MODELS OF L2 SPEECH PERCEPTION AND PRODUCTION

Researchers have proposed a variety of models to account for L2 speech learning, but most of these models were designed to account for performance and/or learning in only one modality. For instance, PAM (Best, Reference Best and Strange1995) was initially developed to explain how naïve listeners assimilate nonnative sounds to native language categories. Best and Tyler (Reference Best, Tyler, Munro and Bohn2007) later applied this model to L2 learning, arguing that perceptual assimilation patterns determine the degree of difficulty L2 listeners experience in learning to perceive L2 contrasts. According to PAM-L2, single category assimilations, where both members of the L2 contrast are assimilated to a single L1 category and perceived to be relatively equal in terms of their goodness of fit, should be especially challenging for L2 learners. On the other hand, two category assimilations, where contrastive L2 sounds are assimilated to distinct L1 categories, should be comparatively easy to learn. PAM and PAM-L2 are built on a direct realist approach to speech perception (e.g., Fowler, Reference Fowler1986), which assumes that listeners directly perceive articulatory gestures. Thus, within this framework, perception and production should be closely intertwined even though neither PAM nor PAM-L2 directly addresses the perception-production link. What’s more, PAM-L2 lacks a strong developmental component, making it difficult to derive testable, longitudinal claims about perception-production relationships.

Another model of L2 perception is the L2 Linguistic Perception model (L2LP; van Leussen & Escudero, Reference van Leussen and Escudero2015), which is grounded in an optimality-based approach to perceptual learning. According to this model, learning is error-driven. When a learner becomes aware of a perceptual error, pathways between phonetic, phonological, and lexical tiers of representation are altered to reduce the likelihood that a similar error will occur in the future. The L2LP is intended to account for the entirety of L2 learning. Yet, like PAM, it does not directly address L2 speech production, nor does it include information on the perception-production link, although presumably accurate perception would be a necessary precondition for accurate production (Escudero, Reference Escudero and Pennington2007). Cognitive scientists working outside of mainstream SLA research have also developed models to explain perceptual learning. For example, Chandrasekaran et al. (Reference Chandrasekaran, Koslov and Maddox2014) argued that L2 speech category learning should be viewed as a general categorization problem involving both reflective and reflexive (i.e., explicit and implicit) learning systems. Overall, then, these models are useful for understanding L2 perception, but their implications for L2 production and the perception-production link are fuzzy.

With respect to speech production, the SLM (Flege, Reference Flege and Strange1995) has received far more attention than any other framework. This model was based on findings showing that even individuals who had immigrated to an L2 environment at a young age had a noticeable foreign accent in the L2 and produced L2 sounds whose phonetic characteristics did not align with those of age-matched monolinguals. The overarching aim of the SLM was, therefore, to explicate the relationship between age of onset of L2 learning, experiential variables such as quantity and quality of L2 input, and L2 pronunciation attainment. One of its key hypotheses was that many production errors have a perceptual basis. More specifically, according to this framework, individuals will be able to produce phonetically accurate L2 sounds only if they have formed a new phonetic category for those sounds. Although phonetic learning remains possible throughout the lifespan (which means that even late-start learners can modify their production), the likelihood of forming a new phonetic category is hypothesized to decrease as age of onset increases. This is because as the L1 becomes more robust, learners will have more and more difficulty discerning the subtle, yet important, phonetic differences that exist between crosslinguistically similar sounds. If learners detect these differences, then they may form a new phonetic category for the L2, which should enable, but not necessarily guarantee, accurate L2 production. If, on the other hand, they do not detect differences between L1 and L2 sounds, then they will associate the L2 sound with the L1 category, leading to accented L2 productions.

The original formulation of the SLM was squarely focused on phonetic ultimate attainment in highly proficient L2 users and posited a unidirectional pathway of accuracy in perception shaping accuracy in production. In contrast, in the revised model (SLM-r), Flege and Bohn (Reference Flege, Bohn and Wayland2021) have posited a bidirectional, co-evolving perception-production link, which means that perception and production should mirror one another (i.e., should be somewhat synchronized) during L2 learning. The SLM-r also moves away from an emphasis on phonological end states, favoring instead a developmental approach that would entail tracking when learners begin to discern differences between phonetically similar L1 and L2 sounds and how doing so facilitates the formation of new L2 phonetic categories.

However, as Flege and Bohn have acknowledged, the time course of L2 category formation is not well understood, nor are the events that potentially catalyze it: “At a later and as-yet undefined moment in phonetic development, the perceptual link between the L2 ‘equivalence’ class and the L1 category will be sundered. We speculate that this delinking may be speeded by growth of the L2 lexicon, at least in literate learners of an L2” (2021, p. 26). The SLM-r also addresses a range of learner differences (e.g., auditory processing) that could account for variation in development.

Although the SLM(-r) offers the most robust starting point for evaluating perception-production links in L2 learning, there are many critical questions that the model does not fully address. For one, the notion of crosslinguistically similar sounds encompasses a range of qualitatively distinct learning targets and crosslinguistic relationships. In some cases, both the L1 and the L2 might contain the same phonological categories, and those categories might be implemented using the same phonetic cue, but precise perceptual crossover boundaries and production values differ. In other cases, the L2 might contain a three-way phonological contrast where the L1 has only a two-way contrast, and the L2 contrast may be implemented using different phonetic cues. Minimally, it would be reasonable to expect different developmental timelines for each scenario. It could also be the case that the functional form of the perception-production relationship varies according to such crosslinguistic relationships.

Finally, it remains an open question whether individual differences affect the speed with which perception and production become aligned and the strength of that alignment. Overall, then, to enhance current models and develop a comprehensive understanding of perception-production relationships in L2 learning, more research is warranted in three areas: tracking the link over time, examining the link for different types of crosslinguistically similar sounds, and investigating individual differences in each modality and in the link itself. In the following sections, we outline how research in each of these areas can inform theory.

DEVELOPMENTAL STAGE

One issue that should be at the forefront of perception-production research is L2 speakers’ developmental stage. Developmental stage is important because the perception-production relationship is time-varying. Conceptually, this means that a complete understanding of the link rests upon examining how it changes over time, that is, how perception and production interact as L2 speakers are exposed to different types of input and engage in varying levels of L1 and L2 use in different learning environments. On a practical level, an inherently time-varying link means that cross-sectional studies may over- or underestimate the link depending on the precise moment at which L2 speakers are measured, leading to a narrow or truncated view of perception-production relationships. Put another way, interpreting cross-sectional perception-production findings as representative of the underlying nature of the perception-production relationship risks reducing a dynamic developmental phenomenon to a static snapshot.

There are several studies that suggest that the perception-production link changes over time. For instance, Rallo Fabra, and Romero (Reference Rallo Fabra and Romero2012) examined L1 Catalan speakers’ perception and production of English vowels. They reported an overall perception-production correlation of r = .26, a small effect that was not statistically significant. However, when they evaluated the link within three distinct learner proficiency groups, they found a large perception-production correlation (r = .76) in intermediate L2 speakers. In another study on L1 Mandarin speakers’ perception and production of English vowels, Jia et al. (Reference Jia, Strange, Collado and Guan2006) compared three groups: foreign language learners in China, second language learners who had lived in the US for less than 2 years, and another group of second language learners who had lived in the US for 3–5 years. The two second language groups were matched across a range of demographic variables, including age of arrival and age of onset of L2 instruction. The overall perception-production correlation was r = .50, and correlations for the foreign language learners and past arrivals were of similar magnitude (r = .42 and r = .46, respectively). However, for recent arrivals, the perception-production link was far weaker (r = .25). Although the goal of the study was to examine age and experience-related changes in the perception and production of L2 sounds, the between-group differences in the strength of the perception-production link are intriguing, insofar as they suggest distinct degrees of perception-production synchronization. In fact, the absence of a strong correlation in the recent arrival group could be interpreted as evidence of a change in the link. For example, a weak correlation could signal a lagged relationship, in which case cross-lagged measures would show a stronger relationship than their time-locked counterparts. To that point, longitudinal developmental studies—studies that observed perception-production relationships over time without providing training in either modality—point to a lagged model (Casillas, Reference Casillas2020a, Reference Casillas2020b; Nagle, Reference Nagle2018).

These intriguing findings underscore the need for a time-sensitive view of the perception-production link that is informed by environmental (e.g., the type of input that learners are exposed to) and speaker (e.g., the frequency and quality of L2 interactions) variables. Cross-sectional studies can be valuable, serving as a first step toward identifying the variables that should be measured longitudinally, but they are limited in terms of the perception-production questions that can be examined. Cross-sectional research can speak to the strength of perception-production relationships in different learner populations, but it cannot shed light on the developmental questions that have the greatest potential to advance theory and practice: Precisely when does perception being to influence production? More specifically, is there a dynamic coupling between perception and production, with changes in perception mirrored in production, or must perception accuracy reach a certain threshold before production begins to improve? To what extent does rate of change in perception predict rate of change in production? And how does the strength of the perception-production relationship change over time?

From a dynamic perspective, it is easy to imagine how perception-production relationships might change. For example, at the outset of learning, perhaps perception and production improve relatively quickly, with most learners reaching a moderate level of accuracy in both modalities within the first year of intensive L2 exposure (a period commonly referred to as the window of maximal opportunity for pronunciation learning; see Derwing & Munro, Reference Derwing and Munro2015). During this period, a large correlation between (cross-lagged) perception and production measures might be observed, with the correlation increasing in strength over time. Perception accuracy might even be the strongest single predictor of production accuracy. Once perception begins to stabilize, entering a developmental plateau, the perception-production link itself might also begin to stabilize, such that no change is observed in the link during, or the effect of perception on production may decrease in strength or disappear altogether.

This stasis might continue until a new experience, such as targeted pronunciation training, catalyzes additional development in either modality. At that point, a similar developmental cycle might ensue: perception and production improve, albeit at different rates, resulting in a relatively strong cross-lagged perception-production link that decays over time as perception fades from the strongest predictor, to one of many predictors, to having little predictive value at all. This scenario is of course hypothetical, but the point is that a developmental approach necessitates a consideration of the timing, strength, and duration of the effect of perception on production. It also demands careful consideration of what time-varying predictors should be sampled. For example, if perception and production learning are lexically driven, as many scholars have suggested (Best & Tyler, Reference Best, Tyler, Munro and Bohn2007; Flege, Reference Flege and Strange1995; Flege & Bohn, Reference Flege, Bohn and Wayland2021), then examining changes in vocabulary size over time would be important.

Thus far, we have focused on a macro-level perspective of the perception-production link, discussing how perception and production change and influence one another on a timescale of months or years. However, perception and production might also display substantial within-subjects variation on much shorter timescales, making it possible to distinguish between broad developmental processes and state-like variation in the perception-production link. A state-level view of the link opens up a new domain of research questions related to the stability of the perception and production systems. For instance, do individual differences in L1 and L2 use trigger temporary variation in perception and production accuracy? Although it can be tempting to attribute variation in either modality to measurement error, variation may in fact be indicative of the volatility of the developing systems. Even in advanced L2 users, it seems reasonable to expect perception, production, and the perception-production link to vary in response to varying patterns of L1 and L2 use. Perhaps L1 use temporarily disrupts, or even decouples, L2 perception and production, in the same way that producing sounds during perceptual training can disrupt perceptual learning (Baese-Berk, Reference Baese-Berk2019; Baese-Berk & Samuel, Reference Baese-Berk and Samuel2016).

In summary, the current state of perception-production research has been informed by a large number of cross-sectional studies, which are not well-suited to capture the time-varying nature of the perception-production link (Nagle, Reference Nagle2021). What is needed then, is a departure from current methods and a reorientation toward a more dynamic developmental approach that is predicated on longitudinal sampling. Such sampling strategies hold the key to understanding both how perception and production change over time and how changes in perception relate to changes in production after controlling for other important developmental phenomena (e.g., vocabulary size, patterns of L1 and L2 use).

LEARNING SCENARIO

A deeper understanding of how the nature of the learning scenario/target structure affects development in each modality and the perception-production link can enhance models of L2 sound learning and inform the development of optimal instructional practices. Current models assume that L1 sound patterns determine how learners perceive and produce L2 sounds. Crosslinguistically similar L2 sounds are viewed as especially challenging because L2 learners may associate those sounds with L1 categories. Yet, crosslinguistic phonetic similarity as a theoretical construct encompasses a variety of qualitatively distinct subcategories, each of which could be characterized by different perception-production relationships. Practically speaking, this means that comparing studies that examine different subclasses of crosslinguistic similarity would be akin to comparing apples and oranges, which could explain divergent results in the literature. Put another way, learning scenario is likely to be an important moderator variable that must be accounted for in perception-production research.

One common scenario a learner is faced with is what the L2LP calls a boundary shift (van Leussen & Escudero, Reference van Leussen and Escudero2015). In this scenario, both the L1 and the L2 contain the same number of phonological categories, and those categories are phonetically implemented using the same cue, but the phonetic boundary is not the same in the two languages. For instance, English and Spanish both contain a two-way stop consonant voicing contrast (e.g., /b/-/p/) that is predominantly cued by differences in voice onset time (Lisker & Abramson, Reference Lisker and Abramson1964). However, they differ with respect to the crossover boundary in perception and average voice onset time values in production. In English, phonologically voiced stops are phonetically realized as prevoiced or short-lag, whereas phonologically voiceless stops are realized as long-lag (/b/ is implemented as either [b] or [p], whereas /p/ is implemented as [pʰ], except after /s/). This contrasts with Spanish, where voiced stops are prevoiced and voiceless stops short-lag (/b/ is implemented as [b] and /p/ as [p]). Thus, an English speaker unfamiliar with Spanish would likely perceive Spanish voiced and voiceless stops as instances of English voiced stops. To perceive Spanish stops correctly—and produce them correctly, assuming that accurate perception is one of the primary determinants of accurate production—English speakers would need to learn to distinguish between prevoiced and short-lag variants, associating prevoiced variants with phonological voicing and short-lag variants with phonological voicelessness. As this example makes apparent, this might involve retuning the relationship between phonetics and phonology to match the L2, or from an SLM(-r) perspective, developing new L2 phonetic categories for Spanish stops.

A second learning scenario is when both languages use the same phonetic cue but the L2 contains more phonological categories than the L1. In this case, learning entails establishing an entirely new phonological category in a region of phonetic space where there is only one category in the L1. For example, Thai has a three-way stop consonant voicing contrast that, like the two-way stop contrast in English, is cued by differences in voice onset time (i.e., same phonetic cue, different number of phonological categories). In this case, English speakers would need to create separate categories for prevoiced and short-lag stops, both of which correspond to English voiced stops (i.e., they need to create a three-way /b/-/p/-/pʰ/ distinction). Creating a new category is qualitatively distinct from shifting boundaries or retuning the phonetics-phonology interface to match the L2. Thus, different perception-production relationships could arise for L2 Spanish and L2 Thai stops.

Learners may sometimes be faced with a scenario in which the L2 contains more phonological categories than the L1 and implements those categories using a novel cue. This would be the case for English and Korean stops, given that Korean contains a three-way stop consonant contrast that is jointly cued by voice onset time and fundamental frequency in the following vowel (see, e.g., Schertz et al., Reference Schertz, Cho, Lotto and Warner2015). As a result, English speakers would need to learn to attend to a novel phonetic cue that is not primary in the L1 to create accurate perceptual representations for Korean stop consonants. Whereas English speakers learning Thai might be able to begin developing new categories relatively quickly because the two languages make use of the same phonetic cue, it is reasonable to hypothesize that perceptual learning in Korean would proceed more gradually given the presence of a novel cue. These crosslinguistic relationships are exemplified in Figure 1.

FIGURE 1.

Crosslinguistic comparison of learning scenarios: English vs. Spanish, Thai, and Korean.

Note. For English, Spanish, and Thai, the primary phonetic cue to the stop consonant contrast is voice onset time. For Korean, the stop consonant contrast is cued by both voice onset time and fundamental frequency (f₀); hence, the two-dimensional space for the new category new cue example involving Korean (f₀ is shown on the vertical axis, and voice onset time on the horizontal axis). Phonological categories appear between forward slashes (/b/) and phonetic categories between brackets ([b]).

It is important to note that the examples above are only one set of possibilities for how the phonetic space may be different between the L1 and L2. All of these cases reflect situations where multiple categories exist in each language. It is also often the case that the L1 contains only one category where the L2 has two. Two commonly researched examples are Japanese speakers’ perception and production of the English /r/-/l/ contrast and non-native English speakers’ perception and production of English tense-lax vowel contrasts such as /i/-/ɪ/. Both of these examples would require learners to begin attending to the relevant phonetic cue in the L2. Although at face value they seem to line up well with the English/Korean stop consonant example (i.e., adding a new category using a new cue), in fact, there is a critical difference: in one-to-two category scenarios, there is not necessarily a preexisting cue in the L1 that could influence L2 perception, whereas in two-to-three category scenarios, there is a preexisting L1 cue that learners may need to retune to fit L2 characteristics. For instance, even though both Korean and English make use of voice onset time, the way the cue is used in the two languages is different. This means that English speakers would need to adjust their perception of voice onset time to fit Korean while simultaneously learning to attend to f₀, a novel cue.

Finally, there are also crosslinguistic differences in phonotactics. For example, English and Arabic show a similar voicing contrast for alveolar stops (/d/-/t/, phonetically [d]-[tʰ]), but Arabic does not have the same contrast at the bilabial and velar places of articulation, where only /b/ and /k/ are instantiated at a phonological level. Thus, Arabic speakers of English would need to create a new phonological contrast, analogous to their native Arabic /t/-/d/ contrast, at two new places of articulation. Likewise, English contrasts initial, medial, and final stops, whereas many languages only show a stop consonant contrast word-initially or word-medially, as in Spanish. Moreover, each of these phonological environments is associated with a different set of cues: in English, voice onset time word-initially, voice onset time and closure duration word-medially, and duration of the preceding vowel word-finally.

In summary, as a conceptual category, crosslinguistic similarity includes a variety of meaningful subclasses, each of which could be associated with distinct perception-production patterns. Future perception-production work must systematically examine these subclasses, taking stock of crosslinguistic differences in phonetic, phonological, and phonotactic patterns, to examine how the perception-production link manifests across a range of learning scenarios.

INDIVIDUAL DIFFERENCES

Individual differences could impact the perception-production link in two different ways: individual differences in the nature of the link itself, in which case it would be appropriate to discuss qualitatively different perception-production links that arise due to individual differences in, for instance, auditory processing ability, or individual differences in the way the link (i.e., a single, invariant perception-production relationship) manifests. There is reason to believe that qualitatively distinct perception-production links might exist, perhaps even within the same learner at different points in time.

Chandrasekaran et al. (Reference Chandrasekaran, Koslov and Maddox2014) have argued that L2 sound learning should be conceptualized as a general categorization problem that involves two neurobiologically distinct systems: a reflective system that relies on working memory to test explicit categorization rules and a reflexive system that associates perceptual stimuli with motor outputs. A dual-learning system approach has implications for the perception-production link. Namely, between- and within-subjects variation in reflective and reflexive strategy use could lead to different perception-production relationships. Some learners may show a greater reliance on one system than the other, and the same learner may show varying levels of reliance on the two systems at different points during L2 learning. A learner might initially use the reflective system before switching to the reflexive system. In fact, the learning environment may stimulate differences in system use. For instance, in many instructed contexts, learners receive explicit pronunciation instruction, which could encourage the use of reflective strategies.

Another possibility is that individual differences influence learning in each modality, which could in turn affect how the perception-production link manifests, that is, the rate at which perception and production become aligned and the strength and duration of that alignment. For example, Cerviño-Povedano and Mora (Reference Cerviño-Povedano, Mora, Dziubalska–Kołaczyk, Wrembel and Kul2010) found that learners with greater phonological short-term memory identified L2 vowels more accurately and were less reliant on secondary (i.e., less informative) phonetic cues in perception. In another study, Darcy et al. (Reference Darcy, Mora and Daidone2016) reported a significant link between inhibitory control and L2 vowel discrimination accuracy. These cross-sectional findings suggest that individual differences in cognitive skills might help learners develop accurate L2 phonological representations and block interference from L1 categories during L2 learning. However, this claim would need to be evaluated longitudinally to determine the extent to which these variables are associated with variation in perception-production learning over time.

Another candidate for research would be individual differences in language use. Although all models of L2 sound learning address this topic, L2LP (van Leussen & Escudero, Reference van Leussen and Escudero2015) and Simulation Theory (Gambi & Pickering, Reference Gambi and Pickering2013) are particularly germane to the present discussion. If perceptual learning is error-driven, as van Leussen and Escudero (Reference van Leussen and Escudero2015) suggested, and if perception facilitates production (Escudero, Reference Escudero and Pennington2007), then learners who engage in the L2 frequently (quantity of L2 use) and extensively (quality of L2 use) might show more rapid development in each modality, and perhaps a stronger perception-production link, than learners whose patterns of L2 use are more circumscribed. L2 use is also central to Simulation Theory, according to which perception is grounded in the covert simulation of speech:

From this viewpoint, the more often an individual interacts in the L2, the more often they would have the opportunity to fine-tune perception-production relationships. A full presentation of Simulation Theory is beyond the scope of this study, but the point is that, according to this and other approaches, L2 use appears to be critical for establishing and adjusting associations between the acoustic signal and motor commands. Many studies have examined the relationship between L2 use and speech production accuracy, but to the best of our knowledge, no study has examined it in relation to the perception-production link itself.

Individual differences might also interact with pronunciation training paradigms, which could influence perception-production relationships through aptitude-treatment interactions. For example, Perrachione et al. (Reference Perrachione, Lee, Ha and Wong2011) found that high aptitude learners (individuals who performed well on pitch perception tests) performed well in high variability phonetic training, whereas for low aptitude learners, certain forms of high variability training were detrimental. Although not strictly an individual difference, one other mediating factor that deserves attention is the nature of the training paradigm itself. Baese-Berk and Samuel (Reference Baese-Berk and Samuel2016) found that an integrated perception-production training paradigm had a negative impact on perceptual learning; perceptual gains were not as robust for the group of learners who had to produce sounds after hearing them compared with the group that engaged in perception-only training. One could imagine how individual differences (e.g., phonological short-term memory, inhibitory skill) might also play a role in mitigating the disruptive effect of production on perception in integrated training paradigms.

From a theoretical perspective, we have identified three key areas that future perception-production research should consider: developmental stage, learning scenario, and individual differences. In the next sections, we address methodological considerations, emphasizing the distinct methodologies that would be needed to carry out robust perception-production research.

MEASURING PERCEPTION

Speech perception is a complex cognitive process that is grounded in the integration of different types of information available at different levels of linguistic structure and memory (e.g., the speech signal itself, phonotactic probability, knowledge of the target variety or even the individual speaker). A variety of tasks have been used to measure speech perception, and each task may tap into slightly different aspects of a listener’s ability to process new speech sounds. It is important to note that it is impossible to divorce the perceptual behaviors we are interested in measuring from the tasks we use to measure them. Therefore, in the current section, we explore methods that are often used to measure speech perception. For each measure, we address what participants are being asked to do, what the outcome measure of the task is, and what this outcome measure is likely to reflect.

We also discuss the stimuli that are used and how they are paired with tasks to yield categorical and gradient perspectives on perceptual processing. Each decision that is made in experimental design has an impact on what is being measured and how the resulting behavioral data can be interpreted. There is significant overlap in what tasks can tap into, so we believe it is important not to divide tasks into strictly independent bins and assume that all behavior within these bins reflects the same perceptual process.

Before we focus on perception tasks as a whole, it is important to consider stimuli characteristics because they can have an impact on the type of perceptual processing in which listeners engage. Stimuli can be naturally produced tokens or tokens drawn from a synthesized phonetic continuum. When exposed to naturally produced tokens that are canonical exemplars of target categories, listeners may respond in a more categorical way, drawing upon the rich acoustic variation present in the stimuli to make perceptual judgments. Likewise, when exposed to stimuli drawn from a continuum, they may respond in a finer-grained manner, directing their perceptual processing to precisely those dimensions along which stimuli vary.

Put another way, synthesized stimuli, which have been created by systematically varying target phonetic dimensions to create a series of steps, are tightly controlled, and, therefore, may elicit a more gradient response from the listener.Footnote ² Similarly, stimuli drawn from multiple talkers may encourage more categorical performance than stimuli drawn from a single talker because of the abstraction processes required to interpret different acoustic signals as members of the same speech sound category. Even the inclusion of different types of filler trials can result in different perceptual performance. For example, Baese-Berk (Reference Baese-Berk2010) demonstrated that native English listeners appeared to be less sensitive to a novel contrast they had recently been trained on (i.e., prevoiced [da] vs. short-lag [ta]) when the test included filler trials corresponding to a known contrast (i.e., [ma] vs. [la]) than when the target stimuli were presented without filler trials. This suggests that what listeners interpret as phonologically and/or phonetically similar may depend on the types of contrasts to which they are exposed, including filler contrasts. Researchers should be sensitive to these issues (natural vs. synthetic, single vs. multiple talkers, filler items, etc.) when developing stimuli and, as laid out below, when pairing stimuli with different perceptual tasks.

Stimuli can be embedded in a variety of perceptual tasks, which means that stimuli and task characteristics necessarily interact to affect the type of processing in which listeners engage. Perceptual tasks can be roughly divided into two types: those that require listeners to compare multiple speech sounds and those that require listeners to engage with a single target sound at a time. Single-target tasks tend to be simpler, insofar as listeners hear a target stimulus (e.g., a sound, syllable, word) and are asked to make a judgment about it. This judgment may require the participant to match the input with an orthographic representation,Footnote ³ a picture, or a motor response (i.e., press right when you hear one sound, and left when you hear another). Regardless of the matching procedure, the fundamental task participants are being asked to complete is one of labelling or identification.

This type of task can be used to evaluate how listeners map auditory exemplars onto representations stored in long-term memory. As a result, it can provide information on the content of emerging perceptual categories. If the task is paired with naturally produced stimuli, it can provide insight into whether learners have begun to create separate perceptual categories for L2 sounds, and when paired with stimuli drawn from a phonetic continuum, it can shed light on phonetic cue use and cue weights. For instance, Schertz et al. (Reference Schertz, Cho, Lotto and Warner2015) created two sets of 141 stimuli varying along three acoustic dimensions to examine Korean speakers’ perception (and production) of Korean and English stop consonants. Listeners heard each stimulus and were asked to classify it using a closed set of options. Similar tasks have been used to evaluate L2 listeners’ reliance on spectral and duration cues in the perception of L2 vowel contrasts (e.g., Flege et al., Reference Flege, Bohn and Jang1997; Sakai, Reference Sakai2016) and to gain insight into the precise location of L2 phonemic boundaries, including how such boundaries change over time as a function of L2 experience (e.g., Casillas, Reference Casillas2020b).

Like their single-stimulus counterparts, perceptual tasks requiring listeners to compare multiple stimuli can take many forms. One common example is an AX discrimination task, where an anchor stimulus is presented (A) followed by a target stimulus (X), and the listener is instructed to decide if the two stimuli are the same or different. Another common task is an ABX categorization task. In this case, two anchor stimuli are presented (A and B) followed by a target stimulus (X), and the listener is asked to assign the target stimulus to one of the anchors. This task is more complex than AX discrimination because it entails goodness of fit comparisons (between X and A and X and B) and binning. That is, listeners need to group like sounds together, which suggests that they have at least tacit knowledge of the dimensions along which X and A/B differ.Footnote ⁴

Oddity tasks are an amalgamation of both discrimination and categorization. On this task, three stimuli are presented, and the listener is asked to determine if they pertain to the same category or if there is an odd item out. If there is an odd item, they are asked to indicate its serial position (1, 2, or 3). As a result, on an oddity task, listeners must compare multiple pairs of stimuli (1 vs. 2, 2 vs. 3, and 1 vs. 3), deciding if they are the same or different, while simultaneously determining if the degree of correspondence between the pairs is roughly equal. These tasks provide insight into how sensitive listeners are to a given contrast. Yet, as with single-stimulus tasks, AX, ABX, and oddity tasks can be rendered more categorical or gradient depending on the stimuli that are used.

Furthermore, the length of the interstimulus interval has been shown to play a key role in the type of processing in which listeners engage (Schouten et al., Reference Schouten, Gerrits and Van Hessen2003; Werker & Logan, Reference Werker and Logan1985). Shorter intervals seem to encourage gradient (i.e., acoustic or phonetic) processing, whereas longer intervals are associated with categorical (i.e., phonological) processing. Because these tasks involve storing and comparing more than one stimulus, individual differences in phonological short-term memory and auditory processing could also affect task performance. Importantly, the ability to discriminate similar sounds under certain conditions does not suggest the ability to identify those sounds correctly; discrimination and identification are not the same.

After developing perception tasks, researchers must decide how they will code and analyze the resulting data. Some tests are amenable to a single outcome measure (e.g., synthetic continua lend themselves to phonemic boundaries and cue weights), but others permit a range of options. Researchers are often interested in quantifying accuracy. Accuracy can be assessed at a trial level, as the probability of a correct response (e.g., Kartushina et al., Reference Kartushina, Hervais-Adelman, Frauenfelder and Golestani2015), or globally, as percent correct or a discrimination index. Sensitivity indices such as d-prime (d’) are advantageous because they take response bias into account.Footnote ⁵ Researchers might also choose to examine reaction times on correct response trials to gain insight into speed of phonological processing. It is important to determine which measure is most appropriate for the test and stimuli used, as these factors are quite likely to influence interpretation of the results. Of course, tasks, stimuli, and outcome measures should also be aligned with research questions.

To this point, we have demonstrated that the tasks used to measure perception vary substantially and that the outcomes being measured and interpreted may differ as a function of task specifics, including the stimuli used. We now shift our attention to how these tasks can be used to answer questions of development of L2 perception. Models of L2 sound learning suggest that multiple measures would be needed to understand L2 perceptual development. For instance, the SLM(-r) posits a specific sequence: learners need to become sensitive to fine-grained crosslinguistic differences before they can begin creating novel categories for L2 sounds. In that case, discrimination and categorization tests could be used to examine sensitivity to such differences, and identification could be used to evaluate the emergence of L2 categories.

PAM-L2 also assigns a central role to crosslinguistic similarity in shaping perceptual learning. According to this model, crosslinguistic perceptual assimilation patterns determine the relative difficulty of L2 contrasts. Research has shown that even learners from the same L1 background show variable patterns, which can affect the initial state of L2 perception and the difficulty that learners experience over time (Mayr & Escudero, Reference Mayr and Escudero2010). Thus, a comprehensive investigative approach to L2 perceptual learning would include examining learners’ L1-L2 perceptual assimilation patterns and tracking their discrimination, categorization, and identification performance longitudinally. Such an approach can provide insight into two key questions: Do individual differences in L1-L2 perceptual assimilation patterns affect the starting point for L2 perception? And does discrimination/categorization accuracy predict identification accuracy?

Indeed, a complete understanding of perceptual development during L2 learning rests upon examining how different perceptual processes unfold over time. Thus, defining perception based on learners’ performance on a single perception test is theoretically and methodologically problematic. For one, different processes likely develop at different rates, which means that the right test must be used at the right time to accurately capture the state of learners’ developing systems. Interpreting the results of any single perception test as representative of the entire system can lead to imprecise generalizations about the nature of perceptual learning. This imprecision is magnified when studies that use fundamentally different tests (e.g., discrimination vs. identification) or implement the same test in different ways are compared. In these cases, performance variation may be due to nontrivial differences in task characteristics. In summary, then, we propose that a variety of perception tasks must be coordinated and deployed to understand L2 perceptual learning. Notably, performance on discrimination and categorization tests should improve before performance on identification tests if discerning crosslinguistic sound differences is a necessary first step toward creating L2 categories. We also advocate for examining performance at various levels of granularity (e.g., by adopting both categorical and fine-grained perspectives on perceptual accuracy and learning).

MEASURING PRODUCTION

Like perception, production is a complex skill that can be measured and defined in a variety of ways. And like perception, different facets of production ability likely develop at different rates, which means that researchers must take care to select and sequence appropriate production tasks over appropriate developmental windows. In their measurement framework, Saito and Plonsky (Reference Saito and Plonsky2019) observed that production accuracy can be measured in controlled and spontaneous speech. They linked control production to declarative pronunciation knowledge and spontaneous production to procedural pronunciation knowledge. They also distinguished between three coding options that yield different perspectives on production accuracy: acoustic measurements, expert ratings of linguistic features, and listener intuition.

Although current models of L2 sound learning do not address how different facets of production ability develop over time, one can imagine a potential developmental sequence based on the factors that Saito and Plonsky identified. For instance, the ability to produce an intelligible L2 contrast might emerge first in controlled speech, when speakers can focus on their pronunciation. Then, they might begin producing an intelligible contrast in spontaneous speech (Saito, Reference Saito2019). Over time, given enough input and the right combination of aptitude and interest, speakers might even begin to produce nativelike sounds, although most will fall short of that mark. In fact, even if L2 speakers are successful at mastering certain phonetic cues, they may struggle to achieve nativelike accuracy in all dimensions of L2 sound production, especially on spontaneous tasks (Saito, Reference Saito2013) and when nativelike accuracy is scrutinized at a phonetic or acoustic level (Stölten et al., Reference Stölten, Abrahamsson and Hyltenstam2014).

To examine controlled production knowledge, researchers commonly rely on word- and sentence-level tasks, including word and sentence reading, word and sentence repetition, and picture naming. These tasks are not without their pitfalls. For instance, reading tasks do not exclusively test production ability, but rather literacy skills (i.e., phoneme-to-grapheme mappings). This is not trivial because many languages, including English, have an opaque orthographic system that learners may not have fully mastered, especially when pronouncing low frequency forms that do not conform to prototypical sound-spelling patterns. Repetition tasks avoid the confounding effect of orthography. At the same time, they confound perception and production skills.

Because listeners are exposed to an auditory stimulus that they must process before repeating it, inaccurate production may be a result of inaccurate perceptual processing rather than production difficulty. What’s more, on an immediate repetition task, speakers may be able to store phonetic forms in short-term memory, potentially bypassing their own phonological system. Thus, both reading and repetition may obscure speakers’ true production ability, which could lead to a fuzzy or inaccurate view of the perception-production link (for studies that address differences in performance after orthographic vs. auditory input, see, e.g., Davidson, Reference Davidson2010; de Jong et al., Reference de Jong, Hao and Park2009; Kato & Baese-Berk, Reference Kato and Baese-Berk2020). One controlled task that avoids both confounds is picture naming. As long as researchers take care to select high frequency, imageable items that are likely to be familiar to speakers, picture naming may provide the clearest perspective on controlled production ability.

On the other side of the controlled-spontaneous spectrum, picture description and story narration have been used to investigate spontaneous production knowledge. These tasks are advantageous for understanding what speakers actually produce under realistic speaking conditions, but they also imply some risk. For instance, participants may not produce enough tokens of the target sound for reliable analysis. Furthermore, other sources of variation in spontaneous production, such as the phonetic context in which the sounds occur and the lexical characteristics of the carrier word, must be accounted for during data analysis. These concerns can be mitigated by, for example, providing speakers with a list of target words that they should use to describe the picture (e.g., Nagle, Reference Nagle2021; Saito & van Poeteren, Reference Saito and van Poeteren2017), but providing written target words introduces the same potential orthographic and literacy confounds that reading tasks do.

After researchers have determined what tasks they will use to examine controlled and/or spontaneous production knowledge, they must determine how they will define accuracy. Acoustic measurements provide an objective measure, but even acoustic measurements involve an element of choice (e.g., what features to measure, how to measure them), and such measurements may not reveal much about how listeners actually process L2 speech. Moreover, L2 speakers may implement L2 contrasts in nonnativelike ways, in which case they might use phonetic cues that acoustic analyses of canonical features (i.e., the features that monolingual speakers use to perceive and produce the target sound) would not detect. For this reason, targeted acoustic analysis may not capture all the differences L2 speakers make between sounds. Furthermore, it is rarely the case that speakers use a single phonetic cue to differentiate sounds, but without complex analytical tools, it can be difficult to integrate multiple cues into a single analysis. Some recent work using linear discriminant analysis (Mairano et al., Reference Mairano, Bouzon, Capliez and De Iacovo2019) has attempted to address this problem. The challenge of an integrated analysis, however, is understanding precisely how each cue contributes to the differentiation of sounds categories in production.

Listener-based measures offer an alternative perspective on production accuracy. Whereas acoustic measurements shed light on phonetic nativelikeness, listener-based approaches are rooted in the notion of intelligibility, or the extent to which speakers can produce an intelligible L2 contrast (the word the speaker intends to produce is beat, and the listener perceives the word as intended, hearing beat instead of bit; see, e.g., Munro & Derwing, Reference Munro and Derwing2008). Intelligibility has broad ecological validity, but it is also a relatively coarse-grained measure that does not provide insight into the precise characteristics of L2 sounds, which is often one of the central goals of (perception-)production research. For instance, L2 speakers may produce a covert acoustic contrast between two L2 sounds even if listeners do not perceive a difference (Song & Eckman, Reference Song and Eckman2019). Listeners can also be asked to evaluate production accuracy using a scalar rating system (see, e.g., Kissling, Reference Kissling2014; Lopez-Soto & Kewely -Port, Reference Lopez-Soto and Kewley-Port2009; Rochet, Reference Rochet and Strange1995; Saito & van Poeteren, Reference Saito and van Poeteren2017), or by comparing pre and posttest productions to one another (Bradlow et al., Reference Bradlow, Pisoni, Akahane-Yamada and Tohkura1997). However, relying on listener perception to assess production accuracy can introduce perceptual biases into production measures. To that point, a wide variety of factors, including lexicality, lexical frequency, semantic plausibility, and social information, are known to influence perception, especially the perception of ambiguous sounds (e.g., Ganong, Reference Ganong1980).

In summary, then, researchers must take care to determine what facet of production they are interested in examining. No single production task or accuracy measure can fully capture production performance at a single time point, much less how L2 production develops over time. Thus, we advise researchers to measure production in both controlled and spontaneous speech (and examine how different views of production accuracy relate to one another), although few studies have done so to date (for notable exceptions, see Kartushina & Frauenfelder, Reference Kartushina and Frauenfelder2014; Lambacher et al., Reference Lambacher, Martens, Kakehi, Marasinghe and Molholt2005; Saito & van Poeteren, Reference Saito and van Poeteren2017; Wang et al., Reference Wang, Jongman and Sereno2003).

EVALUATING THE PERCEPTION-PRODUCTION LINK

To evaluate the perception-production link, researchers must decide what perception and production measures should be compared and what would count as evidence of a robust link. Selecting measures for comparison is not trivial. Perception measures such as d’ represent sensitivity to contrast, whereas production measures represent the ability to produce the target phone intelligibly or accurately. It is, therefore, unclear whether a measure that encodes information about a contrast should be related to the production of either of the phones that make up that contrast. At the most basic level, it seems important to compare like with like; measures that tap into fine-grained aspects of perception (patterns of cue use, precise boundary locations, etc.) should be paired with fine-grained production measures (e.g., Schertz et al. Reference Schertz, Cho, Lotto and Warner2015), and categorical perception measures (discrimination and identification indices) should be paired with categorical production measures (intelligibility). It bears mentioning that intelligibility is often operationalized as the percentage of words that are correctly identified as the target word (i.e., the word the speaker intended to produce). However, intelligibility can also be quantified by means of contrast sensitivity metrics. In this case, the resulting score would represent the extent to which listeners perceive the contrast that the L2 speaker intended to produce. Such a production measure might be more closely aligned with its perceptual counterpart: d’ for perception would represent the L2 learner’s ability to discriminate contrastive L2 sounds, and d’ for production would represent that individual’s ability to produce a discriminable L2 contrast.

Another issue that deserves scrutiny is the fact that perception tests are implemented in a controlled listening context in which the target forms are often presented in isolation without noise. Yet, production can be measured in controlled or spontaneous speech. It stands to reason that perception measures would bear a stronger relationship to controlled production measures than spontaneous production measures (Nagle, Reference Nagle2021). One future goal for perception research should, therefore, be to test perception in more spontaneous listening contexts that require listeners to process speech for meaning (see, e.g., Kim et al., Reference Kim, Cho, Cho, Hong, Kim and Nam2020).

Evaluating perception-production links also entails making choices about statistical tests. This choice is clearly constrained by the research design and research questions, but a variety of descriptive and inferential approaches are nonetheless possible. In early research, it was common to analyze perception and production separately and subsequently compare performance in the two modalities (Bohn & Flege, Reference Bohn, Flege, Allan and Leather1997; Caramazza et al., Reference Caramazza, Yeni-Komshian, Zurif and Carbone1973; Gass, Reference Gass1984; Mack, Reference Mack1989; Sheldon & Strange, Reference Sheldon and Strange1982; Zampini, Reference Zampini1998). This analysis constitutes a type of rank ordering that reveals broad information on perception-production patterns such as accuracy in perception is better than (i.e., leads or precedes) accuracy in production, accuracy in production seems to outpace accuracy in perception, and so on. Few contemporary studies use rank ordering exclusively, but it is often presented as one aspect of a more comprehensive set of analyses. For instance, numerous studies that have rank ordered contrasts according to their difficulty in perception and production (Evans & Alshangiti, Reference Evans and Alshangiti2018; Jia et al., Reference Jia, Strange, Collado and Guan2006; Kartushina & Frauenfelder, Reference Kartushina and Frauenfelder2014; Rallo Fabra & Romero, Reference Rallo Fabra and Romero2012) have shown that difficulty in one modality is not always reflected in the other.

Correlation and regression are the most common inferential tests used to assess the strength of perception-production relationships. It is important to recognize that neither test suggests directionality or causality, and, when applied to cross-sectional data, these techniques run the risk of mischaracterizing the link based on the precise moment at which it was measured.

Some researchers have achieved a finer-grained perspective by using gain scores to examine whether changes in perception are associated with changes in production (Huensch & Tremblay, Reference Huensch and Tremblay2015; Kartushina & Frauenfelder, Reference Kartushina and Frauenfelder2014). Yet, this approach has similar limitations insofar as it renders a relatively circumscribed view of what should be considered a dynamic developmental phenomenon. It comes as no surprise then, that findings range from no significant perception-production link at all in some studies (Hattori & Iverson, Reference Hattori and Iverson2010; Kartushina & Frauenfelder, Reference Kartushina and Frauenfelder2014; Schertz et al., Reference Schertz, Cho, Lotto and Warner2015; Thorin et al., Reference Thorin, Sadakata, Desain and McQueen2018) to medium to large correlations in others (Baker & Trofimovich, Reference Baker and Trofimovich2006; Borden et al., Reference Borden, Gerber and Milsark1983; Lopez-Soto & Kewley-Port, Reference Lopez-Soto and Kewley-Port2009).

Examining the effect of (perception) training is another common means of testing the link. Rather than directly comparing performance in the two modalities (or in addition to this comparison), training studies address how experience in one modality impacts learning in the other. According to Sakai and Moorman’s (Reference Sakai and Moorman2018) meta-analysis of perception training studies, perception training leads to small, yet significant, gains in posttest production accuracy. At the same time, they found that gains in perception were not significantly correlated with gains in production, and individual perception training studies have yielded a wide range of results. Researchers have also investigated the effect of production training and integrated perception-production training paradigms on perceptual learning. Here too, a range of effects have been observed: production training leading to medium gains in perception (Sakai, Reference Sakai2016); no influence of production training on perception (Thorin et al., Reference Thorin, Sadakata, Desain and McQueen2018); and disruption of perceptual learning (Baese-Berk, Reference Baese-Berk2019; Baese-Berk & Samuel, Reference Baese-Berk and Samuel2016). However, as in other areas of perception-production research, training studies vary widely in methodological choices, choices that have a direct impact on perception-production findings (cf. Sakai & Moorman, Reference Sakai and Moorman2018).

Apart from these issues, current approaches suffer from two weaknesses. First, most studies have been relatively short-term, insofar as they have analyzed the link at a single point in time or over two to three data points (e.g., pre-post-delayed training studies). Yet, a robust test of the link would require demonstrating that within-subjects changes in perception guide within-subjects changes in production consistently over longer periods. That is, perception must be conceptualized as a time-varying predictor of production (Nagle, Reference Nagle2021). It would also be important to test how the strength of the link changes over time, as might be accomplished through Time × Perception interaction terms. Another important topic that research has not evaluated statistically is directionality. Directionality in longitudinal studies can be evaluated through cross-lagged analyses, such as cross-lagged panel models (longitudinal structural equation models that can estimate the impact of one variable on another over time; for an introduction to longitudinal structural equation models, see Little, Reference Little2013). Such models can be used to evaluate Flege and Bohn’s (Reference Flege, Bohn and Wayland2021) hypothesis that perception and production co-evolve. They can also provide insight into the weights of reciprocal relationships. One challenge will be achieving the sample size required for statistical power.