1 Introduction

Foreign language instructors and language students are commonly fascinated by the differences between the speech of native and non-native speakers of a language. They may have their intuitions, possibly stemming from their own experiences with teaching and learning foreign languages, about the challenges involved in perceiving and producing the sounds of a new language. The scientific study of the development of perception and production of sounds in a foreign language has a seventy-year-long tradition, and numerous theoretical models as well as empirical research have yielded important insights into various factors that can influence the learning process. In the first decades of research into phonological acquisition of foreign languages, the language that was learned was usually referred to as a second language (L2), which was used as an umbrella term for all foreign languages no matter how many and when they were learned. With the important insight that not only the native language (L1) but also the other languages a speaker knows can influence the phonological acquisition of a target language, a terminological distinction has been proposed between a learner’s first foreign language (L2) and their further foreign languages (L3). In this Element, we will refer to the chronologically first learned foreign language as the L2 and to languages that are learned chronologically later as the L3. L3 learners will also be referred to as multilingual learners.

The observation that a learner’s multiple languages interact has been widely documented in both speech perception and production. This phenomenon, referred to as cross-linguistic influence (CLI), has been shown in the acquisition of target sounds (e.g., Reference Amengual, Meredith, Panelli, Calhoun, Escudero, Tabain and WarrenAmengual et al., 2019; Reference Cabrelli and PichanCabrelli & Pichan, 2019; Reference Gabriel, Kupisch and SeoudyGabriel et al., 2016; Reference Kopečková, Pawlak and AroninKopečková, 2014, Reference Kopečková2016; Reference Kopečková, Marecka, Wrembel and GutKopečková et al., 2016; Reference Llama, Cardoso and CollinsLlama et al., 2010; Reference Llama and CardosoLlama & Cardoso, 2018; Reference Llama and López-MorelosLlama & López-Morelos, 2016; Reference Morales Reyes, Arechabaleta-Regulez and MontrulMorales Reyes et al., 2017; Reference PatiencePatience, 2018; Reference WrembelWrembel, 2014, Reference Wrembel, Gut, Fuchs and Wunder2015) and prosody (e.g., Reference Cabrelli and RothmanCabrelli & Rothman, 2010; Reference GutGut, 2010) as well as phonological processes (e.g., Reference CebrianCebrian, 2000; Reference SatoSato, 1984). Related findings have also revealed that such an influence is multidirectional in nature, i.e., that it concerns not only the influence of an earlier acquired language (whether L1 or L2) on the later acquired language but also vice versa (e.g., Reference Aoki, Nishihara, Cathcart, Chen and FinleyAoki & Nishihara, 2013; Reference BeckmannBeckmann, 2012; Reference CabrelliCabrelli, 2016; Reference Cabrelli and RothmanCabrelli & Rothman, 2010; Reference SypiańskaSypiańska, 2016, Reference Sypiańska2017). Research into L3 phonological learning has further drawn our attention to the possibility that more than one language can act as the source of this influence, i.e., a simultaneous (combined) influence of more than one language at the same time (see Reference De AngelisDe Angelis, 2007; Reference Wrembel, Gut, Fuchs and WunderWrembel, 2015), critically distinguishing the learning constellation of L2 learners from that of L3 learners. Multiple language knowledge thus entails an increase in the number of potential cross-linguistic interactions as well as in the number of potential directions.

Indeed, one of the main tenets of the L3 acquisition field is that learning a second foreign language is fundamentally different from learning a foreign language for the first time. It is not only the broadened phonetic repertoire and hence a potential for a range of cross-linguistic interactions but also the learner’s greater experience with language learning that qualitatively distinguish the learning situations. L3 learners have already learned their first foreign language to some extent and thus are likely to have discovered their individual learning strategies and style to further their own L3 learning success (Reference HufeisenHufeisen, 2010). Studies examining the effects of having learned a first foreign language on L3 speech perception and production have further shown that L3 learners tend to outperform L2 learners in target language phonetic discrimination and sound production, which has been explained by L3 learners’ raised level of metaphonological awareness and/or enhanced perceptual sensitivity (e.g., Reference Antoniou, Liang, Ettlinger and WongAntoniou et al., 2015; Reference KopečkováKopečková, 2016).

Yet one characteristic of non-native speech perception and production arguably common to L2 and L3 learners is the large amount of variability between learners (inter-learner variability) and within the same learner (intra-learner variability). The sources of such variability have been shown to be either linguistic factors, such as segmental and prosodic contexts in which the target occurs, or learner-related factors. The latter can include a myriad of learner-internal factors (e.g., age at onset of language learning, phonological awareness, motivation) as well as learner-external factors (e.g., quality and quantity of phonetic input, opportunities for language use). It has been shown that each of the learner-related factors is subject to change in the course of an individual’s phonological development, and therefore definite predictions about a learner’s pattern of development may be hard to make.

The goal of scientific explorations of L2/L3 phonological acquisition is to gain a systematic understanding of the processes and factors involved in the acquisition of a target sound system and to explain the perception and production outcomes observed for learners. This Element aims to contribute to this goal by presenting findings from a study that investigates the development of phonology in all of the languages of multilingual speakers across a ten-month period. It will do so based on a novel methodological approach that was developed within a relatively recent theoretical framework of foreign language development: Complex Dynamic Systems Theory (CDST) (Reference de Bot, Lowie and Verspoorde Bot et al., 2007; Reference Larsen-Freeman and CameronLarsen-Freeman & Cameron, 2008; Reference Verspoor, de Bot and LowieVerspoor et al., 2011).

In Sections 2 and 3, we review the state of the art with regard to L2 and L3 phonological acquisition. Section 2 presents and compares the currently leading theories of L2 speech learning. We will discuss the extent to which the L2 theories offer suitable frameworks with respect to modelling phonological learning by multilingual speakers, i.e., speakers who speak and learn several foreign languages at the same time. Section 3 explains that the traditional conceptualisations of phonological acquisition which underlie both previous research and the theories presented in Section 2 tend to imply that foreign language speech learning is a linear, incremental process with a stable end point. It then presents Complex Dynamic Systems Theory, which conceptualises language development as an emergent, self-organised process with high variability across learners. It goes on to show which new methodologies are used to investigate foreign language phonological development within this framework. Section 4 presents the methodology of our longitudinal study involving sixteen learners with L1 German, L2 English and L3 Polish. The results are presented in Section 5 and discussed in Section 6. The Element concludes with suggestions for possible directions in future research in the area of L2/L3 phonological acquisition.

2 Theories of L2 Phonological Acquisition

It is widely recognised that L2 speakers normally sound different from L1 speakers of the target language. Although an intuitive account of this could lead one to believe that this is primarily due to production difficulties that affect L2 learners’ performance, most theories of L2 speech learning assume that the learners’ difficulty to produce target language sounds is actually perceptual in nature, i.e., that it is directly linked to the learners’ ability to detect similarities and differences between the sounds of their L1 and the sounds of their L2. Languages differ in their contrastive phonemes (e.g., English and Polish have /w/ and /v/ while German only has /v/), but phonemes may also vary across languages in terms of their phonetic properties (e.g., German /v/ is produced differently from English and Polish /v/). Theories of L2 speech learning describe, predict and explain L2 difficulty based on their particular view regarding the phonological and/or phonetic properties of the learners’ L1 categories (i.e., long-term memory representation of sounds) in comparison to L2 sound categories.

Two perception-based models have inspired the majority of L2 perception and production studies conducted on a variety of first language and target language combinations, namely the Perceptual Assimilation Model (PAM: Reference Best and StrangeBest, 1995; and PAM-L2: Reference Best, Tyler, Bohn and MunroBest & Tyler, 2007) and the Speech Learning Model (SLM: Reference Flege and StrangeFlege, 1995, Reference Flege, Schiller and Meyer2003; and SLM-r: Reference Flege, Bohn and WaylandFlege & Bohn, 2021). While the former aims to explain the development of foreign speech perception at both the phonetic and the phonological level, the latter model aspires to address both the perception and the production development in an L2 at the phonetic level. Both models were developed to predict the influence of an L1 onto an L2 and have so far not been expanded to include potential influences of any additional languages on the learner’s perceptual mappings of the target sounds.

Best’s Perceptual Assimilation Model (PAM: Reference Best and StrangeBest, 1995) and its adaptation to L2 speech learning (PAM-L2: Reference Best, Tyler, Bohn and MunroBest & Tyler, 2007; Reference Tyler, Nyvad, Hejná, Højen, Jespersen and SørensenTyler, 2019) propose that the learner’s ability to perceive L2 sound contrasts depends on the degree to which L2 sounds are assimilated to the exemplars of L1 sound categories. It follows a direct realist or ecological approach to speech perception that argues that listeners detect speech information in terms of articulatory gestures such as the manner and place of articulation rather than on the basis of any mental representation of sounds (Reference Best and StrangeBest, 1995; Reference FowlerFowler, 1986). In other words, the L2 learner engages in perceptually assimilating the articulatory gestures of target sounds to the articulatory gestures of their L1 phonemes. A non-native sound can be heard as a good or a poor example of an L1 phoneme (categorised), as different from any particular L1 phoneme (uncategorised) or as a non-speech sound (non-assimilated). Accordingly, different pairwise assimilation types associated with different levels of L2 discrimination difficulty have been proposed in the model (for a description of all assimilation types and their predictions, see Reference Tyler and WaylandTyler, 2021).

For example, both sounds of an L2 contrast can be judged as members of a single L1 sound category. When both members of the non-native contrast are phonetically and phonologically perceived as good or poor members of a single L1 category (single-category assimilation), then the discrimination in the L2 will be very poor. An example for such a scenario for German learners of Polish might be Polish /w/ and /v/ (as in łata /wata/ ‘patch’ and wata /vata/ ‘cotton wool’), both of which may be perceived as members of the German category /v/ and therefore poorly discriminated at the beginning of L2 learning. However, if one sound of the contrast is perceived as a much poorer member of the L1 category than the other (category-goodness assimilation), then discrimination between these two non-native sounds may range from moderate to very good, depending on how dissimilar the two non-native sounds are from the L1 sound category (Reference Shinohara, Han, Hestvik, Calhoun, Escudero, Tabain and WarrenShinohara et al., 2019). The PAM-L2 predicts that, with increasing L2 experience, new L2 categories can be formed as a result of the learner’s growing L2 lexicon when noticing that there are minimally contrasting words in the target language that are communicatively relevant. As will be explained in Section 4, an interesting learning constellation arises in this connection for German learners of Polish as an L3 and English as an L2. For these experienced language learners, the /v-w/ contrast is shared in both of their foreign languages; if acquired in the L2 as a result of the category-goodness assimilation there is a potential for a facilitative phonetic mapping of the contrast from the learners’ L2 onto their L3 categories rather than non-facilitation stemming from their L1 (an interaction with L2 and L3 orthography has to be further considered in this specific case).

Another possible pattern for discrimination of L2 contrasts occurs when one member of the contrast is perceived as a member of an L1 category and one is perceived as uncategorisable (categorised-uncategorised assimilation). Such an L2 contrast should be discriminated well because it reflects a phonological distinction between an exemplar of a known phoneme and an unknown sound. However, this type can be discriminated poorly when the uncategorised sound is in close phonological space to the categorised sound (Reference Faris, Best and TylerFaris et al., 2016). A possible sound contrast that may fit this pattern for German learners of Polish is Polish /ε/ and /ɛ̃/ (as in step /stɛp/ ‘steppe’ and stęp /stɛ̃p/ ‘gait’) since the Polish nasal vowel is probably perceived as uncategorisable by many L1 German speakers.

The PAM-L2 uses these PAM assimilation types to predict the likelihood of category formation for L2 sounds when an L2 learner (rather than a naive listener/beginner L2 learner in the case of the PAM) is actively using the target language. Perceptual discrimination should thus improve when the assimilation type changes as a result of new category acquisition (e.g., a single-category assimilation becomes a category-goodness assimilation), the prerequisite being that the learner is exposed to input that preserves phonetic differences. Reference Yu and LowieTyler (2019) explains that in classroom settings, in contrast to an immersion environment with rich native speaker input, teachers who speak the target language as an L2 may fail to provide clear phonetic differences between L2 categories and thus inhibit the likelihood of their learners’ L2 category acquisition. The outcome is not necessarily to be linked to the teacher not being a native speaker though, as accented speech may also preserve the phonological distinction between all L2 phonemes and thus facilitate the learner’s formation of relevant L2 categories. Similarly, written language input, which constitutes probably the main difference between immersion and classroom L2 acquisition, may reinforce the perception of L2 categories as based on L1 grapheme-phoneme correspondences and thus hinder L2 category formation. Finally, Reference Yu and LowieTyler (2019) elaborates on the effect of prior foreign language exposure (i.e., target language exposure) on perception learning, which has a special relevance for the predictions of the PAM-L2. For instance, a learner with previous exposure to the L2 may have already established a category-goodness contrast as a single L1–L2 category and therefore their perceptual learning trajectory may be set before classroom exposure. This may happen, for example, when a learner previously learned to read the L2 and, in the absence of spoken input, applied their L1 phonology to a considerable amount of L2 words via orthography (Reference Tyler, Nyvad, Hejná, Højen, Jespersen and SørensenTyler, 2019, p. 617). The presence of previous language exposure can therefore hamper the prediction of category acquisition. In any case, it is conceivable that L2 phonemes included in the single-category assimilation type and the category-goodness assimilation type will be even more difficult to acquire in the classroom setting, whereas the acquisition of the uncategorised L2 phonemes in turn may be supported in such a learning environment, especially where L2 orthography is present and provides an unambiguous distinction.

As stated, neither the PAM nor the PAM-L2 make specific predictions regarding the development of speech perception by L3 learners, i.e., those with prior foreign language exposure to additional languages. In one of the rare studies that extended the predictions of the PAM-L2 to the context of multilingual phonological acquisition, Reference Wrembel, Marecka and KopečkováWrembel et al. (2019) found that multilinguals assimilate L3 vowel sounds to both L1 and L2 categories with a preference for the latter, at least in their population of L1 German, L2 English and L3 Polish learners. The study also showed that even beginner L3 learners formed new L3 categories, distinguishing between highly similar L3 sibilant pairs that would typically follow the single-category assimilation pattern. In other words, in terms of perceptual acquisition, beginner L3 learners were argued to behave similarly to advanced L2 learners in that they were able to discriminate sound contrasts predicted by the PAM-L2 to be challenging.

Unlike the PAM/PAM-L2, which assumes that perceivers extract speech information at a gestural/articulatory level, Flege’s Speech Learning Model (SLM: Reference Flege and StrangeFlege, 1995, Reference Flege, Burmeister, Piske and Rohde2002, Reference Flege, Schiller and Meyer2003) and its revised version (SLM-r: Reference Flege, Bohn and WaylandFlege & Bohn, 2021) assume that perceivers form a long-term mental representation for target sounds from acoustic-phonetic cues. According to the SLM/SLM-r, this is a learning mechanism that is involved in L1 speech learning and that remains intact throughout life. Yet monolingual-like performance is not expected on the part of L2 learners as their L1 and L2 categories are stored in the common phonological space, inevitably influencing each other (Reference Flege and StrangeFlege, 1995, Reference Flege, Burmeister, Piske and Rohde2002; referred to as ‘common phonetic space’ in Reference Flege, Bohn and WaylandFlege & Bohn, 2021). This regularly leads to category assimilation in which L1 and L2 sounds of the learner are perceptually equated; a target language category is blocked and a composite or merged L1–L2 category is then used to perceive and produce both the L2 and the L1 sounds, leading to bidirectional cross-linguistic influence. Consequently, the model predicts that learners who have a merged L1–L2 category will not display monolingual-like performance as they are likely to produce sounds differently from monolinguals of either language.

One of the main hypotheses of the SLM/SLM-r is that the greater the perceived dissimilarity between L1 and L2 sounds, the greater the likelihood that learners establish target-like categories and consequently perceive and produce the target sounds more accurately. However, similar L2 sounds may remain difficult to learn, as L2 learners might not be able to discern the subtle phonetic differences between the L2 and L1 sounds and thus they will readily map the target sounds onto L1 categories, perceiving and producing such L2 sounds in terms of those L1 categories. For example, German learners of Polish are likely to easily detect the phonetic differences between the rhotic sounds of their L1 and L2, the uvular fricative /ʁ/ and alveolar trill /r/, respectively. Although the Polish rhotic may be initially difficult to produce, the likelihood of L2 category formation for such a novel sound is predicted in the model to be high. In contrast, the subtle phonetic differences and complex acoustic cues applied in distinguishing between word-final voiced and voiceless obstruents may be a particular learning challenge for German learners of English because they can be expected to apply their voiceless L1 category (in this word-final position) in perceiving and producing the target obstruents. It is to be noted in this regard that the position of the target language categories within a word is highly relevant, i.e., whether an individual sound occurs in the initial, medial or final position. According to the model, it is at the level of position-sensitive allophones that L1 and L2 categories are compared and refined with experience.

Similarly to the PAM-L2, the SLM/SLM-r assumes a continuous, slow refinement of L2 learners’ perception of phonetic differences between their L1 and L2 as a function of the quality and quantity of phonetic input obtained in meaningful communication. For instance, Reference Flege and LiuFlege and Liu (2001) showed that, for two groups of L2 learners who resided in the target language country for the same amount of time (i.e., as measured by length of residence), it was only the students and not the age-matched learners in other occupations who showed improvement towards native-like performance. The authors argued that the student group was exposed to a substantially greater amount of native-like L2 input, which enhanced their identification of word-final consonants.

In the original SLM, the ability to form new phonetic categories for an L2 sound was further predicted to be mediated by the age (of onset) of L2 speech learning. Adult (late) L2 learners were predicted in the model to readily subsume L2 sounds into L1 categories thanks to the assimilative power of their well-established L1 sound categories, thus failing to form new categories for some L2 sounds. Child (early) learners, in contrast, were predicted to be less influenced by their evolving L1 sound system, and thereby more likely to form separate categories for L2 sounds and thus to learn the target language sounds to native-like levels. Considering the generality of this hypothesis and mainly the difficulty of disentangling the development of the L1 sound system from the overall state of neurocognitive development of child L2 learners, the revised version of the model (SLM-r) replaces the ‘age hypothesis’ with the ‘L1 precision hypothesis’. According to this hypothesis, the more precise the L1 category when L2 learning begins, the more likely the phonetic differences between L1 and L2 sounds will be discerned and a new L2 category formed. In this connection, Reference Flege, Bohn and WaylandFlege and Bohn (2021) point out that individual differences in L1 category precision apply to all age groups and are potentially linked to auditory acuity, early-stage (pre-categorical) auditory processing and auditory working memory. The authors call for future research to evaluate the role of L1 category precision in L2 speech development, also with respect to the question of whether individual differences in L1 category precision affect how much L2 input learners need in order to show consistent patterns of interlingual identification.

Also, the SLM-r no longer focusses primarily on examining highly proficient L2 learners but instead acknowledges that it is the early stages of L2 development that offer an insight into the process of L2 phonetic category formation because it is only then that cross-linguistic measures of dissimilarity can serve as a predictor of later-formed L2 categories. It is likely that, with increasing L2 experience, perceived phonetic dissimilarity between L1 and L2 sounds will grow.

Another major hypothesis of the SLM/SLM-r concerns the link between the development of L2 sound perception and production. In the original version of the model, accurate L2 sound perception was predicted to be a precondition for accurate L2 sound production. Based on the observation of inconsistencies in related L2 research, the SLM-r also revises this hypothesis and proposes that L2 segmental perception and production co-evolve without precedence, i.e., that there exists a strong bidirectional connection between the two modalities although ‘the correspondence between the two is never perfect, even in monolinguals’ (Reference Flege, Bohn and WaylandFlege & Bohn, 2021, p. 19).

Finally, the SLM-r makes a fundamental move in orientation from examining between-group differences towards analysing the learning trajectories of individual L2 learners. Because L1 speakers cannot be all assumed to have formed the very same L1 categories (e.g., Reference Hillenbrand, Getty, Clark and WheelerHillebrand et al., 1995), their L1–L2 mappings and consequently L2 speech learning trajectories and outcomes will vary (e.g., Reference Escudero and WilliamsEscudero & Williams, 2012). Also, Reference Flege, Bohn and WaylandFlege and Bohn (2021, p. 38) admit that two practicalities motivated their call for this methodological shift in L2 speech research: the difficulty to constitute groups differing in one variable only and the challenge to always be able to draw meaningful generalisations from group data. Future L2 research conducted within the SLM-r framework should therefore refocus on individual L2 learners and adapt its research design accordingly, i.e., ‘obtaining enough data from each participant to permit treating each individual as a separate experiment’ (p. 38). Section 3 explains this more specifically and suggests what research designs and statistical analyses may be most conducive to such a new agenda in L2/L3 phonology research.

In summary, the PAM/PAM-L2 and the SLM/SLM-r relate the development of perception and production abilities in an L2 to the perception of cross-language (dis)similarity. Specifically, the PAM/PAM-L2 proposes that L2 contrasts that are phonetically and phonologically assimilated to a single L1 category will be the most challenging for the L2 learners to perceive, whereas the SLM/SLM-r contends that L2 sounds (rather than contrasts) that are perceived as phonetically similar to L1 sounds will be perceived and produced in terms of those L1 categories, and thus be the most challenging to achieve. Both models propose that L2 learners (children and adults alike) will never show monolingual-like performance because L1 and L2 categories coexist in the same phonological space. The models also agree that L2 learners can improve in their L2 speech learning as a result of rich and relevant L2 phonetic input. The models diverge in their explanation of the processes involved in L2 speech learning. Whereas the PAM/PAM-L2 posits that L2 learners extract information about articulatory gestures from the speech signal, the SLM/SLM-r focusses on the development of phonetic categories from acoustic-phonetic cues. Likewise, the models disagree on the linkage between perception and production: while the PAM-L2 postulates no direct link between perception and production (Reference Tyler, Nyvad, Hejná, Højen, Jespersen and SørensenTyler, 2019, p. 619), the SLM-r proposes a co-evolution of the two modalities. Finally, the new goal of the SLM-r is to account for how individual L2 learners reorganise their phonetic system over time in response to the phonetic input received.

Apart from perception-based models explaining the learning trajectory of L2 learners, theoretical models that focus on various aspects of acquiring L2 speech production have also been advanced (e.g., Reference EckmanEckman, 1981; Reference LadoLado, 1957; Reference Major and CookMajor, 2002). Although this theorising is less prominent in present research into L2 phonetic and phonological acquisition, one production-based model seems relevant for its developmental perspective on the process of L2 speech learning: Reference MajorMajor’s (2001, Reference Major and Cook2002) Ontogeny and Phylogeny Model (OPM) posits that L2 speech learning is influenced by a learner’s L1, L2 and certain universal processes common to language learning and human language in general. One example of such a universal process is final obstruent devoicing (e.g., bed pronounced as [bet]), a phonological process resulting in a less marked structure: voiceless codas are unmarked in comparison to their voiced counterparts in that the former could be understood as ‘simpler’. Various criteria have been proposed to determine which structures are less marked, including implicational universals (i.e., ‘if A then B’, but not the other way around), how frequently the structure occurs across languages and the order in which children acquire the structure in their L1.

The OPM proposes an initial dominant influence of the L1 in L2 speech learning, the emergent influence of universals and finally the possibility of target-like L2 speech production, especially for structures dissimilar to L1 and for less marked structures. Interestingly, Reference MajorMajor (2001) claims that the model is equally valid for multilinguals, for whom ‘L1’ can be replaced by ‘L1 plus L2/L3/Ln’. However, there is no proposal in the model for either the mechanisms that trigger the influence of L1 (and L2) and universals or the role of input in L2/L3 speech development.

At present, there is no model of speech learning from a multilingual perspective, largely because of our scarce knowledge about the concurrent phonological development in a multilingual speaker’s different languages. The aim of this Element is therefore to contribute to the investigation of multilingual phonological development with new data, including all languages of the multilingual speakers, both adolescents and adults, and in terms of both their speech perception and production. We will trace the interaction of a range of phonological features and of individual speakers’ languages with a view of offering new insights into the L3 speech learning process.

3 The Conceptualisation of Foreign Language Development

4 Study: Multilingual Development of Phonetics and Phonology

4.1 Method

4.1.1 Design

A longitudinal study design was implemented to shed light on intra- and inter-individual processes of the phonetic and phonological development of multilingual speakers. A group of sixteen learners with a similar language profile (L1 German, L2 English, L3 Polish) was recorded in all of their languages four times (T1–T4) within the first year of L3 learning. In addition, in between T2 and T4, a small subset of these participants consisting of three learners recorded themselves at home in their L2 and L3 every week for eighteen weeks (dense data collection; DDC). All testing sessions were designed to elicit certain phonetic and phonological features which were either shared by the languages or not (see Section 4.1.4). This way, potential interactions between the language subsystems over time can be revealed.

Figure 1 visualises the longitudinal data collection with four main group testing times. It commenced a month after the start of L3 Polish lessons and stretched over the first ten months of L3 learning (T1 at five weeks, T2 at ten weeks, T3 at five months and T4 at ten months into L3 learning). Due to the expectation that there would be less development in the learners’ L1 than in their non-native languages, L1 production data was collected at T1 and T4 only, along with other background measures such as the learners’ language attitudes and cross-linguistic phonological awareness. In their L2 and L3, the participants completed various production and perception tasks (see Section 4.1.3) at all four group testing times. Between the third and seventh month of L3 learning, three adults each participated in eighteen weekly dense testing sessions for which they recorded themselves at home in their L2 and L3. Each week they also provided information on their language learning activity and progress that week. The general language learning profile of these three as well as the other participants are described in the following section.

Figure 1 Visualisation of data collection timeline.

4.1.2 Participants

In total, sixteen beginner learners of L3 Polish took part in the longitudinal study. Nine of them were adolescents of the same age group (12–13 years; 5 females, 4 males), and the remaining seven adults were between 21 and 39 years old (5 females, 2 males). All of them shared German as L1 and English as L2, the latter of which they had learned or were still learning in school within their L1 environment (Germany). In addition, five of the adults also had some knowledge of French and/or Spanish, which they had acquired in secondary school. Although they were tested in these additional languages as well, only English and Polish are the focus of this enquiry, as these are the foreign languages the participants were actively using at the time this study was conducted. The two age groups were matched regarding their L3 Polish input, which amounted to three hours per week. For all participants, Polish was a completely new language, i.e., all heritage speakers were excluded from the final data set, as were all early bilinguals.

The adult participants, aged 21–39 (mean age = 26.4), were recruited from community college and university language classes and received a small financial compensation for participating in the study. They had started learning their L2 English in a formal context (at primary or secondary school) at a mean age of 9.4 years and self-assessed their language skills as upper-intermediate to advanced.

Three out of the seven adults participated in the DDC, recording themselves at home in their L2 and L3 every week for eighteen weeks (see Figure 1). The two females, REBA03 (aged 28) and ROGI18 (aged 32), decided to take Polish classes with the aim to be able to better communicate with their partners’ Polish families (they both spoke German to their Polish partners). REBA03 was a university librarian, and ROGI18 was in the second year of studying to become a secondary school teacher of English and German. The male participant, SYLÜ08 (aged 22), was a politics student who took Polish classes before going to Poland for four months as part of an Erasmus exchange. He was still in Germany at T1 and T2 of the group testing times and the first nine weeks of the DDC. From DDC week 10, he recorded himself from Poland. The final group testing time (T4) took place shortly after his return to Germany. These three participants were chosen for the dense data collection because they had demonstrated at T1 a number of non-target-like pronunciations in both their L2 and L3, which made changes in their phonological system more likely than for the other participants who showed mainly target-like pronunciations in their L2 or L3 or both. Shifts in the phonological system were expected especially for SYLÜ08, given the changes in the language learning environment in the course of the study.

The younger L3 learners, aged 12–13, were recruited at a school located in Germany within close proximity to the Polish border. They all chose Polish as a new subject (over French) and were total beginners as well. All of them had been receiving instruction in their L2 English for several years, their mean age of onset of L2 learning being younger than those of the adults at 6.5 years. The difference of the two groups’ starting age has to do with educational policies in Germany that had changed over the years regarding the school year in which English is introduced as a mandatory subject. Alongside their Polish lessons, the adolescents also had four 45-minute English lessons at school (a total of three hours per week) throughout the time span of the research project. They self-assessed their L2 skills as lower-intermediate.

4.1.3 Data Collection

The data collection comprised a language background interview, three different production tasks and two perception tasks. For each language (L1 German, L2 English and L3 Polish), the tasks were administered on different days and by a research assistant who addressed the participants in the target language only. The research assistants were native speakers of German and Polish respectively and native or near-native speakers of English.

Language background interview.

An extensive language background interview was conducted as part of the L1 recording sessions at both main testing times (T1 and T4). This structured interview covered many aspects considered potentially important with regard to the participants’ language development such as their language learning history, language use and attitudes.

Production task 1: Delayed repetition (DR).

In this task, the participants heard a stimulus word in a carrier phrase in the L1, L2 and L3 (‘Ich sage X zu dir’ in German, ‘I say X to you/again’ in English and ‘Mówię X do ciebie’ in Polish). After a prompt (e.g., ‘And what do you say?’ in English) spoken by a different speaker, the learners repeated the entire first phrase (with the inter-stimulus interval set at 1,000 ms). The delayed repetition task was selected as it effectively elicits speech production even in low-proficiency speakers and reduces the risk of direct imitation by including a distractor phrase. The stimuli were presented in two randomised orders and embedded in a PowerPoint presentation as a self-paced task. The testing block was preceded by a practice block of three trial items.

Production task 2: Picture naming (PN).

In this task, the participants were presented with a series of simple pictures in a PowerPoint presentation (one picture per slide). They were selected from a range of categories that language learners are typically exposed to in the earlier stages of language learning such as colours, animals and foods. Concerning the L3 stimuli, the respective Polish teachers had been consulted to assure that the vocabulary items had been covered in class. This task was self-paced.

Production task 3: Storytelling (ST).

The participants were given a different eight-frame picture story for each language to elicit free speech. They were granted as much time as they needed to familiarise themselves with the pictures and the simple storyline before they were asked to tell it. The picture stories contained a number of objects and actions suitable to elicit the items of interest, and the research assistant administering the task made sure to guide and prompt the participant to name some of them after the participants’ first rendition if necessary.

Perception task 1: Forced-choice goodness (FC).

In the forced-choice goodness task, the participants heard two renditions of the same phrase differing only on the stimulus items embedded in a carrier phase and had to decide which one sounded more natural. One rendition was a target realisation and the other was an accented language realisation in which only the target sound was manipulated, i.e., realised as the L1, L2 or L3 sound respectively. For example, in the L2 English version of the task, the rhotic stimuli included the target-like phrase ‘You will hear the word ring [ɹiŋ]’ followed by the Polish-like realisation of the rhotic sound ‘You will hear the word ring [riŋ]’, as well as stimuli including pairs of English and German-like rhotics. In the L3 Polish session, the stimuli included Polish target rhotics as well as German and English rhotics. The carrier sentences were always consistent with the language of the session.

The stimuli were recorded by female native speakers of the respective languages who were fluent advanced speakers of the other two languages in the triad of languages. The stimuli were randomised and counterbalanced across trials in E-Prime 2.0 (Psychology Software Tools, Pittsburgh, PA). The testing block was preceded by three trial items. The inter-stimulus interval was set at 500 ms and the participants had a 3,000 ms response limit. The response accuracy and reaction time were recorded.

Perception task 2: ABX.

While the forced-choice task was administered to test the learners’ judgement of naturalness employing cross-language sound contrasts, ABX tasks test language-specific contrasts (Reference Strange, Shafer, Hansen and ZampiniStrange & Shafer, 2008). Such timed discrimination tasks assess the learners’ ability to distinguish between two minimally contrasting sounds within a language, as for example /v/ and /w/ in English and Polish. Therefore, this perception task was adopted for testing the phonemic contrast between /v/ and /w/ but not the allophonic contrasts of the other three features investigated in the study. In the ABX task, the participants listened to a pair of stimuli (AB) recorded by one speaker. Subsequently, they heard a third stimulus (X) produced by a different speaker which was either the same as the first (A) or the second one (B). The participants then had to decide whether the final stimulus (X) was more like the first (A) or the second one (B) by pressing a button on a button box (see Figure 2 for a visualisation of the task procedure). The trials were randomised so that the minimal pairs included appeared in all possible combinations (ABA, ABB, BAB, BAA). Like in the case of the FC task, the ABX task was administered through E-Prime 2.0 measuring accuracy and response times. If no response was logged within 3,000 ms after the last stimulus, the trial was coded as incorrect, and the experiment automatically proceeded with the next trial.

Figure 2 ABX task procedure.

(adapted from Reference NelsonNelson, 2020; p. 38). ISI = inter-stimulus interval, RW = response window

Production tasks dense data collection.

The weekly dense data collection included a picture naming, a delayed repetition and a different free speech production task in the L3 every week, with the free speech tasks having been carefully compiled in line with current topics and vocabulary from the participants’ L3 classes. For the L2 session, the free speech element was the only task since the learners were much more fluent in English and were expected to produce sufficient tokens of interest in this single task. To help the participants to tune into the appropriate language mode, the free speech prompt was a short video recording featuring the L2 research assistant (who also collected the L2 data at the group testing times, so they were familiar with the person and already associated them with English). In the video, the research assistant briefly talked about a specific topic and then asked an open-ended question related to that topic which the participants were asked to respond to or talk about for about one minute. Perception dense data in the L2 and L3 was also obtained, but it is not included in our analyses here.

5 Results

First, we present an overview of the whole data set for perception and production results separately. Figure 3 displays the perception accuracy of the /v-w/ contrast, the rhotics, final voiced obstruents and vowel reduction by the sixteen learners in both their L2 English and their L3 Polish at T1, T2, T3 and T4 (for the features rhotics, final voiced obstruents and vowel reduction no data was collected in L3 Polish at T1). Only the perception of the /v-w/ contrast was elicited via an ABX task, while the perception of the remaining features was elicited via an FC task. It shows that the perception accuracy varies across the phonological/phonetic features and processes with the perception of the rhotics having the highest and the perception of final voiced obstruents having the lowest accuracy overall. Only for the perception of /r/ in L3 Polish and the perception of final voiced obstruents in both L2 and L3 can differences between the adolescents and the adult learners be observed. Overall, perception accuracy of a particular phonological feature or process is very similar in both the learners’ L2 and the L3. Variation across the individual speakers, however, is evident for all features except the rhotics, where perception accuracy is above 60 per cent in their L2 English for all learners.

Figure 3 Mean accuracy of perception of /v-w/, the rhotics, final (underlying) voiced obstruents and vowel reduction by the sixteen learners in their L2 English and their L3 Polish at T1, T2, T3 and T4 (at T1 in L3 Polish only for /v-w/). Adult learners are represented by dotted lines. The group mean values are indicated by the bold line.

Figure 4 shows the mean values of accurate production of /w/, the rhotics and vowel reduction (= the mean vowel ratio) by the sixteen learners in both their L2 English and L3 Polish at T1, T2, T3 and T4 (only T2 and T4 for vowel reduction). The production results for final voiced obstruents are not reported at this point as this feature was analysed in the study in terms of three different duration parameters rather than an accuracy score. Both /w/ and the rhotics show distinct patterns from each other but appear similar across both languages and both age groups: while /w/ is produced fairly accurately in L2 English and L3 Polish by most adolescents and adults, the accuracy of the rhotics is much more diverse with individual learners at ceiling level in both languages and others not producing them accurately at all. The two age groups were also comparable in their production of vowel reduction in that they showed some progress, in a target-like direction, with less reduced forms in their L3 Polish than in their L2 English, as suggested by the score for mean vowel duration ratio around 1.0 for L3 Polish. A high degree of inter-learner variability was nevertheless also evident in the L3 Polish.

Figure 4 Mean accuracy of production of /w/, the rhotics and mean vowel duration ratio by the learners in their L2 English and L3 Polish. Values of around 1.0 for mean vowel duration ratio indicate the same length of stressed and unstressed vowels. Adult learners are represented by dotted lines. The group mean values are indicated by the bold line.

Here, the results for each of our hypotheses (see Section 3.4) are presented.

Figures 3 and 4 show that the group average values depicted in bold do not model well the learning processes of the individual learners. In fact, not a single one of the sixteen learners shows a development that matches the group mean value in any of the phonetic or phonological features or two languages.

Figure 5 displays the sixteen learners’ variability of producing the rhotics in their three languages at T1, T2 and T4. The variability was calculated by dividing the number of different sounds per token produced at one point in time, irrespective of them being target-like or not, and multiplied by 100. For example, BISC14 produced five words containing /r/ in Polish at T1 and realised the rhotic in three different ways: three times as the uvular trill [ʀ] and once each as the uvular fricative [ʁ] and the voiceless velar fricative [x]. Her calculated variability for L3 Polish at T1 is thus 3 (different realisations) / 5 (tokens) * 100 = 60.

Figure 5 Variability of rhotic production for all learners in their three languages at T1, T2 and T4 (no L1 data is available for T2).

Figure 5 shows clear differences across the multilinguals in terms of the variability of their realisations at the three data points T1, T2 and T4. DIMO03 and ROGI18 appear to be in a stable period of pronouncing the rhotic sound in their L2 English: they consistently realise it as the approximant [ɹ] during the ten months of our study but show some variation in their pronunciation of the rhotic in both their L1 and L3. While for DIMO03 this variation is entirely target-like in both L1 and L3 – for example, producing both the trill [r] and the tap [ɾ] in Polish (see Table 1 in Section 4.2) – ROGI18’s production of the Polish rhotic varies between [r], [ɾ] and non-target ‘German’ [ʁ]. Similarly, the three multilinguals BISC14, MESC03 and SMSC15 all demonstrate a non-variable realisation of the rhotic in their L1 German but exhibit some variability in the production of the rhotic in both L2 English and L3 Polish. Yet, while BISC14 produces non-target-like rhotics such as [ʁ] in both English and Polish at all data collection points and SMSC15 produces exclusively target-like rhotics in English only at T2 and never in Polish, MESC03 only produces some non-target-like rhotics in English at T2 and T4. For EDMÜ06, variability occurs in all three of his languages at T1 but is restricted to L3 Polish at T2 and T4. Thus, for five out of the sixteen multilinguals, production of the rhotic in one language is stable, which points towards an attractor state of this subsystem during the ten months of our study.

Another interesting observation concerns potential changes in some of the learners’ L1. Six of the multilinguals produce the voiceless velar fricative [x] or a combined [ʁ/x] sound as the German rhotic at T4. In fact, half of the learners produced such sounds in their L3 Polish in the very early stages of learning to attempt the trill [r] when they were not able to produce it at all or not consistently. This non-acceptable realisation of rhotics in Polish, which is used in their L1 German at T4, can thus tentatively be interpreted as a special kind of L3-to-L1 influence, i.e., regressive CLI from an earlier production in the L3 to the L1 at a later stage of learning. This example highlights the complexity of cross-linguistic interaction, where an ‘L1-coloured’ sound from the learners’ L3 repertoire can find its way back into their L1. Similarly, the production of uvular trills [ʀ] in German increases from T1 to T4. This can also be argued to be evidence for regressive CLI, as the learners produce more of the L1 variant with the L3-like manner of articulation. Hence, on this feature, these two languages seem to be more prone to mutual influence for some of the learners with their distinct yet articulatorily more similar rhotics than the English approximant [ɹ]. The L1 rhotic sounds (and other articulatorily similar L1 sounds) clearly seem to be the preferred basis for those struggling with the Polish tap or trill, which in turn also seems to make their L1 German more receptive to Polish-like sounds.

Figure 5 thus also shows highly individual relationships between variability at the micro level, i.e., single realisations at one particular data collection point, and long-term development at the macro level, i.e., a potentially stable state after ten months of learning. Of those multilinguals that show target-like pronunciation of the rhotic in L3 Polish at T4, three had target-like pronunciations from T1 onwards (DIMO03, LYBO29, MESC03). SYLÜ08 produced both [r] and [ɾ] at T1, varied between [r], [ɾ] and [ʁ] at T2 and returned to producing [r] and [ɾ] at T4. DOSC23 shows substantial variability in realising the Polish rhotic at T1, including [ɾ], the voiceless velar fricative [x] and a combined [ʁ/x] sound, but then consistently realised it as [ʁ] at both T2 and T4. For both SYLÜ08 and DOSC23 thus there is cross-linguistic influence from the L1 at one point in time reflecting the competition of the different realisation of the rhotic in their languages, which formed the basis for dynamic developments at later stages. All of the other multilinguals’ pronunciation of the Polish rhotic is still variable and includes non-target forms at T4, hence reflecting a dynamic, non-stable state in their development.

Zooming into the variability at the micro-level and mid-term development, Figure 6 displays min-max graphs for the production of /w/ in L2 English and L3 Polish by SYLÜ08 across the eighteen weeks of dense data collection. It shows that, overall, variability in the accuracy of producing /w/ at each weekly data collection point is higher in L2 English than in L3 Polish. Only at around week 15 of the observation period does the production of English /w/ stabilise.

Figure 6 Accuracy of production of /w/ in L2 English and L3 Polish by SYLÜ08 across the eighteen weeks of dense data collection.

For ROGI18, by contrast, more variability in the production of /w/ occurs in L3 Polish than in L2 English (see Figure 7). She produces English /w/ most accurately at the start of the observation period, which is then followed by some weeks of increased variability that diminishes again at week 14 but ultimately returns. No clear picture arises as to how this variability at the micro level may affect the speakers’ mid/long-term development.

Figure 7 Accuracy of production of /w/ in L2 English and L3 Polish by ROGI18 across the eighteen weeks of dense data collection.

As can be seen in Figure 5, two multilinguals show extreme variability (80 per cent) in their rhotic productions at some point in time: COSC26 in L3 Polish at T2 and CHSC09 in L3 Polish at T1. CHSC09 realises the five Polish words containing a rhotic with the L1 sounds uvular fricative [ʁ] (twice) and uvular trill [ʀ] as well as the combined sounds [k/ʀ] and [x/ʀ] at T1. At T2, she produces three realisations: [ʀ] (three times), [x] (once) and [ʁ] (four times). At T4, however, her Polish rhotics show only two different realisations, [ʁ] six times and [ʀ] three times. While for CHSC09 the extreme variability of rhotic productions at T1 might be interpreted as heralding an upcoming stabilisation of the system, this hypothesis does not hold for COSC26’s development. She produces three different realisations ([ʁ] four times as well as [v] and [k/x] once each) in the six Polish words containing a rhotic at T1 and then shows extreme variability at T2, when she realises the five Polish words containing rhotics with the English approximant [ɹ] twice as well as [w], [ʀ] and [ʁ] once each. At T2, thus, her realisations of the Polish rhotic consist of both L2 and L1 sounds but no target-like sounds. At T4, however, she produces the seven Polish words containing rhotics with a target-like trill [r] four times, as well as [ʁ], [ɹ] and [l] once each. Her subsystem thus has not stabilised after ten months of learning Polish but now contains L1, L2 and L3 realisations of the rhotic. The high variability in her productions observed at T2 can still be interpreted as signalling a major developmental step insofar as her productions of the Polish rhotic at T4 for the first time include some target-like realisations.

To identify possible major developmental changes in the production of the other features also tested in the study, we performed change point analyses of the weekly collected L2 and L3 vowel reduction and final obstruent (de)voicing by SYLÜ08, ROGI18 and REBA03. Figure 8 displays the identified phase shifts or lack thereof for the ratio of vowel duration in stressed/unstressed vowel pairs (values of around 1 indicate the same length) in two speakers. The two lines in the graph are control limits which represent the maximum range over which the values are expected to vary assuming no change has occurred. Points outside the control limits indicate that a change has occurred, visually represented by the dark grey shifts in the background, in which case the accompanying table provides further information describing the change(s).

Figure 8 Change point analyses of vowel reduction ratio in L2 English and L3 Polish by SYLÜ08 (eighteen weeks of dense data collection) and ROGI18 (sixteen weeks of dense data collection).

SYLÜ08 demonstrated no developmental shift in realising vowel reduction in his L2 English in the observation period, as evidenced by the absence of any dark grey shifts in the background; the L2 vowel ratio values still fluctuate on a weekly basis. He did, however, change his realisation of unstressed vowels in relation to the stressed vowels in his L3 Polish at around week 10: prior to the change, the average vowel reduction ratio yielded a value of 0.87, while after the change it was 0.97. This change occurred with a 95 per cent confidence and level 4 was associated with it, indicating the importance of the change (the lower the level, the higher the importance). ROGI18, in turn, showed no significant changes in realising vowel reduction in L3 Polish but a developmental shift in her L2 English instead. At around week 7, her vowel reduction ratio dropped from 0.78 to 0.64, a change that occurred with 94 per cent confidence and was detected on a third pass of the data (level 3). For both speakers, the identified developmental shifts exhibited a target-like direction. Analyses of respective shifts as per standard deviation changes yielded no significant results, which suggests that they were not preceded by extreme levels of increased variability.

Figure 9 presents selected results for significant changes in the duration parameters of final obstruent (underlying) voicing in three speakers in their L2 English or L3 Polish. It shows that all three speakers changed their consonant duration (rather than vowel or consonant burst duration that were also measured) in realising final obstruents in their L2 English or L3 Polish. It is notable that these shifts approach the expected target in that voiced realisations are to be expected to show shorter consonant duration than devoiced realisations (as manifested in REBA03’s production of shorter voiced final obstruents in English, and SYLÜ08’s and ROGI18’s production of prolonged devoiced final obstruents in Polish).

Figure 9 Selected change point analyses for final obstruent (de)voicing in L2 English or L3 Polish by SYLÜ08 (eighteen weeks of data collection), REBA03 and ROGI18 (sixteen weeks of data collection).

SYLÜ08 evidenced one significant shift in his realisation of L3 consonant duration (with 99 per cent confidence, level 1 change) at around the same time as his realisation of vowel reduction changed (see Figure 8), which coincides with his move to Poland for a stay abroad. Interestingly, REBA03 evidenced two significant shifts in her production of L2 English consonants in that she lengthened her word-final obstruents from around week 10 of the data collection (97 per cent confidence, level 1 change) but five weeks later returned to her original manner of realising the sounds (the average length stabilising at 0.14 ms; 94 per cent confidence, level 2 change), which is suggestive of the speaker’s preferred, ‘attractor’ length. The third speaker, ROGI18, significantly changed her production of L3 Polish word-final obstruents in terms of both consonant duration and burst duration at around week 10 of the data collection (with 95 per cent and 97 per cent confidence, respectively, both changes identified at level 1). An inspection of change point analyses for standard deviation and the identified developmental shifts indicated no significant results. There is thus little evidence in the present dense data on vowel reduction and final obstruent (de)voicing for extreme variability as an index of upcoming major developmental changes for the three speakers.

Figures 10 and 11 display the accuracy of producing both /w/ and the rhotics across the four data points for each speaker. A clear interaction can be seen between the development of the two sounds in their L2 English for the four speakers shown on the top row, DIMO03, REBA03, JUEB20 and SMSC15. Although on different levels of accuracy (e.g., ceiling performance for DIMO03 for both sounds but relatively low accuracy for /ɹ/ produced by JUEB20), the production of those two sounds for these four speakers show parallel increases and decreases of accuracy from T1 to T4. No such relationship could be found for the remaining twelve speakers, where the accuracy of their L2 English /w/ and /ɹ/ production seems to develop independently, though perhaps one might suggest some linkage between the two sounds for SYLÜ08.

Figure 10 Relationship between accuracy of producing /w/ (grey line) and /ɹ/ (black line) in L2 English for the sixteen speakers.

Figure 11 Relationship between accuracy of producing /w/ (grey line) and /r/ (black line) in L3 Polish for the sixteen speakers.

There is little evidence for these sixteen speakers exhibiting parallel development in producing /w/ and /r/ in their L3 Polish (see Figure 11). Many speakers never pronounce Polish /r/ correctly during the first ten months of learning this language, especially those shown on the last row, SISC11, CHSC09, REBA03 and EDMÜ06, and also JUEB20 on the top row, while BISC14 and JOHA09 only produce one token correctly at T3. Some tendencies for parallel developments of the production of the two sounds are evident for DIMO03 from T3 and LYBO24 at T4, but in general the two sounds do not develop together in the L3 Polish.

Tables 2–5 show how many individual learners evidenced different patterns of a perception-production relationship for L2 and L3 rhotics and /w/ at each of the main testing times, as well as across the testing times (the last column in the tables). The categorisation of these patterns is based on the calculation of within-subject effect size (Cohen’s d) for perception and production differences in mean scores (d ≤ 0.7 = small to no effects, d ≥ 0.7 = medium effects and d ≥ 1.0 = large effects; cf. Reference Plonsky and OswaldPlonsky & Oswald, 2014). That is, a small within-subject effect size was interpreted as suggesting no meaningful difference between the production and perception scores and thus some evidence for a co-evolution of the skills for a learner (‘perception = production’). Medium and large within-subject effect sizes suggested a precedence of one skill over another for a learner (either ‘perception > production’ or ‘production > perception’). If one or both of the skills equalled or fell below a 50 per cent accuracy threshold, this was interpreted as pointing at a ‘dissociation’ between a learner’s perception and production, suggesting lack of ability and hence little direct relationship. For an overview of individual learner trajectories and perception-production relationships, see the Appendix.

As shown in Table 2, overall, the learners in this study perceived L2 English rhotics equally accurately as they produced the sounds (eleven learners at T1, eight learners at T2, five learners at T3 and ten learners at T4). Six learners (five adults and one adolescent) showed a consistent alignment in perceiving and producing L2 English rhotics across the four testing times (note that the total count across the four testing times, provided in the last column in Tables 2–5, may differ from the number of participants as not all of the learners evidenced a consistent pattern of a perception-production relationship across the main testing times). The young learners tended to manifest more varied patterns at each testing point, with two of them evidencing a consistent dissociation of their perception and production of L2 rhotics, i.e., one or both of their skills in the L2 showed an accuracy level equal or below the 50 per cent threshold. As shown in Table 3, this latter pattern was predominantly found for both learner groups in their L3 Polish: nine out of the sixteen learners consistently perceived alveolar trills fairly accurately (as tested in the FC task) but were not able to produce them accurately. Yet six learners performed equally well in both modalities in their L3 across the testing times, suggesting a co-evolution of their L3 perception and production skills.

The same analyses concerning the perception-production link for rhotic sounds were also carried out for the labiovelar approximant /w/. Tables 4 and 5 show that the two modalities developed in alignment for the majority of the adolescent and adult learners for this phonological feature, in both their L2 English and L3 Polish. For example, ten learners perceived L2/L3 /w/ as accurately as they produced the sound across the four testing times. If a dissociation relationship was identified, this was usually a scenario of a learner producing /w/ in their L2/L3 more accurately than perceiving it (as tested in the ABX task), with one or both of the skills equalling or falling below the 50 per cent threshold though.

Overall, our data suggests that perception and production of speech sounds tend to co-evolve, although the strength of the link can be moderated by the phonetic characteristics of the sound concerned and the stage of language learning.

Figure 12 shows the relationship between the production of /w/ by SYLÜ08 in L2 English and L3 Polish as a moving correlation. It shows that in the first ten weeks the relationship is a competitive one, in that his production of the sound is negatively related in his two foreign languages (r = between −0.20 and −0.80), but the relationship turns into a supportive one from week 13 to 15 (r = between 0.60 and 0.80).

Figure 12 Moving correlation for the production of /w/ in L2 English and L3 Polish by speaker SYLÜ08.

This is in contrast to the relationship found for the production of vowel reduction in L2 English and L3 Polish by the same speaker, which is far more fluctuating (see Figure 13). In the first two weeks as well as weeks 6 to 8 of the observation period, vowel reduction in both of his languages develops in alignment, i.e., the vowel reduction ratio is positively related in his two foreign languages (r = between 0.20 and 0.60), while for weeks 3 to 5 it is negatively related (r = between −0.20 and −0.80). An increased distinction between vowel reduction in L2 and L3 can again be seen for weeks 9 to 12 (r = between −0.20 and −0.80) before it diminishes again towards the end of the observation period.

Figure 13 Moving correlation for the production of vowel reduction in L2 English and L3 Polish by speaker SYLÜ08.

Figure 14 displays the moving correlation for the production of /w/ in L2 English and L3 Polish for ROGI18. Like for SYLÜ08 for this phonological feature, an initially competitive relationship between the two languages turns into a primarily supportive one (weeks 6 to 12). However, it returns to being competitive again towards the end of the observation period. Thus, for both speakers and for the production of /w/, the increase of proficiency in L3 Polish does seem to go hand in hand with a more supportive relationship between the learners’ two foreign languages, which tends to be more competitive at earlier stages. However, the production of vowel reduction in the two languages appears to have a rather more complex relationship, vacillating several times between competitive and supportive.

Figure 14 Moving correlation for the production of /w/ in L2 English and L3 Polish by ROGI18.

6 Discussion

It was one aim of this study to describe the phonetic and phonological development of multilingual speakers within the framework of CDST and to thereby both test some of the main tenets of this theory and contribute to it. Our results showed that, as hypothesised in H1, learning trends in phonetic and phonological development based on group scores do not model well the learning processes of individual learners. Confirming findings by Reference van Dijk, Verspoor, Lowie, Verspoor, de Bot and Lowievan Dijk et al. (2011) for the development of grammar, our data also showed, for each of the four analysed phonetic and phonological features and for both modalities, production and perception, that the learning trajectories of individual learners vary substantially and cannot be represented by the trend that results from calculating group mean values. As proposed within the framework of CDST, speaker variation should thus be treated as a meaningful aspect of language development in L2/L3 research and group means need to be interpreted with care. Moreover, the individual learning trajectories in our data show that development in most cases is non-linear: the accuracy in perception and production of all analysed phonetic and phonological features shows drops across the ten months of the observation period for many of the sixteen learners. The results thus complement the PAM-L2 and the SLM/SLM-r, which propose a continuous, slow refinement of L2 learners’ perception of phonetic differences between their L1 and L2, with a more refined insight into the developmental dynamics of the process. While the upward trend might apply on a very general level of language development (i.e., the macro level; see Hammarberg, 2017), it does not hold entirely true for the actual developmental process at close inspection, which instead involves a back-and-forth progress. The results of the present study thus support the CDST conceptualisation of the language learning process as dynamic, non-linear, variable and highly individual.

The results further show that variability exists not only across different learners and their processes of multilingual language development but also within learners at different phases of their phonetic and phonological development. With the second hypothesis, we wanted to explore a possible relationship between learners’ variability at the micro level and their long-term developments by comparing the variability of production at one point in time (the micro perspective reflecting single events of language use; see Reference HammarbergHammarberg, 2017) with mid-term and long-term development. For both the production of the rhotics and the production of /w/, individual learners showed dynamic phases with increased variability and cross-linguistic influence from their other languages as well as phases of relative stability with less variability. As in Reference HammarbergHammarberg’s (2017) study, there was substantial variation across the sixteen learners in terms of both whether and when variability at the individual data collection points occurred and whether a period of stabilisation could be observed. Yet, no conclusive patterns were found that point towards the role variation at the micro level plays for long-term developmental trends: alternating phases of variability and relative stability in the pronunciation of individual speech sounds were observed for the learners’ L1 German, L2 English and L3 Polish, irrespective of the fact that they were beginners in Polish and much more advanced learners of English. Furthermore, some learners never showed any variability at any of the data collection points and produced the speech sounds in the same (target or non-target) way. Given an observation period of only ten months and a period of dense data collection covering only eighteen weeks, our data might not have been optimally suited for capturing long-term trends. Thus, more research is needed to explore the relationship between variability at the micro level and long-term developments.

Another aspect of variability and its relationship with the dynamicity of the phonetic and phonological learning process that was explored in this study concerns the hypothesis that extreme variability at one point in time signals (upcoming) major developmental changes, as suggested by Reference van Dijk and van Geertvan Dijk and van Geert (2007) and Reference Evans and Larsen-FreemanEvans and Larsen-Freeman (2020). No clear evidence for this hypothesis could be found: while the analysis of the four observation times spread across the first ten months of learning the L3 Polish showed that extreme variability does occur, such as five different realisations of one speech sound at one point in time, it is difficult to link them to major developmental changes. For one learner, extreme variability preceded a restructuring of the L3 Polish subsystem that then included realisations from the L1, L2 and L3; for another learner, extreme variability was followed by a phase of reduced variability. However, due to the fact that the data collection points were spread out over several weeks or even months, which was especially true for the L1 data collection, it is not possible to say what developments occurred between them and what effect the extreme variability may have had. Yet, even in the dense data that included data collected at weekly intervals, no clear evidence was found for increased variability before developmental changes. Future research will have to investigate whether this applies only to phonetic and phonological development or whether this seriously questions the suggestion by Reference van Dijk and van Geertvan Dijk and van Geert (2007) and Reference Evans and Larsen-FreemanEvans and Larsen-Freeman (2020) that high variability indicates developmental changes. In this connection, it will be important in future research to carefully consider the criteria that characterise patterns of discontinuity in the development of a particular feature under scrutiny (cf. Reference van Dijk and van Geertvan Dijk & van Geert, 2007). The present results suggest that extreme variability, understood as an unexpectedly large, local peak in the data, is not a prerequisite for major developmental changes.

This study traced the development of various phonetic and phonological features and, with the fourth hypothesis, aimed at exploring whether these develop in interaction with each other or rather independently. To this end, the development of the pronunciation of /w/ and the rhotics was compared in the learners’ L2 English and also their L3 Polish. The results suggest that the learning of the pronunciation of these two speech sounds proceeds largely independently and thus stands in contrast to findings by Reference Yu and LowieYu and Lowie (2019), who found an interaction between the development of oral complexity and accuracy in their L1 Chinese learners’ L2 English that developed from an initially competitive relationship to a supportive one. The discrepancy between our findings and theirs might be explained by the fact that individual speech sounds form independent subsystems within a learner’s phonology, while grammatical complexity and accuracy in oral language production depend more on each other. Yet the results once more underline the substantial inter-variability across the multilingual learners and suggest a possible role for language proficiency. For four of the learners, parallel developments between the production of /w/ and /ɹ/ were found in their L2 English only, pointing at closer interconnections in a relatively well-established subsystem for these learners and for speech sounds that are phonetically related.

The fifth hypothesis aimed to explore the relationship between the perception and production of individual speech sounds. According to the SLM-r (Reference Flege, Bohn and WaylandFlege & Bohn, 2021), the two modalities co-evolve, implying that there is a close relationship between perception and production development, while the PAM-L2 postulates no direct link between them. Our results show that the two skills do tend to develop in alignment, although the strength of the relationship is moderated by the speech sound as well as the stage of language learning. The different L2 and L3 rhotic sounds investigated in this study arguably present different degrees of perceptual saliency and articulatory challenge than /w/ does (Reference ArchibaldArchibald, 2009; Reference CatfordCatford, 2001; Reference Colantoni and SteeleColantoni & Steele, 2008), which may explain the more varied patterns of a perception-production relationship for the rhotics than for the labiovelar approximant. It is also to be recalled that /w/ is a sound shared in both the L2 and the L3 of the learners in this study, which may have facilitated the multilinguals’ ability to perceive and to produce the familiar phoneme in their newly learnt L3. Our results further suggest that the more established the phonological subsystem of a learner, in this study the L2 in relation to the L3, the greater the likelihood of a close relationship between the development of the two skills, which is in line with studies that have reported a stronger positive correlation between perception and production with increases in L2 proficiency (Reference CardosoCardoso, 2011; Reference Hanulíková, van Alphen, van Goch and WeberHanulíková et al., 2012). In a broader sense, our findings are thus in agreement with the CDST’s theorising on the interconnected nature of learner subsystems (perception and production of speech sounds in this case) and on the potentially competing demands from the different parts of the language subsystems (Reference de Bot, Lowie and Verspoorde Bot et al., 2007) such as the regulation of sensory and motor processes, activation of the lexicon and syntax, which will be especially operative in the beginning stages of learning a foreign language. Our analyses of individual learner performance show yet again that the degree of correspondence between perception and production as well as any changes in that correspondence over time differ for individual multilinguals.

The present results clearly confirm that the languages of multilingual speakers interact as already documented in numerous previous studies (see Section 1). This interaction is reflected in the intra-speaker variability that was observed for many learners in all of their languages when they realise, for example, the rhotics in their L3 Polish variably with the L1 German rhotics [ʀ] and [ʁ] and/or the L2 English rhotic [ɹ] or even combinations and newly created mixed sounds such as [x/ʀ]. The last hypotheses aimed to explore the relationship between the multilingual speakers’ languages with a longitudinal perspective and tested whether the developmental relationship between a learner’s foreign languages is initially competitive and subsequently supportive. This type of interaction had rarely been tested across a multilingual speaker’s different languages (cf. Reference Huang, Steinkrauss and VerspoorHuang et al., 2020) but was found for different features of oral language development by Reference Yu and LowieYu and Lowie (2019) within the learners’ target language. In our densely collected data, we compared the accuracy of the production of various speech sounds in the learners’ L2 English and L3 Polish and found different patterns for different speech sounds. There is some suggestion that, for the production of /w/, which represents a shared novel feature in the learners’ L2 and L3, with increasing proficiency in the L3 Polish a more supportive relationship between the L2 and L3 sets in, though it tended to be more competitive at earlier stages. However, again the time window captured by our analysis might have been too small for stable conclusions as shown in the data on vowel reduction: for this phonological process, whose realisation is dissimilar in the learners’ two foreign languages, the L2 English and the L3 Polish have a rather more complex relationship that changes several times between being competitive and supportive over the course of the eighteen weeks of dense data collection. It will be important in future research to expand on the range of investigated features across a multilingual speaker’s languages to capture (and measure) the range of cross-linguistic similarity types possibly mediating the type of a developmental relationship between the learners’ languages.

7 Conclusion

We have shown that employing longitudinal and dense data with a focus on individual learners contributes important insights for describing and understanding multilingual phonetic and phonological development and that this approach can complement traditional, more product-oriented research designs. The new methods presented in Section 3 and applied in our own study are more suited to showing dynamic development of phonology and phonetics in terms of the interaction between different features as well as the interplay of the different languages than traditional average scores and group comparisons. Moreover, a focus on individual learners as well as on developments in small time windows brings to light both the variability across learners and the individual and dynamic process of their pronunciation learning.

The revised SLM-r (Reference Flege, Bohn and WaylandFlege & Bohn, 2021), with its new orientation towards investigating how the phonetic systems of individuals reorganise over time, would seem to be in line with our proposed avenue of research into multilingual phonetic and phonological learning. Also, the model’s assumption of L1 and L2 sharing the same phonetic space may well be extendable to L3 speech learning contexts: all of the multilingual’s languages share a common phonetic space and are thus subject to (dynamic) interactions in the course of the learner’s development of speech perception and production. In a similar vein, it is to be expected that perceptual assimilation of native and non-native sounds and sound contrasts as theorised in both the SLM/SLM-r (Reference Flege and StrangeFlege, 1995; Reference Flege, Bohn and WaylandFlege & Bohn, 2021) and the PAM/PAM-L2 (Reference Best and StrangeBest, 1995; Reference Best, Tyler, Bohn and MunroBest & Tyler, 2007) works in like manner, the difference being that L3 sounds and sound contrasts can be perceptually associated with either L1, L2 or both languages by L3 learners. The micro-level scrutiny of such a multilingual phonetic and phonological development in individual learners will nevertheless deserve yet greater scholarly attention and focus.

Our study also showed the limitations of a comprehensive, process-oriented approach to studying pronunciation learning by multilinguals. The inclusion of both production and perception data from three languages for each learner in a longitudinal research design naturally limited the number of phonetic and phonological features that could be investigated. Especially with younger learners, the number of tasks they are required to perform at each data collection point needs to be limited to avoid fatigue and a subsequent withdrawal from the study.

Our study constitutes only a first step in the investigation of pronunciation development of multilingual learners. We hope that many further studies using dense data and a CDST framework will follow to shed more light on this complex issue. Our study showed that an even larger time window for the dense data collection as well as its commencement earlier in the L3 learning process might have provided richer data and greater insight into the complex and dynamic process of pronunciation learning. In particular, in this way it might be better understood how individual learner productions and perception at the micro level contribute to long-term learning at the macro level. Finally, an important direction of future research will be to integrate quantitative data with qualitative (biographical) data in order to not only characterise the process of an individual multilingual’s phonetic and phonological development and changes thereof but also to better understand the underlying environmental and affective reasons for such developmental changes.