneous speech should complement rather than replace more traditional lab speech or lab perception experiments. We discussed the feasibility of incorporating natural data in L2 perception experiments, pointing out that a balance needs to be found between using maximally spontaneous speech and keeping control over data to make sure that research questions can be answered.

8.2 Theoretical Challenge: The Second Language as a Linguistic and Social Code

In any stretch of speech, linguistic and social information invariably occur together: whenever speech is produced, the listener can infer information about certain socio-indexical features of the speaker or the context (Foulkes & Docherty, Reference Foulkes and Docherty2006; see Section 5.1). With the growing awareness that socio-indexical features are always present in any stretch of speech has come the realization that the acquisition of these features by L2 learners is part and parcel of the L2 process. To comfortably handle a wide range of communicative contexts and social interactions in the L2, L2 learners need to be able to infer socio-indexical information about, for instance, regional or social background of the speaker. This observation presents us with the theoretical challenge of integrating the acquisition of socio-indexical information in models of L2 phonological acquisition. Chappell and Kanwit (Reference Chappell and Kanwit2022), discussing L2 sociophonetic variation, review the basic tenets of SLM, PAM-L2 and L2LP (cf. Section 2.2) and note that in none of those models is learner perception linked to social characteristics of the speakers (188). In other words, the models allow us to make testable predictions about how L2 learners will perceive L2 sounds on the basis of the relation to their L1 categories and how these learners may be able to acquire L2 categories through, for instance, boundary shifting, but they do not explain how learners come to link linguistic features in the L2 with social characteristics. If the L2 is viewed not just as a linguistic but also as a social code that learners need to master, then integrating social features into L2 speech learning models is vital. In this respect, it is worthwhile mentioning the Attrition & Drift in Access, Perception, and Production Theory (ADAPPT) recently proposed by De Leeuw and Chang (Reference De Leeuw, Chang and Amengual2024). This theory primarily focuses on the impact of additional languages on L1 speech, be it short-term, phonetic changes (‘drift’) or long-term, phonological ones (‘attrition’). In general, ADAPPT stipulates that there will always be impact of additional language exposure (L2, L3, Lx) and acquisition on the L1, though the extent of the impact and the dimensions involved (access, perception or production) may vary. One of the principles of ADAPPT, outlined by De Leeuw and Chang (Reference De Leeuw, Chang and Amengual2024), specifically points at the way in which the social meaning of particular variables in the L1 will play a role in the extent to which they will be open or resistant to influence from an L2. The authors advocate that future studies on L1 drift and attrition should look into the social and indexical meanings in speakers’ L1 and L2 in order to help explain which variables are affected. We believe that, similarly, the acquisition of L2 variables or features can be explained by their socio-indexical meaning and that theories or models focusing specifically on L2 acquisition should hence integrate sociolinguistic information.

The question then arises whether models on the acquisition of socio-lexical information can be integrated in existing L2 speech learning models, or whether new models need to be developed to cope with the development of L2 linguistic and social knowledge together. Chappel and Kanwit (Reference Chappell and Kanwit2022), in their study on the L2 perception of aspiration (/s/ being realized as debuccalized [h]) in certain Spanish accents argue for a union of the L2 perception model they are using (L2LP) with exemplar models of phonological representation and indexical meaning (see Section 2.2 for a discussion on exemplar-based models). They argue that this unified model would account for how L2 learners first modify their L1 boundaries to the perception of the phonetic variants in the L2, after which these new categories can become exemplars in the learners’ mental lexicon. What is needed for this last step to take place is sufficient exposure to the feature:

In this view, the formation of exemplars thus hinges on the perception of the L2 categories. This would imply that less proficient L2 learners, who have not yet developed robust phonological representations in the L2, would be unable to acquire socio-indexical knowledge. This is in line with earlier studies in which it was shown that the development of sociophonetic knowledge goes hand in hand with an increase in proficiency level in the L2, as discussed in Section 5.2.3.

However, more research is needed to examine how powerful (in the sense of being able to formulate verifiable predictions) such a unified model would be. Moreover, alternative models in which socio-indexical features are part of and are acquired hand in hand with phonological representations may need to be explored. Developing and testing predictions of L2 models that unify the acquisition of phonological representations and socio-indexical information presents a challenge for the future.